Daily Papers

Recent thinking models solve complex reasoning tasks by scaling test-time compute, but this scaling must be allocated in line with task difficulty. On one hand, short reasoning (underthinking) leads to errors on harder problems that require extended reasoning steps; but, excessively long reasoning (overthinking) can be token-inefficient, generating unnecessary steps even after reaching a correct intermediate solution. We refer to this as under-adaptivity, where the model fails to modulate its response length appropriately given problems of varying difficulty. To address under-adaptivity and strike a balance between under- and overthinking, we propose TRAAC (Think Right with Adaptive, Attentive Compression), an online post-training RL method that leverages the model's self-attention over a long reasoning trajectory to identify important steps and prune redundant ones. TRAAC also estimates difficulty and incorporates it into training rewards, thereby learning to allocate reasoning budget commensurate with example difficulty. Our approach improves accuracy, reduces reasoning steps, and enables adaptive thinking compared to base models and other RL baselines. Across a variety of tasks (AIME, AMC, GPQA-D, BBEH), TRAAC (Qwen3-4B) achieves an average absolute accuracy gain of 8.4% with a relative reduction in reasoning length of 36.8% compared to the base model, and a 7.9% accuracy gain paired with a 29.4% length drop compared to the best RL baseline. TRAAC also shows strong generalization: although our models are trained on math datasets, they show accuracy and efficiency gains on out-of-distribution non-math datasets like GPQA-D, BBEH, and OptimalThinkingBench. Our analysis further verifies that TRAAC provides fine-grained adjustments to thinking budget based on difficulty and that a combination of task-difficulty calibration and attention-based compression yields gains across diverse tasks.

Pruning deep neural networks is a widely used strategy to alleviate the computational burden in machine learning. Overwhelming empirical evidence suggests that pruned models retain very high accuracy even with a tiny fraction of parameters. However, relatively little work has gone into characterising the small pruned networks obtained, beyond a measure of their accuracy. In this paper, we use the sparse double descent approach to identify univocally and characterise pruned models associated with classification tasks. We observe empirically that, for a given task, iterative magnitude pruning (IMP) tends to converge to networks of comparable sizes even when starting from full networks with sizes ranging over orders of magnitude. We analyse the best pruned models in a controlled experimental setup and show that their number of parameters reflects task difficulty and that they are much better than full networks at capturing the true conditional probability distribution of the labels. On real data, we similarly observe that pruned models are less prone to overconfident predictions. Our results suggest that pruned models obtained via IMP not only have advantageous computational properties but also provide a better representation of uncertainty in learning.

Controlling an acting LLM agent means deciding, at each step of a long-horizon task, whether to trust its next action, abstain, or hand off to a human. That decision needs a per-step confidence signal predicting whether the current rollout will eventually succeed. The agent's own self-reported verbalized confidence is the natural candidate, but existing reinforcement learning methods calibrate it against the correctness of a decision at the current step. We propose a method to calibrate it against the eventual outcome of a multi-step episode: under GRPO, the group of rollouts for a task supplies success labels, and a proper scoring rule scores each step's confidence against the realized outcome. In offline RL on recorded trajectories of a Qwen2.5-7B agent in ALFWorld, TextWorld, and WebShop environments, the AUROC of step confidence against episode outcome rises from near chance to 0.68-0.85, and the top-10% of steps the agent is most confident about succeed ~4.3x more often than average. In online RL on the live ALFWorld agent, four calibration rewards improve early-step AUROC, expected calibration error goes from 0.42 to below 0.12, and selective execution improves over both the untrained agent and a format-only RL control, without degrading task success. We also introduce an evaluation protocol that separates episode depth, task difficulty, and attempt-level foresight, which shows where the gain lies: the trained agent recognizes from its first step which tasks it will fail, while telling a succeeding attempt from a failing one at the same task is harder, with some early-window signal emerging on TextWorld.

Agentic repository-level code understanding is essential for automating complex software engineering tasks, yet the field lacks reliable benchmarks. Existing evaluations often overlook the long tail topics and rely on popular repositories where Large Language Models (LLMs) can cheat via memorized knowledge. To address this, we introduce SWE-QA-Pro, a benchmark constructed from diverse, long-tail repositories with executable environments. We enforce topical balance via issue-driven clustering to cover under-represented task types and apply a rigorous difficulty calibration process: questions solvable by direct-answer baselines are filtered out. This results in a dataset where agentic workflows significantly outperform direct answering (e.g., a ~13-point gap for Claude Sonnet 4.5), confirming the necessity of agentic codebase exploration. Furthermore, to tackle the scarcity of training data for such complex behaviors, we propose a scalable synthetic data pipeline that powers a two-stage training recipe: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning from AI Feedback (RLAIF). This approach allows small open models to learn efficient tool usage and reasoning. Empirically, a Qwen3-8B model trained with our recipe surpasses GPT-4o by 2.3 points on SWE-QA-Pro and substantially narrows the gap to state-of-the-art proprietary models, demonstrating both the validity of our evaluation and the effectiveness of our agentic training workflow.

Large Language Models (LLMs) employ Chain-of-Thought (CoT) reasoning to deconstruct complex problems. While longer CoTs are often presumed superior, this paper challenges that notion, arguing that longer is not always better. Drawing on combined evidence from real-world observations, controlled experiments, and theoretical analysis, we demonstrate that task accuracy typically follows an inverted U-shaped curve with CoT length, where performance initially improves but eventually decreases as the number of CoT steps increases. With controlled experiments, we further uncover the scaling behaviors of the optimal CoT length: it increases with task difficulty but decreases with model capability, exposing an inherent simplicity bias where more capable models favor shorter, more efficient CoT reasoning. This bias is also evident in Reinforcement Learning (RL) training, where models gravitate towards shorter CoTs as their accuracy improves. To have a deep understanding of these dynamics, we establish a simple theoretical model that formally proves these phenomena, including the optimal length's scaling laws and the emergence of simplicity bias during RL. Guided by this framework, we demonstrate significant practical benefits from training with optimally-lengthed CoTs and employing length-aware filtering at inference. These findings offer both a principled understanding of the "overthinking" phenomenon and multiple practical guidelines for CoT calibration, enabling LLMs to achieve optimal reasoning performance with adaptive CoTs tailored to task complexity and model capability.

Protein language models (PLMs) have advanced computational protein science through large-scale pretraining and scalable architectures. In parallel, reinforcement learning (RL) has broadened exploration and enabled precise multi-objective optimization in protein design. Yet whether RL can push PLMs beyond their pretraining priors to uncover latent sequence-structure-function rules remains unclear. We address this by pairing RL with PLMs across four domains: antimicrobial peptide design, kinase variant optimization, antibody engineering, and inverse folding. Using diverse RL algorithms and model classes, we ask if RL improves sampling efficiency and, more importantly, if it reveals capabilities not captured by supervised learning. Across benchmarks, RL consistently boosts success rates and sample efficiency. Performance follows a three-factor interaction: task headroom, reward fidelity, and policy capacity jointly determine gains. When rewards are accurate and informative, policies have sufficient capacity, and tasks leave room beyond supervised baselines, improvements scale; when rewards are noisy or capacity is constrained, gains saturate despite exploration. This view yields practical guidance for RL in protein design: prioritize reward modeling and calibration before scaling policy size, match algorithm and regularization strength to task difficulty, and allocate capacity where marginal gains are largest. Implementation is available at https://github.com/chq1155/RL-PLM.

Following the standard supervised fine-tuning (SFT) paradigm, in-context learning (ICL) has become an efficient approach propelled by the recent advancements in large language models (LLMs), yielding promising performance across various tasks in few-shot data setups. However, both paradigms are prone to suffer from the critical problem of overconfidence (i.e., miscalibration), especially in such limited data setups. In this work, we deliver an in-depth analysis of the behavior across different choices of learning methods from the perspective of both performance and calibration, as well as their interplay. Through extensive controlled experiments, we find that simultaneous gains for both task performance and calibration are difficult to achieve, and the problem of miscalibration exists across all learning methods in low-resource scenarios. To address this challenging trade-off between performance and calibration, we then investigate the potential of self-ensembling techniques applied at different modeling stages (e.g., variations of in-context examples or variations in prompts or different ensembling strategies). We justify the feasibility of self-ensembling on SFT in addition to ICL, to make the predictions more calibrated and have comparable or even better performance. Our work sheds light on which learning paradigm to choose and how to enhance both task performance and calibration of LLMs.

Reasoning models have demonstrated impressive performance on difficult tasks that traditional language models struggle at. However, many are plagued with the problem of overthinking--generating large amounts of unnecessary tokens which don't improve accuracy on a question. We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists, and evaluate how well calibrated a variety of reasoning models are in terms of efficiently allocating the optimal token count. We find that in general, reasoning models are poorly calibrated, particularly on easy problems. To evaluate calibration on easy questions we introduce DUMB500, a dataset of extremely easy math, reasoning, code, and task problems, and jointly evaluate reasoning model on these simple examples and extremely difficult examples from existing frontier benchmarks on the same task domain. Finally, we introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.

Metacognition is a critical component of intelligence that describes the ability to monitor and regulate one's own cognitive processes. Yet LLMs exhibit systemic deficiencies in key metacognitive faculties: they hallucinate with high confidence, fail to recognize knowledge boundaries, and misrepresent their internal uncertainty--undermining trustworthiness and reliability. Since monitoring task performance and adapting behavior accordingly are central to metacognition, we posit that models capable of accurately judging their own performance are better positioned to improve it. We operationalize this idea via two novel mechanisms: reinforcement learning with metacognitive feedback (RLMF), a paradigm to refine completion rankings during preference optimization based on the quality of a model's self-judgments of performance, and metacognitive data selection, which uses similar self-judgments to identify high-value training examples, outperforming naive active learning. We apply these innovations to the problem of faithful calibration (FC), a task that is itself fundamentally metacognitive: the goal is to align expressed with intrinsic uncertainty, difficult even for frontier LLMs. We adopt a two-stage, decoupled approach, first using these methods to calibrate the faithfulness of models' self-reported confidence scores, then mapping to natural, context-adaptable linguistic uncertainty via targeted output editing. Extensive experiments show RLMF achieves generalizable, state-of-the-art FC on diverse tasks while preserving accuracy. Further, RLMF surpasses standard RL by up to 63% while enhancing models' ability to assess and express their own capability limits. This positions RLMF as a promising paradigm to enhance LLM metacognition toward improved abilities and alignment, and suggests metacognitive performance as an effective RL signal to overcome limits of prior intrinsic feedback methods.

Computer-use agents must solve long-horizon tasks through repeated interaction with partially observable, multimodal desktop environments. Although imitation learning and offline trajectory refinement provide strong priors, static traces cannot cover the causal feedback loop of real computer use: each action changes the screen state, future action space, and recovery options. EvoCUA-1.5 extends self-evolving computer-use agents from offline experience learning to online reinforcement learning, where policies interact with executable sandbox environments and improve from verifiable task outcomes. Online RL in this setting requires more than directly reusing single-turn language-RL recipes. Multi-turn interaction introduces context-managed observations, sparse terminal rewards, variable-length trajectories, and slow environment feedback. EvoCUA-1.5 addresses these challenges with Step-Level Policy Optimization (STEPO), which preserves trajectory-level advantage balance after decomposition into step-level samples; policy-aware filtering and pass-rate calibration over verifiable synthesized tasks; Dynamic Tri-Adaptive Curriculum (DTAC), which combines learnable tasks, difficult positive replay, and controlled infeasible-task exposure; and a fully asynchronous RL infrastructure with staleness control and mini-group batching. Experiments show that these components improve training stability and downstream performance. EvoCUA-1.5 achieves 63.2\% success on OSWorld-Verified, outperforming comparable 32B/35B-scale open-weight baselines and even approaching models with significantly larger parameter counts. Overall, EvoCUA-1.5 provides a practical framework for scaling online RL in multi-turn computer-use agents.

Accurate estimation of item (question or task) difficulty is critical for educational assessment but suffers from the cold start problem. While Large Language Models demonstrate superhuman problem-solving capabilities, it remains an open question whether they can perceive the cognitive struggles of human learners. In this work, we present a large-scale empirical analysis of Human-AI Difficulty Alignment for over 20 models across diverse domains such as medical knowledge and mathematical reasoning. Our findings reveal a systematic misalignment where scaling up model size is not reliably helpful; instead of aligning with humans, models converge toward a shared machine consensus. We observe that high performance often impedes accurate difficulty estimation, as models struggle to simulate the capability limitations of students even when being explicitly prompted to adopt specific proficiency levels. Furthermore, we identify a critical lack of introspection, as models fail to predict their own limitations. These results suggest that general problem-solving capability does not imply an understanding of human cognitive struggles, highlighting the challenge of using current models for automated difficulty prediction.

We consider the issue of calibration in large language models (LLM). Recent studies have found that common interventions such as instruction tuning often result in poorly calibrated LLMs. Although calibration is well-explored in traditional applications, calibrating LLMs is uniquely challenging. These challenges stem as much from the severe computational requirements of LLMs as from their versatility, which allows them to be applied to diverse tasks. Addressing these challenges, we propose THERMOMETER, a calibration approach tailored to LLMs. THERMOMETER learns an auxiliary model, given data from multiple tasks, for calibrating a LLM. It is computationally efficient, preserves the accuracy of the LLM, and produces better-calibrated responses for new tasks. Extensive empirical evaluations across various benchmarks demonstrate the effectiveness of the proposed method.

We introduce NOLLI, a procedurally generated English-Korean puzzle benchmark designed to diagnose where Korean performance gaps arise. It comprises 15 puzzle types (25 tasks; 7,500 items), with every instance seed-regenerable, verified to have a unique solution, and scored deterministically. Rather than equating harder with bigger, we calibrate difficulty behaviorally, tuning each generator until a fixed reference model lands in target accuracy bands. Its three-level design combines matched direct translations, script adaptations over Hangul jamo (sub-syllabic letters), and Korean-only tasks grounded in Korean culture or orthography. We evaluate 15 frontier, open-weight, and Korean-developed models; among the 12 above a 3% overall-accuracy floor, matched English-Korean accuracy is statistically equivalent within a +/- 10 pp margin (TOST), suggesting little cost from presentation language alone. Writing-system-intensive tasks show sharper gaps: Korean Cipher falls behind English by up to 68.7 pp, whereas Cryptarithmetic over the same jamo shows no systematic penalty, and Jamo Composition accuracy predicts Korean Cipher accuracy. These contrasts are diagnostic rather than causal, consistent with difficulty in multi-step sub-syllabic execution. Korean-only tasks separate rule-application deficits, which vary in sign, from a Kinship deficit positive in all 12. Finally, a salient size measure fails to grow from Easy to Hard in 7 of 15 types, making structural size an unreliable proxy for empirical difficulty.

