NeurIPS 2026 Papers

"Hearing is never like being.'' —Persian Proverb Large language models have become the default advisors for life-critical human problems. While they democratize access to personal counseling, they suffer from a silent foundation bias: the internet is a record of people with the freedom to act. Current benchmarks assume users have this same agency, ignoring the reality of those in war zones or navigating statelessness. These long-tail experiences are not only missing from training data, but authentic evaluation data to measure them is equally scarce. We introduce HearSayBench, a human-verified dataset of 400 scenarios from respected archives like the United Nations, covering 80 regions across three specific barriers: social, personal, and environmental. Our work uses Capabilities Approach to test if a model can distinguish between what a person is legally promised and what they are actually free to do in their specific environment. While these models may have "heard'' about global inequality during training, we find that this knowledge is merely hearsay. Across 11 frontier and open-weight models, we identify a systemic %37 performance drop between situational comprehension and structural reasoning. When faced with the most vulnerable users, models consistently offer a "Checklist of Impossible Things'': polite, fluent advice that is physically impossible or legally suicidal to follow. Ultimately, we show that the true digital divide is no longer about access to technology, but about whether an AI can recognize the reality of your life.

Aligning autonomous agents with human intent remains a central challenge in modern AI. A key manifestation of this challenge is reward hacking, whereby agents appear successful under the evaluation signal while violating the intended objective. Reward hacking has been observed across a wide range of settings, yet methods for reliably measuring it at scale remain lacking. In this work, we introduce a new evaluation paradigm for measuring reward hacking. Whereas prior studies have primarily analyzed it post hoc by inspecting agent trajectories, we instead embed detectable reward hacking opportunities directly into environments. This makes their exploitation verifiable by design, enabling deterministic and automated measurement of whether and how agents exploit such vulnerabilities. We instantiate this approach in TextArena and release Hack-Verifiable TextArena, a testbed in which reward hacking can be measured reliably. Using this benchmark, we analyze reward hacking behavior across language models in diverse environments and settings.

Reliable generative AI models critically rely on expert human annotations to evaluate output quality, yet these "gold'' labels are expensive to collect and limited in quantity. Organizations thus often turn to collecting vast but noisy "silver'' labels from crowdsourced workers or vendor annotators as proxies for gold labels. Because gold remains the evaluation target, naively aggregating noisy silver labels may introduce bias, and estimators built on sparsely observed gold labels may have high variance to resolve the model performance gaps that guide practical decisions. Model evaluation has become an ongoing operational practice rather than a one-time exercise, with evaluation rounds repeating across model versions, releases, and content domains. A natural question is whether the previous historical evaluation data can be used to improve each new round of evaluation. We introduce HERO (History Enhanced RObust model evaluation), a novel framework that uses historical data to reduce both bias and variance in model performance evaluation. HERO calibrates silver labelers' performance learned from historical gold annotations, and stabilizes the resulting estimator by anchoring it to covariate information measured with high precision in the historical data. HERO can be broadly applied across multiple common evaluation tasks, and remains valid when only a subset of historical labelers appears in the current round. We establish conditions under which the bias and variance reductions hold, showcase HERO's performance in simulation studies, and demonstrate its effectiveness on real-world model evaluation benchmarking datasets.

General-purpose agents perform tasks in unfamiliar environments without domain-specific manual customization. Yet no study has systematically measured how agent architecture shapes performance across heterogeneous protocols and diverse unfamiliar environments. This is the first systematic study, comparing tool-calling, MCP, code-generation, and CLI agents on the same benchmarks with the same models. Two gaps blocked such a study: existing harnesses require per-benchmark wiring or fixed protocol classes (web for BrowserGym, CLI for Harbor), and benchmarks themselves expect human-authored prompts, context, and integration glue. To enable this study, we contribute (1) a unifying protocol that bridges existing benchmark and agent protocols; (2) an evaluation harness that surfaces any benchmark to any general-purpose agent and backbone model; and (3) the first Open General Agent Leaderboard of agent configurations, a full factorial over 5 agent architectures × 5 backbone LLMs (three closed-source, two open-weight) × 6 benchmarks spanning software engineering, customer service, deep research, and personal assistance. We find that (i) general agents adapt to every tested domain without per-domain customization; (ii) agent architecture choice swings results by up to 12pp within a single model, yet backbone model choice dominates overall performance; (iii) on 4 of 6 tested benchmarks, top general agents are indistinguishable from the leading heavily-customized domain-specific agents; (iv) open-weight models tested exhibit "generality sinks" absent from frontier closed-source models: they consistently collapse on specific agent architectures or benchmarks. Code, harness, leaderboard, and traces will be released upon acceptance.