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is mixed regarding whether training on easier or harder data leads to better results, and whether those gains come on easier or harder test data. We address this question by conducting a systematic evaluation of LLMs' generalization across models, datasets, and fine-grained groups of example difficulty. We rank examples in six datasets using the outputs of thousands of different LLMs and Item Response Theory (IRT), a well-established difficulty metric in educational testing. Unlike prior work, our difficulty ratings are therefore determined solely by the abilities of many different LLMs, excluding human opinions of difficulty. With a more objective, larger-scale, and finer-grained analysis, we show that cross-difficulty generalization is often limited; training on either easy or hard data cannot achieve consistent improvements across the full range of difficulties. These results show the importance of having a range of difficulties in both training and evaluation data for LLMs, and that taking shortcuts with respect to difficulty is risky.

A wide body of evidence shows that human language processing difficulty is predicted by the information-theoretic measure surprisal, a word's negative log probability in context. However, it is still unclear how to best estimate these probabilities needed for predicting human processing difficulty -- while a long-standing belief held that models with lower perplexity would provide more accurate estimates of word predictability, and therefore lead to better reading time predictions, recent work has shown that for very large models, psycholinguistic predictive power decreases. One reason could be that language models might be more confident of their predictions than humans, because they have had exposure to several magnitudes more data. In this paper, we test what effect temperature-scaling of large language model (LLM) predictions has on surprisal estimates and their predictive power of reading times of English texts. Firstly, we show that calibration of large language models typically improves with model size, i.e. poorer calibration cannot account for poorer fit to reading times. Secondly, we find that temperature-scaling probabilities lead to a systematically better fit to reading times (up to 89% improvement in delta log likelihood), across several reading time corpora. Finally, we show that this improvement in fit is chiefly driven by words that are composed of multiple subword tokens.

We study whether language models can evaluate the validity of their own claims and predict which questions they will be able to answer correctly. We first show that larger models are well-calibrated on diverse multiple choice and true/false questions when they are provided in the right format. Thus we can approach self-evaluation on open-ended sampling tasks by asking models to first propose answers, and then to evaluate the probability "P(True)" that their answers are correct. We find encouraging performance, calibration, and scaling for P(True) on a diverse array of tasks. Performance at self-evaluation further improves when we allow models to consider many of their own samples before predicting the validity of one specific possibility. Next, we investigate whether models can be trained to predict "P(IK)", the probability that "I know" the answer to a question, without reference to any particular proposed answer. Models perform well at predicting P(IK) and partially generalize across tasks, though they struggle with calibration of P(IK) on new tasks. The predicted P(IK) probabilities also increase appropriately in the presence of relevant source materials in the context, and in the presence of hints towards the solution of mathematical word problems. We hope these observations lay the groundwork for training more honest models, and for investigating how honesty generalizes to cases where models are trained on objectives other than the imitation of human writing.

Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns the latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR's exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.

Predicting human item difficulty is central to educational assessment, where reliable estimates support fairness and effective test construction. Existing methods often depend on costly human calibration or item-level textual representations, providing limited evidence about the cognitive processes that make items difficult. We argue that difficulty should be viewed not only as a property of item text, but also as an observable consequence of the problem-solving burden an item induces. Large Reasoning Models (LRMs) offer scalable process evidence through reasoning traces, but such evidence must be structured to support interpretable modeling. To this end, we introduce Epi2Diff (Episode to Difficulty), a framework that maps LRM reasoning traces into cognitively grounded episode sequences. These episodes group trace segments into functional problem-solving states, enabling difficulty to be modeled through reasoning scale, effort allocation, and state transitions. Epi2Diff extracts compact episode-dynamic features and combines them with semantic item representations for human difficulty prediction. Experiments on four real-world human difficulty datasets show that Epi2Diff consistently outperforms strong baselines, including fine-tuned small language models, LLM in-context learning, and supervised LLM adaptation. On SAT-derived classification benchmarks, Epi2Diff achieves an 8.1% average relative gain over supervised LLM fine-tuning baselines. Further analyses show that harder items induce more effortful, iterative, and implementation-centered episode dynamics, rather than merely longer responses. These results demonstrate that cognitive episodes in LRM reasoning traces provide a predictive and interpretable process representation for human item difficulty, offering a new lens for educational measurement with reasoning models.

Large language models (LLMs) have the potential to boost human productivity by speeding up task completion -- provided users know when to offload cognitive work to them. But we do not know if users are well-calibrated in estimating these potential time savings. We conducted a preregistered large-scale behavioral study (N = 1237) to characterize mismatches between expectations and reality, with a focus on simple cognitive tasks. While actual completion times between independent completion and AI-assisted completion did not differ, participants predicted AI to be significantly faster. The same bias was not observed when imagining help from another human participant. We identify a speedup illusion where people have accurate forecasts of independent completion times but significantly underestimate AI-assisted times. Additionally, time and effort dissociate: participants reported lower subjective effort with AI despite equivalent completion times. This suggests that completion time itself is not sufficient to characterize efficiency gains.

While generalization over tasks from easy to hard is crucial to profile language models (LLMs), the datasets with fine-grained difficulty annotations for each problem across a broad range of complexity are still blank. Aiming to address this limitation, we present Easy2Hard-Bench, a consistently formatted collection of 6 benchmark datasets spanning various domains, such as mathematics and programming problems, chess puzzles, and reasoning questions. Each problem within these datasets is annotated with numerical difficulty scores. To systematically estimate problem difficulties, we collect abundant performance data on attempts to each problem by humans in the real world or LLMs on the prominent leaderboard. Leveraging the rich performance data, we apply well-established difficulty ranking systems, such as Item Response Theory (IRT) and Glicko-2 models, to uniformly assign numerical difficulty scores to problems. Moreover, datasets in Easy2Hard-Bench distinguish themselves from previous collections by a higher proportion of challenging problems. Through extensive experiments with six state-of-the-art LLMs, we provide a comprehensive analysis of their performance and generalization capabilities across varying levels of difficulty, with the aim of inspiring future research in LLM generalization. The datasets are available at https://huggingface.co/datasets/furonghuang-lab/Easy2Hard-Bench.

Large language models must satisfy hard orthographic constraints during controlled text generation, yet systematic cross-architecture evaluation remains limited. We evaluate 28 configurations spanning three model families (Qwen3, Claude Haiku-4.5, GPT-5-mini) on 58 word puzzles requiring character-level constraint satisfaction. Architectural differences produce substantially larger performance gaps (2.0-2.2x, F1=0.761 vs. 0.343) than parameter scaling within families (83% gain from eightfold scaling), suggesting that constraint satisfaction may require specialized architectural features or training objectives beyond standard language model scaling. Thinking budget sensitivity proves heterogeneous: high-capacity models show strong returns (+0.102 to +0.136 F1), while mid-sized variants saturate or degrade. These patterns are inconsistent with uniform compute benefits. Using difficulty ratings from 10,000 human solvers per puzzle, we establish modest but consistent calibration (r=0.24-0.38) across all families, yet identify systematic failures on common words with unusual orthography ("data", "poop", "loll": 86-95% human success, 89-96% model miss rate). These failures reveal over-reliance on distributional plausibility that penalizes orthographically atypical but constraint-valid patterns, suggesting architectural innovations may be required beyond simply scaling parameters or computational budgets.

Overconfidence and underconfidence in machine learning classifiers is measured by calibration: the degree to which the probabilities predicted for each class match the accuracy of the classifier on that prediction. How one measures calibration remains a challenge: expected calibration error, the most popular metric, has numerous flaws which we outline, and there is no clear empirical understanding of how its choices affect conclusions in practice, and what recommendations there are to counteract its flaws. In this paper, we perform a comprehensive empirical study of choices in calibration measures including measuring all probabilities rather than just the maximum prediction, thresholding probability values, class conditionality, number of bins, bins that are adaptive to the datapoint density, and the norm used to compare accuracies to confidences. To analyze the sensitivity of calibration measures, we study the impact of optimizing directly for each variant with recalibration techniques. Across MNIST, Fashion MNIST, CIFAR-10/100, and ImageNet, we find that conclusions on the rank ordering of recalibration methods is drastically impacted by the choice of calibration measure. We find that conditioning on the class leads to more effective calibration evaluations, and that using the L2 norm rather than the L1 norm improves both optimization for calibration metrics and the rank correlation measuring metric consistency. Adaptive binning schemes lead to more stablity of metric rank ordering when the number of bins vary, and is also recommended. We open source a library for the use of our calibration measures.

Terminal-agent benchmarks have become a primary signal for measuring the coding and system-administration capabilities of large language models. As the market for evaluation environments grows, so does the pressure to ship tasks quickly, often without thorough adversarial review of the verification logic. This paper is a guideline for writing good benchmark tasks, drawn from over a year of contributing to and reviewing tasks for Terminal Bench. Most people write benchmark tasks the way they write prompts. They shouldn't. A prompt is designed to help the agent succeed; a benchmark is designed to find out if it can. We argue that good tasks are adversarial, difficult, and legible, and that a large class of common failure modes -- AI-generated instructions, over-prescriptive specifications, clerical difficulty, oracle solutions that assume hidden knowledge, tests that validate the wrong things, and reward-hackable environments -- are predictable consequences of treating task authoring as prompt authoring. We catalog these failure modes, argue that real difficulty is conceptual rather than environmental, and discuss recent empirical evidence that over 15% of tasks in popular terminal-agent benchmarks are reward-hackable. We hope this serves as a useful reference for benchmark maintainers, task contributors, and researchers using benchmark scores as evidence.

Large Language Models (LLMs) changed the way we design and interact with software systems. Their ability to process and extract information from text has drastically improved productivity in a number of routine tasks. Developers that want to include these models in their software stack, however, face a dreadful challenge: debugging LLMs' inconsistent behavior across minor variations of the prompt. We therefore introduce two metrics for classification tasks, namely sensitivity and consistency, which are complementary to task performance. First, sensitivity measures changes of predictions across rephrasings of the prompt, and does not require access to ground truth labels. Instead, consistency measures how predictions vary across rephrasings for elements of the same class. We perform an empirical comparison of these metrics on text classification tasks, using them as guideline for understanding failure modes of the LLM. Our hope is that sensitivity and consistency will be helpful to guide prompt engineering and obtain LLMs that balance robustness with performance.

As the field progresses toward Artificial General Intelligence (AGI), there is a pressing need for more comprehensive and insightful evaluation frameworks that go beyond aggregate performance metrics. This paper introduces a unified rating system that jointly models the difficulty of individual test cases and the competency of AI models (or humans) across vision, language, and action domains. Unlike existing metrics that focus solely on models, our approach allows for fine-grained, difficulty-aware evaluations through competitive interactions between models and tasks, capturing both the long-tail distribution of real-world challenges and the competency gap between current models and full task mastery. We validate the generalizability and robustness of our system through extensive experiments on multiple established datasets and models across distinct AGI domains. The resulting rating distributions offer novel perspectives and interpretable insights into task difficulty, model progression, and the outstanding challenges that remain on the path to achieving full AGI task mastery.

Mixed-precision quantization (MPQ) has become a key technique for deploying large language models under stringent memory and compute constraints. We first identify a phenomenon that we term the Perplexity Illusion: layers ranked as important by perplexity-based sensitivity show little rank correlation with those that are most influential for complex reasoning performance, with Kendall τapprox 0 in our analysis. We further reveal an Alignment-Diversity Tradeoff: using only target-task calibration data can degrade post-quantization performance, whereas incorporating general-domain data stabilizes sensitivity estimation and improves robustness across tasks. Based on these observations, we propose TASA (Task-Aware Sensitivity Analysis), a two-level framework that jointly optimizes calibration-data composition and mixed-precision bit allocation. Specifically, TASA searches for a calibration-data mixture using a training-free gradient-trace alignment criterion, and then aggregates perplexity and reasoning-oriented sensitivity signals to guide both inter-layer and intra-layer bit allocation. Experiments on LLaMA-3-8B and Qwen2.5-7B reveal a precision inversion: appropriately allocated 3.5-bit models can match or surpass less task-aware 4-bit baselines. At an average precision of 3.5 bits, TASA matches or outperforms several competitive 4-bit uniform baselines in aggregate accuracy, and improves over the strongest W3 baseline on GSM8K by more than 20 absolute points on LLaMA-3-8B. These results show that calibration-data composition substantially affects task-sensitive quantization, a factor underexplored in prior work.

Single-prompt accuracy is the dominant way to benchmark language models, but it can miss reliability failures that matter. We evaluate a 15-model open-weight corpus, with the main reliability analyses focused on 10 instruct models across five classification and reasoning benchmarks under five prompt variants each, measuring accuracy, token-probability calibration, verbal-confidence calibration, verbal parse rate, and prompt-perturbation spread for every (model x dataset x variant) cell. We find three broad results. First, evaluation design can materially change the conclusion. Switching Expected Calibration Error (ECE) token from a raw to a label-set-normalised definition changes per-cell calibration by a mean absolute 0.149. More strikingly, pairing a chain-of-thought prompt with a first-character evaluator on ARC-Challenge reduces apparent accuracy by 72-88% across all five primary models; two independent repair procedures recover 93.8% and 102.7% of the lost performance, indicating an evaluator-side rather than model-side failure. Second, confidence signals are fragile. On MMLU-Pro, every primary model verbally reports confidence substantially above both its accuracy and its token-probability confidence on the same rows, and verbal parse rate can collapse for a single model on a single prompt variant. Third, prompt robustness does not track parameter count reliably. Across 10 instruct models, the correlation between model size and prompt-perturbation spread ranges from -0.244 to 0.474 across benchmarks. Taken together, these results show that reliability conclusions for small language models depend not only on the model being evaluated, but also on the evaluation pipeline used to measure it. We argue that calibration definitions, evaluator logic, verbal parseability, and prompt robustness should be reported explicitly when making reliability claims.

Large language models exhibit a puzzling inconsistency: they solve complex problems yet frequently fail on seemingly simpler ones. We investigate whether LLMs internally encode problem difficulty in a way that aligns with human judgment, and whether this representation tracks generalization during reinforcement learning post-training. We train linear probes across layers and token positions on 60 models, evaluating on mathematical and coding subsets of Easy2HardBench. We find that human-labeled difficulty is strongly linearly decodable (AMC: rho approx 0.88) and exhibits clear model-size scaling, whereas LLM-derived difficulty is substantially weaker and scales poorly. Steering along the difficulty direction reveals that pushing models toward "easier" representations reduces hallucination and improves accuracy. During GRPO training on Qwen2.5-Math-1.5B, the human-difficulty probe strengthens and positively correlates with test accuracy across training steps, while the LLM-difficulty probe degrades and negatively correlates with performance. These results suggest that human annotations provide a stable difficulty signal that RL amplifies, while automated difficulty estimates derived from model performance become misaligned precisely as models improve. We release probe code and evaluation scripts to facilitate replication.

Prompting and in-context learning (ICL) have become efficient learning paradigms for large language models (LLMs). However, LLMs suffer from prompt brittleness and various bias factors in the prompt, including but not limited to the formatting, the choice verbalizers, and the ICL examples. To address this problem that results in unexpected performance degradation, calibration methods have been developed to mitigate the effects of these biases while recovering LLM performance. In this work, we first conduct a systematic analysis of the existing calibration methods, where we both provide a unified view and reveal the failure cases. Inspired by these analyses, we propose Batch Calibration (BC), a simple yet intuitive method that controls the contextual bias from the batched input, unifies various prior approaches, and effectively addresses the aforementioned issues. BC is zero-shot, inference-only, and incurs negligible additional costs. In the few-shot setup, we further extend BC to allow it to learn the contextual bias from labeled data. We validate the effectiveness of BC with PaLM 2-(S, M, L) and CLIP models and demonstrate state-of-the-art performance over previous calibration baselines across more than 10 natural language understanding and image classification tasks.

Machine translation quality has steadily improved over the years, achieving near-perfect translations in recent benchmarks. These high-quality outputs make it difficult to distinguish between state-of-the-art models and to identify areas for future improvement. In this context, automatically identifying texts where machine translation systems struggle holds promise for developing more discriminative evaluations and guiding future research. In this work, we address this gap by formalizing the task of translation difficulty estimation, defining a text's difficulty based on the expected quality of its translations. We introduce a new metric to evaluate difficulty estimators and use it to assess both baselines and novel approaches. Finally, we demonstrate the practical utility of difficulty estimators by using them to construct more challenging benchmarks for machine translation. Our results show that dedicated models outperform both heuristic-based methods and LLM-as-a-judge approaches, with Sentinel-src achieving the best performance. Thus, we release two improved models for difficulty estimation, Sentinel-src-24 and Sentinel-src-25, which can be used to scan large collections of texts and select those most likely to challenge contemporary machine translation systems.

Transferring learned patterns from pretrained neural language models has been shown to significantly improve effectiveness across a variety of language-based tasks, meanwhile further tuning on intermediate tasks has been demonstrated to provide additional performance benefits, provided the intermediate task is sufficiently related to the target task. However, how to identify related tasks is an open problem, and brute-force searching effective task combinations is prohibitively expensive. Hence, the question arises, are we able to improve the effectiveness and efficiency of tasks with no training examples through selective fine-tuning? In this paper, we explore statistical measures that approximate the divergence between domain representations as a means to estimate whether tuning using one task pair will exhibit performance benefits over tuning another. This estimation can then be used to reduce the number of task pairs that need to be tested by eliminating pairs that are unlikely to provide benefits. Through experimentation over 58 tasks and over 6,600 task pair combinations, we demonstrate that statistical measures can distinguish effective task pairs, and the resulting estimates can reduce end-to-end runtime by up to 40%.

As LLM benchmarks saturate, the evaluation community has pursued two strategies to increase difficulty: escalating knowledge demands (GPQA, HLE) or removing knowledge entirely in favor of abstract reasoning (ARC-AGI). The first conflates memorization with capability; the second divorces reasoning from the practical contexts in which it matters. We take a different approach. The Grounded Integration Measure (GIM) is a benchmark of 820 original problems (615 public, 205 private) where difficulty comes from integration; individual problems require coordinating multiple cognitive operations (constraint satisfaction, state tracking, epistemic vigilance, audience calibration) over broadly accessible knowledge, so that reasoning stays grounded in realistic tasks without being gated on specialized expertise. Each problem is an original expert-authored composition, majority with rubric-decomposed scoring (median 6 independently judged criteria). A balanced public--private split provides built-in contamination diagnostic. We calibrate a continuous response 2-parameter logistic (2PL) IRT model over >200k prompt-response pairs across 28 models, producing robust ability estimates that correctly order test-configurations even when raw accuracy is distorted by errors or missing data, addressing a common challenge in benchmark reporting. Using this framework, we present a comprehensive leaderboard spanning 22 models and 47 test-configurations (unique model, thinking-level pairs), and conduct what is to our knowledge the most extensive published study of how test-time compute trades off against model capability on a fixed benchmark: 11 models swept across 35 test-configurations. We observe that within-family configuration choices, such as thinking budget and quantization, matter as much as model selection. We release the evaluation framework, calibrated IRT parameters, and all public problems.

We show that a GPT-3 model can learn to express uncertainty about its own answers in natural language -- without use of model logits. When given a question, the model generates both an answer and a level of confidence (e.g. "90% confidence" or "high confidence"). These levels map to probabilities that are well calibrated. The model also remains moderately calibrated under distribution shift, and is sensitive to uncertainty in its own answers, rather than imitating human examples. To our knowledge, this is the first time a model has been shown to express calibrated uncertainty about its own answers in natural language. For testing calibration, we introduce the CalibratedMath suite of tasks. We compare the calibration of uncertainty expressed in words ("verbalized probability") to uncertainty extracted from model logits. Both kinds of uncertainty are capable of generalizing calibration under distribution shift. We also provide evidence that GPT-3's ability to generalize calibration depends on pre-trained latent representations that correlate with epistemic uncertainty over its answers.

Large Language Models (LLMs) have demonstrated great capabilities in solving a wide range of tasks in a resource-efficient manner through prompting, which does not require task-specific training, but suffers from performance fluctuation when there are multiple prompt candidates. Previous works have introduced gradient-free probability-based prompt selection methods that aim to choose the optimal prompt among the candidates for a given task but fail to provide a comprehensive and fair comparison between each other. In this paper, we propose a unified framework to interpret and evaluate the existing probability-based prompt selection methods by performing extensive experiments on 13 common NLP tasks. We find that all existing methods can be unified into some variant of the method that maximizes the mutual information between the input and the corresponding model output (denoted as MI). Using the finding, we develop several variants of MI and increases the effectiveness of the best prompt selection method from 87.79% to 94.98%, measured as the ratio of the performance of the selected prompt to that of the optimal oracle prompt. Furthermore, we propose a novel calibration method called Calibration by Marginalization (CBM) that is orthogonal to existing methods and helps increase the prompt selection effectiveness of the best method by 99.44%. The code and datasets used in our work will be released at https://github.com/soheeyang/unified-prompt-selection.

The vulnerability of machine learning models to membership inference attacks has received much attention in recent years. However, existing attacks mostly remain impractical due to having high false positive rates, where non-member samples are often erroneously predicted as members. This type of error makes the predicted membership signal unreliable, especially since most samples are non-members in real world applications. In this work, we argue that membership inference attacks can benefit drastically from difficulty calibration, where an attack's predicted membership score is adjusted to the difficulty of correctly classifying the target sample. We show that difficulty calibration can significantly reduce the false positive rate of a variety of existing attacks without a loss in accuracy.

Text-to-Image (T2I) models have recently achieved impressive visual fidelity, yet their evaluation remains constrained by benchmarks that are often difficult to interpret and insufficiently diagnostic. Existing skill-based evaluations tend to overlook critical failure modes that strongly impact usability but fall outside standard taxonomies, such as global incoherence arising from missing parts or physically implausible configurations (e.g., floating objects). In addition, prompt difficulty is typically controlled along a single dimension; either prompt length or the number of elements to generate. To address these limitations, we introduce Imag-Eval, a controlled benchmark designed to assess how T2I models ground compositional natural-language instructions into visual outputs. Unlike prior work that conflates surface linguistic complexity with compositional difficulty, Imag-Eval explicitly seeks to disentangles these factors by independently varying both the number of instances and the combination of constraints (rules), while avoiding error propagation. This design enables fine-grained and interpretable analysis of where cross-modal instruction following fails. Our benchmark comprises 1,140 prompts and 8,842 combined rules, and we evaluate it on several state-of-the-art models. Complementing this analysis with an additional study of over 2,000 prompts from a concurrent benchmark, our results suggest that, for structured skills, compositional difficulty is primarily governed by the number of grounded rules and their binding to instances,, rather than by prompt length alone.

Foundation models have revolutionized general-purpose problem-solving, offering rapid task adaptation through pretraining, meta-training, and finetuning. Recent crucial advances in these paradigms reveal the importance of challenging task prioritized sampling to enhance adaptation robustness under distribution shifts. However, ranking task difficulties over iteration as a preliminary step typically requires exhaustive task evaluation, which is practically unaffordable in computation and data-annotation. This study provides a novel perspective to illuminate the possibility of leveraging the dual importance of adaptation robustness and learning efficiency, particularly in scenarios where task evaluation is risky or costly, such as iterative agent-environment interactions for robotic policy evaluation or computationally intensive inference steps for finetuning foundation models. Firstly, we introduce Model Predictive Task Sampling (MPTS), a framework that bridges the task space and adaptation risk landscape, providing a theoretical foundation for robust active task sampling. MPTS employs a generative model to characterize the episodic optimization process and predicts task-specific adaptation risk via posterior inference. The resulting risk learner amortizes the costly evaluation of task adaptation performance and provably approximates task difficulty rankings. MPTS seamlessly integrates into zero-shot, few-shot, and supervised finetuning settings. Empirically, we conduct extensive experiments in pattern recognition using foundation models and sequential decision-making. Our results demonstrate that MPTS significantly enhances adaptation robustness for tail or out-of-distribution (OOD) tasks and improves learning efficiency compared to state-of-the-art (SOTA) methods. The code is available at the project site https://github.com/thu-rllab/MPTS.

Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.

Depth pruning improves the inference efficiency of large language models by removing Transformer blocks. Prior work typically treats layer redundancy as an inherent structural property of pretrained networks, emphasizing importance criteria and search algorithms to identify removable layers. In this study, we empirically investigate depth pruning from a functional perspective. Evaluating representative LLM families across diverse calibration configurations and multiple search algorithms, we show that different configurations produce different pruning patterns. Furthermore, under a fixed calibration configuration, complex search algorithms yield marginal performance improvements over simple one-shot methods, converging to similar pruned subsets. Overall, our results suggest that the calibration configuration plays a substantially larger role than the choice of search algorithm in shaping pruning patterns and calibration perplexity, while contributing comparably to variance in downstream reasoning accuracy. This indicates that future pruning efforts may benefit from prioritizing the calibration configuration over search complexity.

GPT-3 can perform numerous tasks when provided a natural language prompt that contains a few training examples. We show that this type of few-shot learning can be unstable: the choice of prompt format, training examples, and even the order of the training examples can cause accuracy to vary from near chance to near state-of-the-art. We demonstrate that this instability arises from the bias of language models towards predicting certain answers, e.g., those that are placed near the end of the prompt or are common in the pre-training data. To mitigate this, we first estimate the model's bias towards each answer by asking for its prediction when given the training prompt and a content-free test input such as "N/A". We then fit calibration parameters that cause the prediction for this input to be uniform across answers. On a diverse set of tasks, this contextual calibration procedure substantially improves GPT-3 and GPT-2's average accuracy (up to 30.0% absolute) and reduces variance across different choices of the prompt.

In this paper, we uncover a systematic bias in the evaluation paradigm of adopting large language models~(LLMs), e.g., GPT-4, as a referee to score and compare the quality of responses generated by candidate models. We find that the quality ranking of candidate responses can be easily hacked by simply altering their order of appearance in the context. This manipulation allows us to skew the evaluation result, making one model appear considerably superior to the other, e.g., Vicuna-13B could beat ChatGPT on 66 over 80 tested queries with ChatGPT as an evaluator. To address this issue, we propose a calibration framework with three simple yet effective strategies: 1) Multiple Evidence Calibration, which requires the evaluator model to generate multiple evaluation evidence before assigning ratings; 2) Balanced Position Calibration, which aggregates results across various orders to determine the final score; 3) Human-in-the-Loop Calibration, which introduces a balanced position diversity entropy to measure the difficulty of each example and seeks human assistance when needed. We also manually annotate the "win/tie/lose" outcomes of responses from ChatGPT and Vicuna-13B in the Vicuna Benchmark's question prompt, and extensive experiments demonstrate that our approach successfully mitigates evaluation bias, resulting in closer alignment with human judgments. We release our code and human annotation at https://github.com/i-Eval/FairEval to facilitate future research.

We develop task scaling laws and model ladders to predict the individual task performance of pretrained language models (LMs) in the overtrained setting. Standard power laws for language modeling loss cannot accurately model task performance. Therefore, we leverage a two-step prediction approach: first use model and data size to predict a task-specific loss, and then use this task loss to predict task performance. We train a set of small-scale "ladder" models, collect data points to fit the parameterized functions of the two prediction steps, and make predictions for two target models: a 7B model trained to 4T tokens and a 13B model trained to 5T tokens. Training the ladder models only costs 1% of the compute used for the target models. On four multiple-choice tasks written in ranked classification format, we can predict the accuracy of both target models within 2 points of absolute error. We have higher prediction error on four other tasks (average absolute error 6.9) and find that these are often tasks with higher variance in task metrics. We also find that using less compute to train fewer ladder models tends to deteriorate predictions. Finally, we empirically show that our design choices and the two-step approach lead to superior performance in establishing scaling laws.

In this work, we investigate how explicitly modeling problem's difficulty prior information shapes the effectiveness of reinforcement learning based fine-tuning for multimodal reasoning. Our exploration mainly comprises of following three perspective: First, through offline data curation, we analyze the U-shaped difficulty distribution of two given datasets using the base model by multi-round sampling, and then filter out prompts that are either too simple or extremely difficult to provide meaningful gradients and perform subsequent two-stage training. Second, we implement an online advantage differentiation, computing group-wise empirical accuracy as a difficulty proxy to adaptively reweight advantages estimation, providing stronger learning signals for more challenging problems. Finally, we introduce difficulty hints as explicit prompts for more complex samples in the second training stage, encouraging the model to calibrate its reasoning depth and perform reflective validation checks. Our comprehensive approach demonstrates significant performances across various multi-modal mathematical reasoning benchmarks with only 2K+0.6K two-stage training data.

A critical component in the trustworthiness of LLMs is reliable uncertainty communication, yet LLMs often use assertive language when conveying false claims, leading to over-reliance and eroded trust. We present the first systematic study of faithful confidence calibration of LLMs, benchmarking models' ability to use linguistic expressions of uncertainty that faithfully reflect their intrinsic uncertainty, across a comprehensive array of models, datasets, and prompting strategies. Our results demonstrate that LLMs largely fail at this task, and that existing interventions are insufficient: standard prompt approaches provide only marginal gains, and existing, factuality-based calibration techniques can even harm faithful calibration. To address this critical gap, we introduce MetaFaith, a novel prompt-based calibration approach inspired by human metacognition. We show that MetaFaith robustly improves faithful calibration across diverse models and task domains, enabling up to 61% improvement in faithfulness and achieving an 83% win rate over original generations as judged by humans.

Estimating the difficulty of multiple-choice questions would be great help for educators who must spend substantial time creating and piloting stimuli for their tests, and for learners who want to practice. Supervised approaches to difficulty estimation have yielded to date mixed results. In this contribution we leverage an aspect of generative large models which might be seen as a weakness when answering questions, namely their uncertainty, and exploit it towards exploring correlations between two different metrics of uncertainty, and the actual student response distribution. While we observe some present but weak correlations, we also discover that the models' behaviour is different in the case of correct vs wrong answers, and that correlations differ substantially according to the different question types which are included in our fine-grained, previously unused dataset of 451 questions from a Biopsychology course. In discussing our findings, we also suggest potential avenues to further leverage model uncertainty as an additional proxy for item difficulty.

Large language models (LLMs) have become increasingly capable of following instructions and complex reasoning, making prompting a flexible interface for adapting models without parameter updates. Yet prompt design remains labor-intensive and highly sensitive to formatting, phrasing, and instruction order, motivating automated prompt optimization methods that reduce manual effort while preserving inference-time flexibility. However, existing methods often search over prompt candidates or use fixed critique-refine pipelines driven by individual examples or small batches, limiting their ability to capture systematic error patterns and make targeted edits grounded in failure history. We propose Reflective Prompt Tuning (RPT), a framework that uses LLM function calling to simulate the iterative workflow of human prompt engineers. An LLM optimizer calls a diagnostic function that evaluates the target model over an entire optimization set, summarizes recurring failure modes, and returns a structured diagnostic report. The optimizer uses this report, together with an accumulated memory of prior reports, to revise the prompt for the next iteration. RPT further supports confidence-aware optimization by using calibration signals in diagnostic feedback and final prompt selection. Across three reasoning tasks, RPT improves over initial prompts by up to 12.9 points, remains competitive with state of the art, and improves confidence calibration. Our analyses show that RPT is especially effective on multi-hop and mathematical reasoning, producing targeted prompt revisions that align with diagnosed failure patterns and lead to gains in task performance and calibration.

In an educational setting, an estimate of the difficulty of multiple-choice questions (MCQs), a commonly used strategy to assess learning progress, constitutes very useful information for both teachers and students. Since human assessment is costly from multiple points of view, automatic approaches to MCQ item difficulty estimation are investigated, yielding however mixed success until now. Our approach to this problem takes a different angle from previous work: asking various Large Language Models to tackle the questions included in two different MCQ datasets, we leverage model uncertainty to estimate item difficulty. By using both model uncertainty features as well as textual features in a Random Forest regressor, we show that uncertainty features contribute substantially to difficulty prediction, where difficulty is inversely proportional to the number of students who can correctly answer a question. In addition to showing the value of our approach, we also observe that our model achieves state-of-the-art results on the BEA publicly available dataset.

Training terminal agents requires executable and verifiable tasks that are not merely solvable, but appropriately challenging for learning. Executable validation establishes feasibility, yet does not reveal how a task behaves relative to a given solver setting. In this paper, we present CalibForge, an autonomous terminal-task synthesis system that uses verified solver behavior to revise candidate tasks through adversarial solver calibration. Multi-solver calibration targets disagreement within a heterogeneous solver pool, whereas contrastive solver calibration targets a designated strong-pass/weak-fail relation; both operationalize a solver-relative learnable zone anchored in demonstrated solvability. Using CalibForge, we construct 5,431 calibrated terminal tasks. Our ablations show that both strategies yield more effective supervision than authoring and validation alone or ordinary single-solver feedback. Models trained on the full collection achieve 32.58% and 47.57% on Terminal-Bench 2.0. The largest improvements over the corresponding base model reach 24.71 percentage points on Terminal-Bench 2.0, 27.68 points on SWE-bench Pro, and 30.04 points on Doc2Repo. Together, these results support solver-relative learnability as a practical target for constructing effective and transferable agent training data.

Calibration, the alignment of predicted probabilities with true outcome frequencies, is essential for reliable decision-making. While extensively studied for classification and regression, calibration has not been formally addressed for probabilistic label ranking, where the goal is to predict a distribution over orderings of a label set. Naively treating rankings as classes ignores their structure and fails to capture important modalities such as pairwise and top-k predictions. We formalize calibration for label ranking and develop a hierarchy of notions covering full rankings, sub-rankings, and top-k rankings. We prove that full-rank calibration implies the others but not conversely, and sub-ranking and top-k calibration are incomparable. Empirically, we find popular label ranking models are often poorly calibrated, with substantial differences between sub-ranking and top-k metrics. Applying our framework to RLHF reward models, we find that calibration correlates strongly but not perfectly with benchmark accuracy, suggesting it captures a meaningful quality dimension beyond top-1 accuracy. These findings motivate future work on understanding the downstream effects of miscalibration and developing methods to correct it.

Agentic benchmarks increasingly rely on LLM-simulated users to scalably evaluate agent performance, yet the robustness, validity, and fairness of this approach remain unexamined. Through a user study with participants across the United States, India, Kenya, and Nigeria, we investigate whether LLM-simulated users serve as reliable proxies for real human users in evaluating agents on τ-Bench retail tasks. We find that user simulation lacks robustness, with agent success rates varying up to 9 percentage points across different user LLMs. Furthermore, evaluations using simulated users exhibit systematic miscalibration, underestimating agent performance on challenging tasks and overestimating it on moderately difficult ones. African American Vernacular English (AAVE) speakers experience consistently worse success rates and calibration errors than Standard American English (SAE) speakers, with disparities compounding significantly with age. We also find simulated users to be a differentially effective proxy for different populations, performing worst for AAVE and Indian English speakers. Additionally, simulated users introduce conversational artifacts and surface different failure patterns than human users. These findings demonstrate that current evaluation practices risk misrepresenting agent capabilities across diverse user populations and may obscure real-world deployment challenges.

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 442 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Various design settings for in-context learning (ICL), such as the choice and order of the in-context examples, can bias a model toward a particular prediction without being reflective of an understanding of the task. While many studies discuss these design choices, there have been few systematic investigations into categorizing them and mitigating their impact. In this work, we define a typology for three types of label biases in ICL for text classification: vanilla-label bias, context-label bias, and domain-label bias (which we conceptualize and detect for the first time). Our analysis demonstrates that prior label bias calibration methods fall short of addressing all three types of biases. Specifically, domain-label bias restricts LLMs to random-level performance on many tasks regardless of the choice of in-context examples. To mitigate the effect of these biases, we propose a simple bias calibration method that estimates a language model's label bias using random in-domain words from the task corpus. After controlling for this estimated bias when making predictions, our novel domain-context calibration significantly improves the ICL performance of GPT-J and GPT-3 on a wide range of tasks. The gain is substantial on tasks with large domain-label bias (up to 37% in Macro-F1). Furthermore, our results generalize to models with different scales, pretraining methods, and manually-designed task instructions, showing the prevalence of label biases in ICL.

The measure of a machine learning algorithm is the difficulty of the tasks it can perform, and sufficiently difficult tasks are critical drivers of strong machine learning models. However, quantifying the generalization difficulty of machine learning benchmarks has remained challenging. We propose what is to our knowledge the first model-agnostic measure of the inherent generalization difficulty of tasks. Our inductive bias complexity measure quantifies the total information required to generalize well on a task minus the information provided by the data. It does so by measuring the fractional volume occupied by hypotheses that generalize on a task given that they fit the training data. It scales exponentially with the intrinsic dimensionality of the space over which the model must generalize but only polynomially in resolution per dimension, showing that tasks which require generalizing over many dimensions are drastically more difficult than tasks involving more detail in fewer dimensions. Our measure can be applied to compute and compare supervised learning, reinforcement learning and meta-learning generalization difficulties against each other. We show that applied empirically, it formally quantifies intuitively expected trends, e.g. that in terms of required inductive bias, MNIST < CIFAR10 < Imagenet and fully observable Markov decision processes (MDPs) < partially observable MDPs. Further, we show that classification of complex images < few-shot meta-learning with simple images. Our measure provides a quantitative metric to guide the construction of more complex tasks requiring greater inductive bias, and thereby encourages the development of more sophisticated architectures and learning algorithms with more powerful generalization capabilities.

Large language models (LLMs) are increasingly used in decision-making contexts, but when they present answers without signaling low confidence, users may unknowingly act on erroneous outputs. Prior work shows that LLMs maintain internal uncertainty signals, yet their expressed confidence is often miscalibrated and poorly discriminates between correct and incorrect answers. We investigate whether supervised fine-tuning can improve models' ability to communicate uncertainty and whether such improvements generalize across tasks and domains. We fine-tune LLMs on datasets spanning general knowledge, mathematics, and open-ended trivia, and evaluate two metacognitive tasks: (1) single-question confidence estimation, where the model assigns a numeric certainty to its answer, and (2) pairwise confidence comparison, where the model selects which of two answers it is more likely to answer correctly. We assess generalization to unseen domains, including medical and legal reasoning. Results show that fine-tuning improves calibration (alignment between stated confidence and accuracy) and discrimination (higher confidence for correct vs. incorrect responses) within and across domains. However, gains are task-specific: training on single-question calibration does not transfer to pairwise comparison, and vice versa. Multitask fine-tuning yields broader gains, lowering calibration error and strengthening discrimination in out-of-domain evaluations. This suggests that uncertainty communication in LLMs is trainable but requires multitask training to generalize effectively.

Estimating the difficulty of input questions as perceived by large language models (LLMs) is essential for accurate performance evaluation and adaptive inference. Existing methods typically rely on repeated response sampling, auxiliary models, or fine-tuning the target model itself, which may incur substantial computational costs or compromise generality. In this paper, we propose a novel approach for difficulty estimation that leverages only the hidden representations produced by the target LLM. We model the token-level generation process as a Markov chain and define a value function to estimate the expected output quality given any hidden state. This allows for efficient and accurate difficulty estimation based solely on the initial hidden state, without generating any output tokens. Extensive experiments across both textual and multimodal tasks demonstrate that our method consistently outperforms existing baselines in difficulty estimation. Moreover, we apply our difficulty estimates to guide adaptive reasoning strategies, including Self-Consistency, Best-of-N, and Self-Refine, achieving higher inference efficiency with fewer generated tokens.

The ability of Large Language Models (LLMs) to use external tools unlocks powerful real-world interactions, making rigorous evaluation essential. However, current benchmarks primarily report final accuracy, revealing what models can do but obscuring the cognitive bottlenecks that define their true capability boundaries. To move from simple performance scoring to a diagnostic tool, we introduce a framework grounded in Cognitive Load Theory. Our framework deconstructs task complexity into two quantifiable components: Intrinsic Load, the inherent structural complexity of the solution path, formalized with a novel Tool Interaction Graph; and Extraneous Load, the difficulty arising from ambiguous task presentation. To enable controlled experiments, we construct ToolLoad-Bench, the first benchmark with parametrically adjustable cognitive load. Our evaluation reveals distinct performance cliffs as cognitive load increases, allowing us to precisely map each model's capability boundary. We validate that our framework's predictions are highly calibrated with empirical results, establishing a principled methodology for understanding an agent's limits and a practical foundation for building more efficient systems.

Large Language Models are known to capture real-world knowledge, allowing them to excel in many downstream tasks. Despite recent advances, these models are still prone to what are commonly known as hallucinations, causing them to emit unwanted and factually incorrect text. In this work, we propose a novel calibration method that can be used to combat hallucinations. We add a special [IDK] ("I don't know") token to the model's vocabulary and introduce an objective function that shifts probability mass to the [IDK] token for incorrect predictions. This approach allows the model to express uncertainty in its output explicitly. We evaluate our proposed method across multiple model architectures and factual downstream tasks. We find that models trained with our method are able to express uncertainty in places where they would previously make mistakes while suffering only a small loss of encoded knowledge. We further perform extensive ablation studies of multiple variations of our approach and provide a detailed analysis of the precision-recall tradeoff of our method.

With the increasing adoption of vision-language models (VLMs) in critical decision-making systems such as healthcare or autonomous driving, the calibration of their uncertainty estimates becomes paramount. Yet, this dimension has been largely underexplored in the VLM test-time prompt-tuning (TPT) literature, which has predominantly focused on improving their discriminative performance. Recent state-of-the-art advocates for enforcing full orthogonality over pairs of text prompt embeddings to enhance separability, and therefore calibration. Nevertheless, as we theoretically show in this work, the inherent gradients from fully orthogonal constraints will strongly push semantically related classes away, ultimately making the model overconfident. Based on our findings, we propose Semantic Orthogonal Calibration (SoC), a Huber-based regularizer that enforces smooth prototype separation while preserving semantic proximity, thereby improving calibration compared to prior orthogonality-based approaches. Across a comprehensive empirical validation, we demonstrate that SoC consistently improves calibration performance, while also maintaining competitive discriminative capabilities.

When using large language models (LLMs) in high-stakes applications, we need to know when we can trust their predictions. Some works argue that prompting high-performance LLMs is sufficient to produce calibrated uncertainties, while others introduce sampling methods that can be prohibitively expensive. In this work, we first argue that prompting on its own is insufficient to achieve good calibration and then show that fine-tuning on a small dataset of correct and incorrect answers can create an uncertainty estimate with good generalization and small computational overhead. We show that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance and tractable for large open-source models when using LoRA. We also investigate the mechanisms that enable reliable LLM uncertainty estimation, finding that many models can be used as general-purpose uncertainty estimators, applicable not just to their own uncertainties but also the uncertainty of other models. Lastly, we show that uncertainty estimates inform human use of LLMs in human-AI collaborative settings through a user study.

Professional agent tasks often depend on conventions that are absent from public corpora, yet benchmarks rarely control whether an agent has access to those conventions. We introduce a knowledge-gated task-construction protocol that separates a task instruction from a compact artefact containing private conventions, reference tables, and utility operators. Construction-time provenance, byte-identical task instructions across the provided- and withheld-artefact conditions, leak audits, and executable witnesses make dependence on the artefact explicit and testable. Across fifteen calibration tasks, one frontier agent configuration achieves a 68.0% pass rate with the artefact and 0% without it; on one task, a plausible but incorrect artefact also yields 0% across five trials. Deterministic solvers and rule corpora provide exact ground truth for structured tasks, while named criterion-level rubrics support outputs that cannot be checked by a single executable oracle. A configuration-relative calibration screen retains seven tasks satisfying our five-trial empirical knowledge-gating screen. These experiments validate the behavior of the construction protocol; they do not establish that the retained tasks improve post-training. We publicly release part of the task suite and supporting tooling at https://github.com/DatagridsAI/Knowledge-Gated-Task-Construction.

Estimating the difficulty of a dataset typically involves comparing state-of-the-art models to humans; the bigger the performance gap, the harder the dataset is said to be. However, this comparison provides little understanding of how difficult each instance in a given distribution is, or what attributes make the dataset difficult for a given model. To address these questions, we frame dataset difficulty -- w.r.t. a model V -- as the lack of V-usable information (Xu et al., 2019), where a lower value indicates a more difficult dataset for V. We further introduce pointwise \mathcal{V-information} (PVI) for measuring the difficulty of individual instances w.r.t. a given distribution. While standard evaluation metrics typically only compare different models for the same dataset, V-usable information and PVI also permit the converse: for a given model V, we can compare different datasets, as well as different instances/slices of the same dataset. Furthermore, our framework allows for the interpretability of different input attributes via transformations of the input, which we use to discover annotation artefacts in widely-used NLP benchmarks.

Alzheimer's disease (AD) constitutes a complex neurocognitive disease and is the main cause of dementia. Although many studies have been proposed targeting at diagnosing dementia through spontaneous speech, there are still limitations. Existing state-of-the-art approaches, which propose multimodal methods, train separately language and acoustic models, employ majority-vote approaches, and concatenate the representations of the different modalities either at the input level, i.e., early fusion, or during training. Also, some of them employ self-attention layers, which calculate the dependencies between representations without considering the contextual information. In addition, no prior work has taken into consideration the model calibration. To address these limitations, we propose some new methods for detecting AD patients, which capture the intra- and cross-modal interactions. First, we convert the audio files into log-Mel spectrograms, their delta, and delta-delta and create in this way an image per audio file consisting of three channels. Next, we pass each transcript and image through BERT and DeiT models respectively. After that, context-based self-attention layers, self-attention layers with a gate model, and optimal transport domain adaptation methods are employed for capturing the intra- and inter-modal interactions. Finally, we exploit two methods for fusing the self and cross-attended features. For taking into account the model calibration, we apply label smoothing. We use both performance and calibration metrics. Experiments conducted on the ADReSS Challenge dataset indicate the efficacy of our introduced approaches over existing research initiatives with our best performing model reaching Accuracy and F1-score up to 91.25% and 91.06% respectively.

Vision-Language Models (VLMs) have advanced across multimodal benchmarks but still show clear gaps in ordinal number understanding, i.e., the ability to track relative positions and generalize to large indices. We present OrdinalBench, a diagnostic benchmark that standardizes ordinal number understanding as an evaluation task for VLMs. The core task is N-th object identification, defined by a starting reference and traversal rule. Task difficulty is controlled along three axes: (i) ordinal magnitude, from small numbers to extreme cases up to 300; (ii) arrangement complexity, from single loops to maze-like paths; and (iii) object count. The benchmark provides 39,000 question-answer pairs, each annotated with a ground-truth reasoning trajectory and balanced across difficulty levels for controlled large-scale testing. Beyond answer-only evaluation, our framework requires models to generate structured stepwise traces of the counting process and provides an open evaluation toolkit that measures both final accuracy and step-level path consistency. Zero-shot evaluations of GPT-5, Gemini 2.5 Flash Lite, Qwen2.5-VL, InternVL3.5, and Molmo reveal sharp degradation under large-ordinal and complex-path conditions, highlighting weak generalization despite strong scores on standard multimodal tasks. By framing ordinal number understanding as a core target, OrdinalBench provides a reproducible benchmark and diagnostic framework for developing VLMs with stronger sequential reasoning. All data and code are available at https://ordinalbench.github.io/

Recent benchmarks rely on user simulators to evaluate AI agents in multi-turn interaction. While existing simulation techniques demonstrate surface fidelity to human style and behavior, ecologically valid interactive benchmarking also requires alignment in when and how agents fail across simulated and real user populations. We find that existing simulators lack outcome calibration: agreement with observed success rates and failure patterns when real users interact with the same agent. We introduce Calibrated User Embeddings (CUE), a framework that both encodes observed sessions and samples continuous representations, then decodes them into persona commands to steer LLMs to act as user simulators without training. Through this, we evaluate user-conditioned replay of past sessions and aggregate metric agreement when sampling novel personas for the same tasks. On τ^2-Bench, CUEd simulators commit fewer simulator-attributed errors and more faithfully reproduce real-user agent failure modes, aggregate success rates, and outcomes for specific task-user pairs than other persona-based simulation methods. These gains coexist with competitive user fidelity as measured using metrics established in prior work. After being fit to mostly customer support interactions, the same CUEd simulators generalize to document creation, math tutoring, and casual conversation, and remain effective across different simulator LLMs without CUE retraining.

Pre-trained Transformers are now ubiquitous in natural language processing, but despite their high end-task performance, little is known empirically about whether they are calibrated. Specifically, do these models' posterior probabilities provide an accurate empirical measure of how likely the model is to be correct on a given example? We focus on BERT and RoBERTa in this work, and analyze their calibration across three tasks: natural language inference, paraphrase detection, and commonsense reasoning. For each task, we consider in-domain as well as challenging out-of-domain settings, where models face more examples they should be uncertain about. We show that: (1) when used out-of-the-box, pre-trained models are calibrated in-domain, and compared to baselines, their calibration error out-of-domain can be as much as 3.5x lower; (2) temperature scaling is effective at further reducing calibration error in-domain, and using label smoothing to deliberately increase empirical uncertainty helps calibrate posteriors out-of-domain.

We introduce TuneJury, an open, instance-level pairwise reward model for text-to-music that predicts a music preference score from a text prompt and an audio clip. The released checkpoint is trained on publicly available human-preference labels covering arena-style (A vs. B) votes, metric-alignment preference pairs, crowdsourced pairwise comparisons, and expert aesthetic ratings. The predicted score margin between two clips is well calibrated on our held-out test split, supporting data filtering via a simple score threshold. TuneJury generalizes to both held-out test pairs and out-of-distribution benchmarks, remaining competitive with prior baselines on the latter. For generators released after training, we introduce anchor calibration, a post-hoc, per-system Bradley-Terry calibration that recovers agreement at substantially better data efficiency than from-scratch retraining. The same frozen reward drives consistent reward-axis gains across three downstream applications: inference-time best-of-N selection, DITTO-style latent optimization, and expert-iteration post-training. TuneJury is available at https://github.com/yonghyunk1m/TuneJury.

In legal decisions, split votes (SV) occur when judges cannot reach a unanimous decision, posing a difficulty for lawyers who must navigate diverse legal arguments and opinions. In high-stakes domains, understanding the alignment of perceived difficulty between humans and AI systems is crucial to build trust. However, existing NLP calibration methods focus on a classifier's awareness of predictive performance, measured against the human majority class, overlooking inherent human label variation (HLV). This paper explores split votes as naturally observable human disagreement and value pluralism. We collect judges' vote distributions from the European Court of Human Rights (ECHR), and present SV-ECHR, a case outcome classification (COC) dataset with SV information. We build a taxonomy of disagreement with SV-specific subcategories. We further assess the alignment of perceived difficulty between models and humans, as well as confidence- and human-calibration of COC models. We observe limited alignment with the judge vote distribution. To our knowledge, this is the first systematic exploration of calibration to human judgements in legal NLP. Our study underscores the necessity for further research on measuring and enhancing model calibration considering HLV in legal decision tasks.

We present an online post-hoc calibration method, called Online Platt Scaling (OPS), which combines the Platt scaling technique with online logistic regression. We demonstrate that OPS smoothly adapts between i.i.d. and non-i.i.d. settings with distribution drift. Further, in scenarios where the best Platt scaling model is itself miscalibrated, we enhance OPS by incorporating a recently developed technique called calibeating to make it more robust. Theoretically, our resulting OPS+calibeating method is guaranteed to be calibrated for adversarial outcome sequences. Empirically, it is effective on a range of synthetic and real-world datasets, with and without distribution drifts, achieving superior performance without hyperparameter tuning. Finally, we extend all OPS ideas to the beta scaling method.

Progress in large language models is constrained by an evaluation bottleneck: build a benchmark, evaluate models and settings, then iterate. We therefore ask a simple question: can we forecast outcomes before running any experiments? We study text-only performance forecasting: estimating a model's score from a redacted task description and intended configuration, with no access to dataset instances. To support systematic study, we curate PRECOG, a corpus of redacted description-performance pairs spanning diverse tasks, domains, and metrics. Experiments show the task is challenging but feasible: models equipped with a retrieval module that excludes source papers achieve moderate prediction performance with well-calibrated uncertainty, reaching mean absolute error as low as 8.7 on the Accuracy subset at high-confidence thresholds. Our analysis indicates that stronger reasoning models engage in diverse, iterative querying, whereas current open-source models lag and often skip retrieval or gather evidence with limited diversity. We further test a zero-leakage setting, forecasting on newly released datasets or experiments before their papers are indexed, where GPT-5 with built-in web search still attains nontrivial prediction accuracy. Overall, our corpus and analyses offer an initial step toward open-ended anticipatory evaluation, supporting difficulty estimation and smarter experiment prioritization.

Crowdsourcing has emerged as an effective platform for labeling large amounts of data in a cost- and time-efficient manner. Most previous work has focused on designing an efficient algorithm to recover only the ground-truth labels of the data. In this paper, we consider multi-choice crowdsourcing tasks with the goal of recovering not only the ground truth, but also the most confusing answer and the confusion probability. The most confusing answer provides useful information about the task by revealing the most plausible answer other than the ground truth and how plausible it is. To theoretically analyze such scenarios, we propose a model in which there are the top two plausible answers for each task, distinguished from the rest of the choices. Task difficulty is quantified by the probability of confusion between the top two, and worker reliability is quantified by the probability of giving an answer among the top two. Under this model, we propose a two-stage inference algorithm to infer both the top two answers and the confusion probability. We show that our algorithm achieves the minimax optimal convergence rate. We conduct both synthetic and real data experiments and demonstrate that our algorithm outperforms other recent algorithms. We also show the applicability of our algorithms in inferring the difficulty of tasks and in training neural networks with top-two soft labels.

The superficial alignment hypothesis (SAH) posits that large language models learn most of their knowledge during pre-training, and that post-training merely surfaces this knowledge. The SAH, however, lacks a precise definition, which has led to (i) different and seemingly orthogonal arguments supporting it, and (ii) important critiques to it. We propose a new metric called task complexity: the length of the shortest program that achieves a target performance on a task. In this framework, the SAH simply claims that pre-trained models drastically reduce the complexity of achieving high performance on many tasks. Our definition unifies prior arguments supporting the SAH, interpreting them as different strategies to find such short programs. Experimentally, we estimate the task complexity of mathematical reasoning, machine translation, and instruction following; we then show that these complexities can be remarkably low when conditioned on a pre-trained model. Further, we find that pre-training enables access to strong performances on our tasks, but it can require programs of gigabytes of length to access them. Post-training, on the other hand, collapses the complexity of reaching this same performance by several orders of magnitude. Overall, our results highlight that task adaptation often requires surprisingly little information -- often just a few kilobytes.

Deep neural networks are increasingly utilized in various machine learning tasks. However, as these models grow in complexity, they often face calibration issues, despite enhanced prediction accuracy. Many studies have endeavored to improve calibration performance through the use of specific loss functions, data preprocessing and training frameworks. Yet, investigations into calibration properties have been somewhat overlooked. Our study leverages the Neural Architecture Search (NAS) search space, offering an exhaustive model architecture space for thorough calibration properties exploration. We specifically create a model calibration dataset. This dataset evaluates 90 bin-based and 12 additional calibration measurements across 117,702 unique neural networks within the widely employed NATS-Bench search space. Our analysis aims to answer several longstanding questions in the field, using our proposed dataset: (i) Can model calibration be generalized across different datasets? (ii) Can robustness be used as a calibration measurement? (iii) How reliable are calibration metrics? (iv) Does a post-hoc calibration method affect all models uniformly? (v) How does calibration interact with accuracy? (vi) What is the impact of bin size on calibration measurement? (vii) Which architectural designs are beneficial for calibration? Additionally, our study bridges an existing gap by exploring calibration within NAS. By providing this dataset, we enable further research into NAS calibration. As far as we are aware, our research represents the first large-scale investigation into calibration properties and the premier study of calibration issues within NAS. The project page can be found at https://www.taolinwei.com/calibration-study

Current benchmarks for language models primarily evaluate execution on fully specified tasks. However, real user tasks are often ambiguous. Users arrive with incomplete, exploratory, or even inconsistent goals, requiring the assistant to first determine the intended task before carrying it out. We study this problem as task alignment: the ability to align with a user on their intended task. We introduce a general framework for converting specified tasks into underspecified interactions, formalized as a POMDP in which the model must infer a latent task from partial and evolving user intent. We validate our user simulator post hoc with a human user study. Across shopping, coding, and professional work settings, we find that while models often perform well once the task is specified, models still struggle with task alignment: current models act prematurely, interact ineffectively, and fail to resolve ambiguous requests. Models on average recover the user's intended task only 22-32% of the time under ambiguity. In a human study in the same setting, humans reach 48%, outperforming all evaluated models. We show that post-training with supervised fine-tuning and reinforcement learning improves task alignment, but models still lag behind humans in resolving uncertainty through interaction. Together, our results suggest that current models still lack key interaction abilities required for reliable agency.

Scaling up language models has been empirically shown to improve performance on a wide range of downstream tasks. However, if we were to observe worse performance as a function of scale ("inverse scaling") on certain tasks, this would indicate that scaling can also encourage behaviors that are misaligned with human preferences. The Inverse Scaling Prize (McKenzie et al. 2022) identified eleven such inverse scaling tasks, evaluated on models of up to 280B parameters and up to 500 zettaFLOPs of training compute. This paper takes a closer look at these inverse scaling tasks. We evaluate models of up to 540B parameters, trained on five times more compute than those evaluated in the Inverse Scaling Prize. With this increased range of model sizes and training compute, only four out of the eleven tasks remain inverse scaling. Six out of the eleven tasks exhibit "U-shaped scaling", where performance decreases up to a certain size, and then increases again up to the largest model evaluated (the one remaining task displays positive scaling). In addition, we find that 1-shot examples and chain-of-thought can help mitigate undesirable scaling patterns even further. U-shaped scaling suggests that the inverse scaling trend observed in McKenzie et al. (2022) may not continue to hold for larger models, which we attribute to the presence of distractor tasks that only sufficiently large models can avoid.

While dialogue remains an important end-goal of natural language research, the difficulty of evaluation is an oft-quoted reason why it remains troublesome to make real progress towards its solution. Evaluation difficulties are actually two-fold: not only do automatic metrics not correlate well with human judgments, but also human judgments themselves are in fact difficult to measure. The two most used human judgment tests, single-turn pairwise evaluation and multi-turn Likert scores, both have serious flaws as we discuss in this work. We instead provide a novel procedure involving comparing two full dialogues, where a human judge is asked to pay attention to only one speaker within each, and make a pairwise judgment. The questions themselves are optimized to maximize the robustness of judgments across different annotators, resulting in better tests. We also show how these tests work in self-play model chat setups, resulting in faster, cheaper tests. We hope these tests become the de facto standard, and will release open-source code to that end.

Large language models (LLMs) offer powerful capabilities but incur substantial computational costs, driving the need for efficient compression techniques. This study evaluates the impact of popular compression methods - Magnitude Pruning, SparseGPT, and Wanda - on the LLaMA-2-7B model, focusing on the trade-offs between model size reduction, downstream task performance, and the role of calibration data. Our findings reveal that while SparseGPT and Wanda preserve perplexity even at 50% sparsity, they suffer significant degradation on downstream tasks, highlighting the inadequacy of perplexity as the sole evaluation metric. To address this, we introduce Jensen-Shannon (JS) Divergence as a more comprehensive metric that captures nuanced changes in model behavior post-compression. We further demonstrate that task-specific calibration data significantly enhances the downstream performance of compressed models compared to general calibration data. This research underscores the necessity for diverse evaluation metrics and careful calibration data selection to fully understand the complexities of LLM compression and its implications for practical applications.

Large Language Models deployed as question answering tools require robust calibration to avoid overconfidence. We systematically evaluate how reasoning capabilities and budget affect confidence assessment accuracy, using the ClimateX dataset (Lacombe et al., 2023) and expanding it to human and planetary health. Our key finding challenges the "test-time scaling" paradigm: while recent reasoning LLMs achieve 48.7% accuracy in assessing expert confidence, increasing reasoning budgets consistently impairs rather than improves calibration. Extended reasoning leads to systematic overconfidence that worsens with longer thinking budgets, producing diminishing and negative returns beyond modest computational investments. Conversely, search-augmented generation dramatically outperforms pure reasoning, achieving 89.3% accuracy by retrieving relevant evidence. Our results suggest that information access, rather than reasoning depth or inference budget, may be the critical bottleneck for improved confidence calibration of knowledge-intensive tasks.

The ability to determine when a person struggles during skill acquisition is crucial for both optimizing human learning and enabling the development of effective assistive systems. As skills develop, the type and frequency of struggles tend to change, and understanding this evolution is key to determining the user's current stage of learning. However, existing manipulation datasets have not focused on how struggle evolves over time. In this work, we collect a dataset for struggle determination, featuring 61.68 hours of video recordings, 2,793 videos, and 5,385 annotated temporal struggle segments collected from 76 participants. The dataset includes 18 tasks grouped into four diverse activities -- tying knots, origami, tangram puzzles, and shuffling cards, representing different task variations. In addition, participants repeated the same task five times to capture their evolution of skill. We define the struggle determination problem as a temporal action localization task, focusing on identifying and precisely localizing struggle segments with start and end times. Experimental results show that Temporal Action Localization models can successfully learn to detect struggle cues, even when evaluated on unseen tasks or activities. The models attain an overall average mAP of 34.56% when generalizing across tasks and 19.24% across activities, indicating that struggle is a transferable concept across various skill-based tasks while still posing challenges for further improvement in struggle detection. Our dataset is available at https://github.com/FELIXFENG2019/EvoStruggle.

Visual instruction tuning is a key training stage of large multimodal models. However, when learning multiple visual tasks simultaneously, this approach often results in suboptimal and imbalanced overall performance due to latent knowledge conflicts across tasks. To mitigate this issue, we propose a novel Adaptive Task Balancing approach tailored for visual instruction tuning (VisATB). Specifically, we measure two critical dimensions for visual task balancing based on validation performance: (1) Inter-Task Contribution, the mechanism where learning one task enhances the performance on others owing to shared knowledge across tasks, and (2) Intra-Task Difficulty, which denotes the inherent learning difficulty of a single task. Furthermore, we propose prioritizing three categories of tasks with greater weight: those that offer substantial contributions to others, those that receive minimal contributions from others, and those that present high learning difficulties. Among these three task weighting strategies, the first and third focus on improving overall performance, and the second targets the mitigation of performance imbalance. Extensive experiments on three benchmarks demonstrate that our VisATB approach consistently achieves superior and more balanced overall performance in visual instruction tuning. The data, code, and models are available at https://github.com/YanqiDai/VisATB.

Deep neural networks have demonstrated remarkable performance across numerous learning tasks but often suffer from miscalibration, resulting in unreliable probability outputs. This has inspired many recent works on mitigating miscalibration, particularly through post-hoc recalibration methods that aim to obtain calibrated probabilities without sacrificing the classification performance of pre-trained models. In this study, we summarize and categorize previous works into three general strategies: intuitively designed methods, binning-based methods, and methods based on formulations of ideal calibration. Through theoretical and practical analysis, we highlight ten common limitations in previous approaches. To address these limitations, we propose a probabilistic learning framework for calibration called h-calibration, which theoretically constructs an equivalent learning formulation for canonical calibration with boundedness. On this basis, we design a simple yet effective post-hoc calibration algorithm. Our method not only overcomes the ten identified limitations but also achieves markedly better performance than traditional methods, as validated by extensive experiments. We further analyze, both theoretically and experimentally, the relationship and advantages of our learning objective compared to traditional proper scoring rule. In summary, our probabilistic framework derives an approximately equivalent differentiable objective for learning error-bounded calibrated probabilities, elucidating the correspondence and convergence properties of computational statistics with respect to theoretical bounds in canonical calibration. The theoretical effectiveness is verified on standard post-hoc calibration benchmarks by achieving state-of-the-art performance. This research offers valuable reference for learning reliable likelihood in related fields.

Scholastic trivia competitions test knowledge and intelligence through mastery of question answering. Modern question answering benchmarks are one variant of the Turing test. Specifically, answering a set of questions as well as a human is a minimum bar towards demonstrating human-like intelligence. This paper makes the case that the format of one competition -- where participants can answer in the middle of hearing a question (incremental) -- better differentiates the skill between (human or machine) players. Additionally, merging a sequential decision-making sub-task with question answering (QA) provides a good setting for research in model calibration and opponent modeling. Thus, embedded in this task are three machine learning challenges: (1) factoid QA over thousands of Wikipedia-like answers, (2) calibration of the QA model's confidence scores, and (3) sequential decision-making that incorporates knowledge of the QA model, its calibration, and what the opponent may do. We make two contributions: (1) collecting and curating a large factoid QA dataset and an accompanying gameplay dataset, and (2) developing a model that addresses these three machine learning challenges. In addition to offline evaluation, we pitted our model against some of the most accomplished trivia players in the world in a series of exhibition matches spanning several years. Throughout this paper, we show that collaborations with the vibrant trivia community have contributed to the quality of our dataset, spawned new research directions, and doubled as an exciting way to engage the public with research in machine learning and natural language processing.

While text-to-image (T2I) generative models have become ubiquitous, they do not necessarily generate images that align with a given prompt. While previous work has evaluated T2I alignment by proposing metrics, benchmarks, and templates for collecting human judgements, the quality of these components is not systematically measured. Human-rated prompt sets are generally small and the reliability of the ratings -- and thereby the prompt set used to compare models -- is not evaluated. We address this gap by performing an extensive study evaluating auto-eval metrics and human templates. We provide three main contributions: (1) We introduce a comprehensive skills-based benchmark that can discriminate models across different human templates. This skills-based benchmark categorises prompts into sub-skills, allowing a practitioner to pinpoint not only which skills are challenging, but at what level of complexity a skill becomes challenging. (2) We gather human ratings across four templates and four T2I models for a total of >100K annotations. This allows us to understand where differences arise due to inherent ambiguity in the prompt and where they arise due to differences in metric and model quality. (3) Finally, we introduce a new QA-based auto-eval metric that is better correlated with human ratings than existing metrics for our new dataset, across different human templates, and on TIFA160.

Robustness is often regarded as a critical future challenge for real-world applications, where stability is essential. However, as models often learn tasks in a similar order, we hypothesize that easier tasks will be easier regardless of how they are presented to the model. Indeed, in this paper, we show that as models approach high performance on a task, robustness is effectively achieved. Through an empirical analysis of multiple models across diverse datasets and configurations (e.g., paraphrases, different temperatures), we find a strong positive correlation. Moreover, we find that robustness is primarily driven by task-specific competence rather than inherent model-level properties, challenging current approaches that treat robustness as an independent capability. Thus, from a high-level perspective, we may expect that as new tasks saturate, model robustness on these tasks will emerge accordingly. For researchers, this implies that explicit efforts to measure and improve robustness may warrant reduced emphasis, as such robustness is likely to develop alongside performance gains. For practitioners, it acts as a sign that indeed the tasks that the literature deals with are unreliable, but on easier past tasks, the models are reliable and ready for real-world deployment.

Progress in AI is often demonstrated by new models claiming improved performance on tasks measuring model capabilities. Evaluating language models in particular is challenging, as small changes to how a model is evaluated on a task can lead to large changes in measured performance. There is no common standard setup, so different models are evaluated on the same tasks in different ways, leading to claims about which models perform best not being reproducible. We propose OLMES, a completely documented, practical, open standard for reproducible LLM evaluations. In developing this standard, we identify and review the varying factors in evaluation practices adopted by the community - such as details of prompt formatting, choice of in-context examples, probability normalizations, and task formulation. In particular, OLMES supports meaningful comparisons between smaller base models that require the unnatural "cloze" formulation of multiple-choice questions against larger models that can utilize the original formulation. OLMES includes well-considered recommendations guided by results from existing literature as well as new experiments investigating open questions.

Instruction embedding models have become common among state-of-the-art models, however are evaluated using a single prompt per task. The single-point evaluation ignores a main problem of the instruction-based approach namely: sensitivity to the phrasing of the instruction. We present an empirical study of prompt sensitivity across 6 embedding models, 11 datasets, and 15 task-specific prompts per dataset, a total of 990. We show that reported scores misrepresent the distribution of scores over plausible prompts. The default prompt can both systematically understate or overstate performance. Furthermore, we show that the leaderboard ranking is not robust to prompt selection: by choosing prompts favorably, any model in our study can be promoted to first place. Our findings suggest that single-prompt evaluation is insufficient for instruction-tuned embedding models and that benchmarks should incorporate prompt robustness, either by evaluating over multiple prompts or by reporting sensitivity alongside point estimates.

We propose a multitask pretraining approach ZeroPrompt for zero-shot generalization, focusing on task scaling and zero-shot prompting. While previous models are trained on only a few dozen tasks, we scale to 1,000 tasks for the first time using real-world data. This leads to a crucial discovery that task scaling can be an efficient alternative to model scaling; i.e., the model size has little impact on performance with an extremely large number of tasks. Our results show that task scaling can substantially improve training efficiency by 30 times in FLOPs. Moreover, we present a prompting method that incorporates a genetic algorithm to automatically search for the best prompt for unseen tasks, along with a few other improvements. Empirically, ZeroPrompt substantially improves both the efficiency and the performance of zero-shot learning across a variety of academic and production datasets.

Process reward models (PRMs) play a central role in guiding inference-time scaling algorithms for large language models (LLMs). However, we observe that even state-of-the-art PRMs can be poorly calibrated and often overestimate success probabilities. To address this, we present a calibration approach, performed via quantile regression, that adjusts PRM outputs to better align with true success probabilities. Leveraging these calibrated success estimates and their associated confidence bounds, we introduce an instance-adaptive scaling (IAS) framework that dynamically adjusts the inference budget based on the estimated likelihood that a partial reasoning trajectory will yield a correct final answer. Unlike conventional methods that allocate a fixed number of reasoning trajectories per query, this approach successfully adapts to each instance and reasoning step when using our calibrated PRMs. Experiments on mathematical reasoning benchmarks show that (i) our PRM calibration method successfully achieves small calibration error, outperforming the baseline methods, (ii) calibration is crucial for enabling effective adaptive scaling, and (iii) the proposed IAS strategy reduces inference costs while maintaining final answer accuracy, utilizing less compute on more confident problems as desired.

Large language models (LLMs) have demonstrated impressive capabilities across various tasks, but their performance is highly sensitive to the prompts utilized. This variability poses challenges for accurate assessment and user satisfaction. Current research frequently overlooks instance-level prompt variations and their implications on subjective evaluations. To address these shortcomings, we introduce ProSA, a framework designed to evaluate and comprehend prompt sensitivity in LLMs. ProSA incorporates a novel sensitivity metric, PromptSensiScore, and leverages decoding confidence to elucidate underlying mechanisms. Our extensive study, spanning multiple tasks, uncovers that prompt sensitivity fluctuates across datasets and models, with larger models exhibiting enhanced robustness. We observe that few-shot examples can alleviate this sensitivity issue, and subjective evaluations are also susceptible to prompt sensitivities, particularly in complex, reasoning-oriented tasks. Furthermore, our findings indicate that higher model confidence correlates with increased prompt robustness. We believe this work will serve as a helpful tool in studying prompt sensitivity of LLMs. The project is released at: https://github.com/open-compass/ProSA .

Large language models (LLMs) increasingly exhibit human-like patterns of pragmatic and social reasoning. This paper addresses two related questions: do LLMs approximate human social meaning not only qualitatively but also quantitatively, and can prompting strategies informed by pragmatic theory improve this approximation? To address the first, we introduce two calibration-focused metrics distinguishing structural fidelity from magnitude calibration: the Effect Size Ratio (ESR) and the Calibration Deviation Score (CDS). To address the second, we derive prompting conditions from two pragmatic assumptions: that social meaning arises from reasoning over linguistic alternatives, and that listeners infer speaker knowledge states and communicative motives. Applied to a case study on numerical (im)precision across three frontier LLMs, we find that all models reliably reproduce the qualitative structure of human social inferences but differ substantially in magnitude calibration. Prompting models to reason about speaker knowledge and motives most consistently reduces magnitude deviation, while prompting for alternative-awareness tends to amplify exaggeration. Combining both components is the only intervention that improves all calibration-sensitive metrics across all models, though fine-grained magnitude calibration remains only partially resolved. LLMs thus capture inferential structure while variably distorting inferential strength, and pragmatic theory provides a useful but incomplete handle for improving that approximation.

Foundation models, specifically Large Language Models (LLMs), have lately gained wide-spread attention and adoption. Reinforcement Learning with Human Feedback (RLHF) involves training a reward model to capture desired behaviors, which is then used to align LLM's. These reward models are additionally used at inference-time to estimate LLM responses' adherence to those desired behaviors. However, there is little work measuring how robust these reward models are to distribution shifts. In this work, we evaluate how reward model performance - measured via accuracy and calibration (i.e. alignment between accuracy and confidence) - is affected by distribution shift. We show novel calibration patterns and accuracy drops due to OOD prompts and responses, and that the reward model is more sensitive to shifts in responses than prompts. Additionally, we adapt an OOD detection technique commonly used in classification to the reward model setting to detect these distribution shifts in prompts and responses.

Running LLMs with extended reasoning on every problem is expensive, but determining which inputs actually require additional compute remains challenging. We investigate whether their own likelihood of success is recoverable from their internal representations before generation, and if this signal can guide more efficient inference. We train linear probes on pre-generation activations to predict policy-specific success on math and coding tasks, substantially outperforming surface features such as question length and TF-IDF. Using E2H-AMC, which provides both human and model performance on identical problems, we show that models encode a model-specific notion of difficulty that is distinct from human difficulty, and that this distinction increases with extended reasoning. Leveraging these probes, we demonstrate that routing queries across a pool of models can exceed the best-performing model whilst reducing inference cost by up to 70\% on MATH, showing that internal representations enable practical efficiency gains even when they diverge from human intuitions about difficulty. Our code is available at: https://github.com/KabakaWilliam/llms_know_difficulty

Graph-augmented retrieval combines dense similarity with graph-based relevance signals such as Personalized PageRank (PPR), but these scores have different distributions and are not directly comparable. We study this as a score calibration problem for heterogeneous retrieval fusion in multi-hop question answering. Our method, PhaseGraph, maps vector and graph scores to a common unit-free scale using percentile-rank normalization (PIT) before fusion, enabling stable combination without discarding magnitude information. Across MuSiQue and 2WikiMultiHopQA, calibrated fusion improves held-out last-hop retrieval on HippoRAG2-style benchmarks: LastHop@5 increases from 75.1% to 76.5% on MuSiQue (8W/1L, p=0.039) and from 51.7% to 53.6% on 2WikiMultiHopQA (11W/2L, p=0.023), both on independent held-out test splits. A theory-driven ablation shows that percentile-based calibration is directionally more robust than min-max normalization on both tune and test splits (1W/6L, p=0.125), while Boltzmann weighting performs comparably to linear fusion after calibration (0W/3L, p=0.25). These results suggest that score commensuration is a robust design choice, and the exact post-calibration operator appears to matter less on these benchmarks.

As browser-use agents improve, benchmarks keep pace by collecting new tasks, websites, and applications, often making tasks longer or more novel. This makes difficulty expensive to refresh and difficult to control: when many aspects change at once, it is unclear what actually makes a task challenging. We instead construct challenging instances from tasks agents already solve, turning difficulty into a programmable property of the environment. BreakingWeb pairs every base task with an intervention condition that preserves the user instruction, latent target, and backend success criterion while changing the environment at different web stack layers. Each intervention is deterministic, detectable, and recoverable, and is annotated with the cognitive primitive it primarily loads. The benchmark contains 519 clean/intervention task pairs across seven self-hosted websites and 29 intervention families, all graded against outcomes. We evaluate six strong browser-use agents, three GUI-only agents that see only screenshots, and humans. The construction is effective: interventions cut agent pass rate by 22.9% on average and overturn nearly half of the tasks each agent solves cleanly, whereas humans lose 10.0% on a first attempt and 5.7% after one familiarisation attempt. The dominant failure is belief failure: 75% of the six agents' failures end with a declared success although the required change never happened. Our code, data and environment are publicly available at www.breakingweb.app.

When language models (LMs) are trained via reinforcement learning (RL) to generate natural language "reasoning chains", their performance improves on a variety of difficult question answering tasks. Today, almost all successful applications of RL for reasoning use binary reward functions that evaluate the correctness of LM outputs. Because such reward functions do not penalize guessing or low-confidence outputs, they often have the unintended side-effect of degrading calibration and increasing the rate at which LMs generate incorrect responses (or "hallucinate") in other problem domains. This paper describes RLCR (Reinforcement Learning with Calibration Rewards), an approach to training reasoning models that jointly improves accuracy and calibrated confidence estimation. During RLCR, LMs generate both predictions and numerical confidence estimates after reasoning. They are trained to optimize a reward function that augments a binary correctness score with a Brier score -- a scoring rule for confidence estimates that incentivizes calibrated prediction. We first prove that this reward function (or any analogous reward function that uses a bounded, proper scoring rule) yields models whose predictions are both accurate and well-calibrated. We next show that across diverse datasets, RLCR substantially improves calibration with no loss in accuracy, on both in-domain and out-of-domain evaluations -- outperforming both ordinary RL training and classifiers trained to assign post-hoc confidence scores. While ordinary RL hurts calibration, RLCR improves it. Finally, we demonstrate that verbalized confidence can be leveraged at test time to improve accuracy and calibration via confidence-weighted scaling methods. Our results show that explicitly optimizing for calibration can produce more generally reliable reasoning models.

Pruning and quantization form the foundation of model compression for neural networks, enabling efficient inference for large language models (LLMs). Recently, various quantization and pruning techniques have demonstrated state-of-the-art performance in a post-training setting. They rely upon calibration data, a small set of unlabeled examples, to generate layer activations. However, no prior work has systematically investigated how the calibration data impacts the effectiveness of model compression methods. In this paper, we present the first extensive empirical study on the effect of calibration data upon LLM performance. We trial a variety of pruning and quantization methods, tasks, models, and datasets. Surprisingly, we find substantial variations in downstream task performance, contrasting existing work that suggests a greater level of robustness to the calibration data. Finally, we make a series of recommendations for the effective use of calibration data in LLM quantization and pruning.

While multimodal large language models have advanced across text, image, and audio, personalization research has remained primarily vision-language, with unified omnimodal benchmarking that jointly covers text, image, and audio still limited, and lacking the methodological rigor to account for absent-persona scenarios or systematic grounding studies. We introduce Omni-Persona, the first comprehensive benchmark for omnimodal personalization. We formalize the task as cross-modal routing over the Persona Modality Graph, encompassing 4 task groups and 18 fine-grained tasks across {sim}750 items. To rigorously diagnose grounding behavior, we propose Calibrated Accuracy (mathrm{Cal)}, which jointly rewards correct grounding and appropriate abstention, incorporating absent-persona queries within a unified evaluation framework. On our dedicated experiments, three diagnostic findings emerge: (i) open-source models show a consistent audio-vs-visual grounding gap that RLVR partially narrows via dense rule-based supervision; (ii) answerable recall and parameter scale are incomplete diagnostics, since strong recall can coexist with absent-persona hallucination and larger models do not always achieve higher Cal, exposing calibration as a separate evaluation axis; and (iii) SFT is bounded by the difficulty of constructing annotated ground-truth supervision at scale, while RLVR generalizes more consistently through outcome-level verifiable feedback yet drifts toward conservative behavior and lower generation quality under our reward design. Omni-Persona thus serves as a diagnostic framework that surfaces the pitfalls of omnimodal personalization, guiding future post-training and reward design.

LLM agents often operate over multi-turn interactions in which user intent changes before execution. We study intent drift: the failure mode in which superseded parts of the user's intent continue to influence the final answer or tool action. We introduce IntentFlux, an executable benchmark that converts verifiable tasks into dialogues with controlled intent changes while preserving their original graders. In a 627-case calibration, mean task score falls from 0.476 to 0.384 as dialogues contain more superseded and withdrawn information. Across eight models, the rate of fully correct solutions is significantly lower when the same final task must be recovered from an evolving dialogue rather than given directly in a single turn. We further introduce StateForge, which explicitly maintains the active requirements before generation. On General-Test, it improves mean task score from 0.367 to 0.467. Providing the ground-truth final state improves performance further but still does not recover single-turn performance, indicating that state-estimation errors explain only part of the gap. These results establish intent drift as a measurable multi-turn failure mode and explicit state maintenance as a partial mitigation.

Improvements in language models' capabilities have pushed their applications towards longer contexts, making long-context evaluation and development an active research area. However, many disparate use-cases are grouped together under the umbrella term of "long-context", defined simply by the total length of the model's input, including - for example - Needle-in-a-Haystack tasks, book summarization, and information aggregation. Given their varied difficulty, in this position paper we argue that conflating different tasks by their context length is unproductive. As a community, we require a more precise vocabulary to understand what makes long-context tasks similar or different. We propose to unpack the taxonomy of long-context based on the properties that make them more difficult with longer contexts. We propose two orthogonal axes of difficulty: (I) Diffusion: How hard is it to find the necessary information in the context? (II) Scope: How much necessary information is there to find? We survey the literature on long-context, provide justification for this taxonomy as an informative descriptor, and situate the literature with respect to it. We conclude that the most difficult and interesting settings, whose necessary information is very long and highly diffused within the input, is severely under-explored. By using a descriptive vocabulary and discussing the relevant properties of difficulty in long-context, we can implement more informed research in this area. We call for a careful design of tasks and benchmarks with distinctly long context, taking into account the characteristics that make it qualitatively different from shorter context.

Lexical difficulty prediction is a fundamental problem in language learning and readability assessment, requiring models to estimate word difficulty across different first-language (L1) backgrounds. However, existing approaches rely on regression-only training with scalar supervision, which does not explicitly structure the representation space, limiting their ability to capture cross-lingual alignment and ordinal difficulty. To mitigate these issues, we propose Context-Aligned Contrastive Regression, which integrates Ridge regression ensemble with two complementary objectives, i.e., Cross-View Context and Ordinal Soft Contrastive Learning. Experiments on three L1 datasets show that (i) contrastive objectives improve cross-lingual representation alignment while preserving language-specific nuances, (ii) the learned representations capture the ordinal structure of lexical difficulty, and (iii) the ensemble effectively mitigates systematic biases of individual models, leading to more stable performance across difficulty levels.

Machine unlearning aims to remove the influence of specific training data from a model while preserving reliable behavior on the remaining data, making reliable prediction and uncertainty estimation essential for evaluation. Calibration is commonly used as a proxy for reliability in language models, but low calibration error does not necessarily imply reliable decision rules, as models may rely on spurious correlations while remaining well calibrated. We investigate this gap in generative language models using the multiple-choice question-answering evaluation protocol on the TOFU benchmark, measuring probabilistic reliability with calibration metrics (ECE, MCE, Brier) and decision-rule reliability via attribution-based shortcut detection with Integrated Gradients and Local Mutual Information. We find that fine-tuned models achieve low calibration error (ECE ~ 0.04) compared to pretrained models (ECE > 0.5), and models after unlearning retain similarly low calibration despite reduced accuracy on the forget split, while attribution analysis shows increased reliance on correlation-based tokens. These results demonstrate that good calibration can coexist with shortcut-based decision rules after unlearning, extending the reliability paradox to the machine unlearning setting.

Calibration of language models -- the alignment between expressed or implicit confidence and empirical correctness -- is a well-studied subfield within NLP. Methods to measure it already exist. The problem is adoption: outside this subfield, NLP research regularly introduces new models, datasets, and benchmarks without checking whether the model's confidence scores are meaningful. We argue that this adoption gap is a major obstacle to trustworthy LLM evaluation. Miscalibration causes problems in two distinct areas: at deployment, where overconfident mistakes cause real harm, and inside the research pipeline, where methods like LLM-as-a-judge, synthetic data generation, and active learning rely on calibrated confidence without verifying it. Standard calibration metrics only require two inputs per example: a confidence score and a correctness judgment. Most benchmarks in use today already provide both, meaning calibration can be reported immediately. For open-ended generation, however, defining these two inputs is still an open challenge. We argue that each NLP subfield should pair its main performance metric with a calibration score and call for treating calibration as an essential property of every model rather than a niche topic.

In this paper, we investigate which questions are challenging for retrieval-based Question Answering (QA). We (i) propose retrieval complexity (RC), a novel metric conditioned on the completeness of retrieved documents, which measures the difficulty of answering questions, and (ii) propose an unsupervised pipeline to measure RC given an arbitrary retrieval system. Our proposed pipeline measures RC more accurately than alternative estimators, including LLMs, on six challenging QA benchmarks. Further investigation reveals that RC scores strongly correlate with both QA performance and expert judgment across five of the six studied benchmarks, indicating that RC is an effective measure of question difficulty. Subsequent categorization of high-RC questions shows that they span a broad set of question shapes, including multi-hop, compositional, and temporal QA, indicating that RC scores can categorize a new subset of complex questions. Our system can also have a major impact on retrieval-based systems by helping to identify more challenging questions on existing datasets.

Evaluating software engineering agents on realistic benchmarks is costly, since each task may require multi-step code exploration, modification, and test execution. Existing efficient evaluation methods select representative subsets to estimate full-benchmark performance, but are largely result-only: they fit historical pass/fail response matrices or static task semantics, discarding how agents solve problems. We propose PTA-IRT, a Privileged Trajectory-Aware Item Response Theory framework that fuses process and outcome signals. Historical execution trajectories supply process-level evidence beyond pass/fail, such as explored context, attempted edits, and solving paths, which PTA-IRT uses as privileged information for calibration subset selection and ability estimation. Under low calibration budgets, PTA-IRT consistently outperforms prior IRT baselines on score and ranking recovery across four SWE benchmarks. Code and data are publicly available at https://github.com/DeepSoftwareAnalytics/PTA-IRT.

Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models possess numerous capabilities (e.g., mathematical reasoning, legal support, or medical diagnostic) as well as safety risks (e.g., racial bias, toxicity, or misinformation). The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice. Currently, holistic evaluations are costly due to the large volume of benchmark questions, making frequent evaluations impractical. A popular attempt to lower the cost is to compute the average score on a subset of the benchmark. This approach, unfortunately, often renders an unreliable measure of LM performance because the average score is often confounded with the difficulty of the questions in the benchmark subset. Item response theory (IRT) was designed to address this challenge, providing a reliable measurement by careful controlling for question difficulty. Unfortunately, question difficulty is expensive to estimate. Facing this challenge, we train a model that predicts question difficulty from its content, enabling a reliable measurement at a fraction of the cost. In addition, we leverage this difficulty predictor to further improve the evaluation efficiency through training a question generator given a difficulty level. This question generator is essential in adaptive testing, where, instead of using a random subset of the benchmark questions, informative questions are adaptively chosen based on the current estimation of LLM performance. Experiments on 22 common natural language benchmarks and 172 LMs show that this approach is more reliable and efficient compared to current common practice.

A well-calibrated model should express confidence that matches its actual accuracy -- when it claims 80\% confidence, it should be correct 80\% of the time. While large language models (LLMs) have achieved remarkable performance across diverse tasks, their epistemic calibration remains poorly understood. We introduce KalshiBench, a benchmark of 300 prediction market questions from Kalshi, a CFTC-regulated exchange, with verifiable real-world outcomes occurring after model training cutoffs. Unlike traditional benchmarks measuring accuracy on static knowledge, KalshiBench evaluates whether models can appropriately quantify uncertainty about genuinely unknown future events. We evaluate five frontier models -- Claude Opus 4.5, GPT-5.2, DeepSeek-V3.2, Qwen3-235B, and Kimi-K2 -- and find systematic overconfidence across all models. Even the best-calibrated model (Claude Opus 4.5, ECE=0.120) shows substantial calibration errors, while reasoning-enhanced models like GPT-5.2-XHigh exhibit worse calibration (ECE=0.395) despite comparable accuracy. Critically, only one model achieves a positive Brier Skill Score, indicating most models perform worse than simply predicting base rates. Our findings suggest that scaling and enhanced reasoning do not automatically confer calibration benefits, highlighting epistemic calibration as a distinct capability requiring targeted development.

Low-precision Transformer systems increasingly quantize attention matrix multiplications, while softmax often remains at higher precision. During pretraining, an approximate softmax changes the gradients that train the model as well as its forward computation. We study this interaction with K-interval attention, which approximates the exponential using K+1 grid values. We vary per-row grid calibration, interpolation versus hard rounding, and the placement of a straight-through surrogate relative to normalization. We derive the corresponding backward rules, including calibration derivatives, and compare these choices in pretraining experiments matched on model, data, and optimizer. Detaching the row extrema leaves the forward computation unchanged but produces a delayed increase in validation loss. With hard rounding at K=4, min-max calibration and a pre-normalization surrogate incur a large loss gap; changing either choice substantially reduces it. At 124M parameters and 2.5B training tokens, fixed-window calibration with a post-normalization surrogate yields a validation loss gap of +0.019 nats relative to softmax at K=4, and with a pre-normalization surrogate yields +0.004 nats at K=16.

We present Top-Theta (Top-theta) Attention, a training-free method for sparsifying transformer attention during inference. Our key insight is that static, per-head thresholds can be calibrated to retain the desired constant number of significant elements per attention row. This approach enables content-based sparsity without retraining, and it remains robust across data domains. We further introduce compensation techniques to preserve accuracy under aggressive sparsification, establishing attention thresholding as a practical and principled alternative to top-k attention. We provide extensive evaluation on natural language processing tasks, showing that Top-theta achieves 3-10x reduction in V-cache usage and up to 10x fewer attention elements during inference while degrading no more than 1% in accuracy.

State-of-the-art reasoning models utilize long chain-of-thought (CoT) to solve increasingly complex problems using more test-time computation. In this work, we explore a long CoT setting where the model makes up to K successive attempts at solving a problem, in which each attempt is allowed to build on earlier ones after the model receives a hard verifier feedback. This motivates RL methods that can harness per-attempt rewards by carefully weighting individual attempts. We study optimizing the Verification@K reward (the model succeeds by the K-th attempt) and show that naively weighing the attempts by their pass/fail results in biased gradients. We introduce Calibrated Attempt-Level (CAL) GRPO by devising a weighing strategy to obtain unbiased gradients while maintaining small variance. Our theory reveals how incorporating per-attempt rewards influence the training and the eventual Verification@K performance. Experiments, baselines, and ablations on synthetic and real data corroborate our theory and the benefits of CAL-GRPO over vanilla GRPO as well as naive weighting.

Large language models (LLMs) have shown remarkable performance in various tasks but often fail to handle queries that exceed their knowledge and capabilities, leading to incorrect or fabricated responses. This paper addresses the need for LLMs to recognize and refuse infeasible tasks due to the required skills surpassing their capabilities. We first systematically conceptualize infeasible tasks for LLMs, providing formal definitions and categorizations that cover a spectrum of related hallucinations. We develop and benchmark a new dataset comprising diverse infeasible and feasible tasks to test multiple LLMs' abilities on task feasibility. Furthermore, we explore the potential of training enhancements to increase LLMs' refusal capabilities with fine-tuning. Experiments validate the effectiveness of our methods, offering promising directions for refining the operational boundaries of LLMs in real applications.

The rapid release of both language models and benchmarks makes it increasingly costly to evaluate every model on every dataset. In practice, models are often evaluated on different samples, making scores difficult to compare across studies. To address this, we propose a framework based on multidimensional Item Response Theory (IRT) that uses anchor items to calibrate new benchmarks to the evaluation suite while holding previously calibrated item parameters fixed. Our approach supports a realistic evaluation setting in which datasets are introduced over time and models are evaluated only on the datasets available at the time of evaluation, while a fixed anchor set for each dataset is used so that results from different evaluation periods can be compared directly. In large-scale experiments on more than 400 models, our framework predicts full-evaluation performance within 2-3 percentage points using only 100 anchor questions per dataset, with Spearman ρgeq 0.9 for ranking preservation, showing that it is possible to extend benchmark suites over time while preserving score comparability, at a constant evaluation cost per new dataset. Code available at https://github.com/eliyahabba/growing-pains

Standard evaluation of LLM confidence relies on calibration metrics (ECE, Brier score) that conflate two distinct capacities: how much a model knows (Type-1 sensitivity) and how well it knows what it knows (Type-2 metacognitive sensitivity). We introduce an evaluation framework based on Type-2 Signal Detection Theory that decomposes these capacities using meta-d' and the metacognitive efficiency ratio M-ratio. Applied to four LLMs (Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.3, Llama-3-8B-Base, Gemma-2-9B-Instruct) across 224,000 factual QA trials, we find: (1) metacognitive efficiency varies substantially across models even when Type-1 sensitivity is similar -- Mistral achieves the highest d' but the lowest M-ratio; (2) metacognitive efficiency is domain-specific, with different models showing different weakest domains, invisible to aggregate metrics; (3) temperature manipulation shifts Type-2 criterion while meta-d' remains stable for two of four models, dissociating confidence policy from metacognitive capacity; (4) AUROC_2 and M-ratio produce fully inverted model rankings, demonstrating these metrics answer fundamentally different evaluation questions. The meta-d' framework reveals which models "know what they don't know" versus which merely appear well-calibrated due to criterion placement -- a distinction with direct implications for model selection, deployment, and human-AI collaboration. Pre-registered analysis; code and data publicly available.

Aligning large language models (LLMs) with human preferences has been recognized as the key to improving LLMs' interaction quality. However, in this pluralistic world, human preferences can be diversified due to annotators' different tastes, which hinders the effectiveness of LLM alignment methods. This paper presents the first quantitative analysis of commonly used human feedback datasets to investigate the impact of diversified preferences on reward modeling. Our analysis reveals a correlation between the calibration performance of reward models (RMs) and the alignment performance of LLMs. We find that diversified preference data negatively affect the calibration performance of RMs on human-shared preferences, such as Harmless\&Helpful, thereby impairing the alignment performance of LLMs. To address the ineffectiveness, we propose a novel Multi-Objective Reward learning method (MORE) to enhance the calibration performance of RMs on shared preferences. We validate our findings by experiments on three models and five human preference datasets. Our method significantly improves the prediction calibration of RMs, leading to better alignment of the Alpaca-7B model with Harmless\&Helpful preferences. Furthermore, the connection between reward calibration and preference alignment performance suggests that calibration error can be adopted as a key metric for evaluating RMs. The open-source code and data are available at https://github.com/dunzeng/MORE.

Summarization models often generate text that is poorly calibrated to quality metrics because they are trained to maximize the likelihood of a single reference (MLE). To address this, recent work has added a calibration step, which exposes a model to its own ranked outputs to improve relevance or, in a separate line of work, contrasts positive and negative sets to improve faithfulness. While effective, much of this work has focused on how to generate and optimize these sets. Less is known about why one setup is more effective than another. In this work, we uncover the underlying characteristics of effective sets. For each training instance, we form a large, diverse pool of candidates and systematically vary the subsets used for calibration fine-tuning. Each selection strategy targets distinct aspects of the sets, such as lexical diversity or the size of the gap between positive and negatives. On three diverse scientific long-form summarization datasets (spanning biomedical, clinical, and chemical domains), we find, among others, that faithfulness calibration is optimal when the negative sets are extractive and more likely to be generated, whereas for relevance calibration, the metric margin between candidates should be maximized and surprise--the disagreement between model and metric defined candidate rankings--minimized. Code to create, select, and optimize calibration sets is available at https://github.com/griff4692/calibrating-summaries

To be discoverable in an embedding-based search process, each part of a document should be reflected in its embedding representation. To quantify any potential reflection biases, we introduce a permutation-based evaluation framework. With this, we observe that state-of-the-art embedding models exhibit systematic positional and language biases when documents are longer and consist of multiple segments. Specifically, early segments and segments in higher-resource languages like English are over-represented, while later segments and segments in lower-resource languages are marginalized. In our further analysis, we find that the positional bias stems from front-loaded attention distributions in pooling-token embeddings, where early tokens receive more attention. To mitigate this issue, we introduce an inference-time attention calibration method that redistributes attention more evenly across document positions, increasing discoverabiltiy of later segments. Our evaluation framework and attention calibration is available at https://github.com/impresso/fair-sentence-transformers

Activation-based steering can personalize large language models at inference time, but its effects in educational settings remain unclear. We study persona vectors for seven character traits in short-answer generation and automated scoring on the ASAP-SAS benchmark across three models spanning two architectures. Persona steering lowers answer quality overall, with much larger effects on open-ended English Language Arts (ELA) prompts than on factual science prompts; interpretive and argumentative tasks are up to 11x more sensitive. On the scoring side, we observe predictable valence-aligned calibration shifts: evil and impolite scorers grade more harshly, while good and optimistic scorers grade more leniently. ELA tasks are 2.5-3x more susceptible to scorer personalization than science tasks, and the Mixture-of-Experts model shows roughly 6x larger calibration shifts than the dense models. To our knowledge, this is the first study to systematically examine the effects of activation-steered persona traits in educational generation and scoring, and the results highlight the need for task-aware and architecture-aware calibration when deploying steered models in educational settings.

Reliable probability estimates are critical in many machine learning applications, yet modern classifiers are often poorly calibrated. Post-hoc calibration provides a simple and widely used solution, but the large number of proposed methods, combined with small-scale and inconsistent evaluations, makes it difficult to determine which approaches are truly effective in practice. We introduce a large-scale, standardized benchmark for post-hoc calibration, covering nearly 2000 experiments across tabular and computer vision tasks, including binary, multiclass, and large-scale classification settings. Our benchmark aggregates predictions from a diverse set of classical models, modern deep learning architectures, and foundation models, and provides unified, reproducible implementations of dozens of calibration methods within a common evaluation framework. We argue that Post-Hoc Improvement (PHI) in proper scoring rules offers a principled alternative to traditional calibration error estimators for comparing post-hoc methods, capturing both calibration quality and potential degradation to the model's predictive performance. Using this framework, we conduct the most comprehensive empirical study of post-hoc calibration to date. Our results reveal consistent patterns across domains: smooth calibration functions outperform binning-based approaches, dedicated multiclass methods are essential in high-dimensional settings, and generic machine learning models are not competitive without calibration-specific design. To facilitate future research, we release all data, code, and evaluation tools, providing a plug-and-play benchmark for developing and comparing calibration methods.

Large language model (LLM)-based agents have shown promise in tackling complex tasks by interacting dynamically with the environment. Existing work primarily focuses on behavior cloning from expert demonstrations or preference learning through exploratory trajectory sampling. However, these methods often struggle to address long-horizon tasks, where suboptimal actions accumulate step by step, causing agents to deviate from correct task trajectories. To address this, we highlight the importance of timely calibration and the need to automatically construct calibration trajectories for training agents. We propose Step-Level Trajectory Calibration (STeCa), a novel framework for LLM agent learning. Specifically, STeCa identifies suboptimal actions through a step-level reward comparison during exploration. It constructs calibrated trajectories using LLM-driven reflection, enabling agents to learn from improved decision-making processes. We finally leverage these calibrated trajectories with successful trajectories for reinforced training. Extensive experiments demonstrate that STeCa significantly outperforms existing methods. Further analysis highlights that timely calibration enables agents to complete tasks with greater robustness. Our code and data are available at https://github.com/WangHanLinHenry/STeCa.

Task semantics can be expressed by a set of input-to-output examples or a piece of textual instruction. Conventional machine learning approaches for natural language processing (NLP) mainly rely on the availability of large-scale sets of task-specific examples. Two issues arise: first, collecting task-specific labeled examples does not apply to scenarios where tasks may be too complicated or costly to annotate, or the system is required to handle a new task immediately; second, this is not user-friendly since end-users are probably more willing to provide task description rather than a set of examples before using the system. Therefore, the community is paying increasing interest in a new supervision-seeking paradigm for NLP: learning from task instructions. Despite its impressive progress, there are some common issues that the community struggles with. This survey paper tries to summarize and provide insights into the current research on instruction learning, particularly by answering the following questions: (i) What is task instruction, and what instruction types exist? (ii) How to model instructions? (iii) What factors influence and explain the instructions' performance? (iv) What challenges remain in instruction learning? To our knowledge, this is the first comprehensive survey about textual instructions.

Predicting not only the target but also an accurate measure of uncertainty is important for many machine learning applications and in particular safety-critical ones. In this work we study the calibration of uncertainty prediction for regression tasks which often arise in real-world systems. We show that the existing definition for calibration of a regression uncertainty [Kuleshov et al. 2018] has severe limitations in distinguishing informative from non-informative uncertainty predictions. We propose a new definition that escapes this caveat and an evaluation method using a simple histogram-based approach. Our method clusters examples with similar uncertainty prediction and compares the prediction with the empirical uncertainty on these examples. We also propose a simple, scaling-based calibration method that preforms as well as much more complex ones. We show results on both a synthetic, controlled problem and on the object detection bounding-box regression task using the COCO and KITTI datasets.

Large language models (LLMs) have demonstrated remarkable proficiency in in-context learning (ICL), where models adapt to new tasks through example-based prompts without requiring parameter updates. However, understanding how tasks are internally encoded and generalized remains a challenge. To address some of the empirical and technical gaps in the literature, we introduce an automated formulation for encoding task information in ICL prompts as a function of attention heads within the transformer architecture. This approach computes a single task vector as a weighted sum of attention heads, with the weights optimized causally via gradient descent. Our findings show that existing methods fail to generalize effectively to modalities beyond text. In response, we also design a benchmark to evaluate whether a task vector can preserve task fidelity in functional regression tasks. The proposed method successfully extracts task-specific information from in-context demonstrations and excels in both text and regression tasks, demonstrating its generalizability across modalities. Moreover, ablation studies show that our method's effectiveness stems from aligning the distribution of the last hidden state with that of an optimally performing in-context-learned model.