Evaluating document OCR systems traditionally requires ground-truth (GT) text annotations aligned to each recognised element. Producing such annotations demands expert human labour, scales poorly with corpus size, varies between annotators on ambiguous content, ages quickly as document distributions and OCR systems evolve, and is often infeasible for sensitive or proprietary material. These constraints prevent continuous OCR quality monitoring at production scale and force practitioners to rely on small audited samples or proxy heuristics. To overcome these inherent limitations of ground-truth-based evaluations, we present a new class of reference-free OCR metrics based on render-and-compare that do not require any ground-truth annotations: OCR output is rendered back into a page image and the result is compared visually against the masked original using no-reference or learned metrics. The underlying principle is straightforward: the original page image already serves as a pixel-level ground truth for the content the OCR system tries to extract (text, formulas, and tables), so visual agreement between the original and a re-rendering of the OCR output is itself a proxy for recognition fidelity. To our knowledge, no prior work has systematically benchmarked reference-free visual metrics as proxies for whole-document OCR quality at scale. We conduct the first such study, evaluating 147 method implementations spanning 15 broad categories of techniques on 1,355 OmniDocBench pages across 5 OCR-output variants (text-only, formula-only, table-only, combined, combined without image masking). We measure every method by Spearman rank correlation against ground-truth-based metrics (edit distance for text and formulae, TEDS for tables), quantifying whether a reference-free metric correctly orders pages by OCR quality without any annotations. A naïve CLIP-cosine full-page similarity baseline achieves Spearman correlation $\rho = 0.339$; our best composite method reaches $\rho = 0.494$ (mean across the 5 variants; significance $p < 0.001$, computed over $N = 1{,}355$ pages), with a per-variant peak of $\rho = 0.605$ on the formula-only variant. Both numbers indicate that reference-free OCR quality estimation is genuinely viable: a meaningful page-level ranking signal can be recovered without any ground-truth annotations, and the gap to the ground-truth-based ranking is consistent with the inherent noise of a multi-domain document corpus where OCR errors are heterogeneous across text, formula, and table content. We also discuss inherent limitations that bound what any reference-free metric can achieve in this setting. To support reproducibility and for future research in this direction we are releasing three public HuggingFace datasets (rendered page pairs, per-element OCR log-probabilities, document-similarity triplets), and a trained document-similarity (DocSim) LoRA head (a DreamSim-style recipe adapted to document page reconstructions) along with the associated code.

Evaluating new large language models typically requires costly human annotation campaigns at scale. LLM-as-a-judge offers a cheaper alternative, but judge scores carry systematic errors — such as position bias, self-preference, or intransitivity — that can strongly miscalibrate the resulting rankings. We quantify the resulting judge--human disagreement at two complementary levels. At the local level, we estimate per-battle uncertainty from the judge's own score differences by propagating calibrated win probabilities rather than hard labels into the Bradley--Terry procedure. This alone provides a drastic improvement to Elo estimation accuracy, bringing LLM-derived ratings within $17.9$ Elo MAE of human-derived ones when averaged over 55 held-out models on LMArena. At the global level, we apply split conformal prediction to the residual gap between LLM-derived and human-derived Elo ratings across held-out models, producing prediction intervals with distribution-free marginal coverage guarantees that account for irreducible LLM--human disagreement. Together, these two layers yield a low-cost evaluation tool that provides developers with calibrated Elo estimates and honest uncertainty bounds, without access to large-scale human annotations.

This position paper argues that the evaluation of modern visual processing systems should no longer be driven primarily by single-metric image quality assessment benchmarks, particularly in the era of generative and perception-oriented methods. Image restoration exemplifies this divergence: while objective IQA metrics enable reproducible, scalable evaluation, they have increasingly drifted apart from human perception and user preferences. We contend that this mismatch risks constraining innovation and misguiding research progress across visual processing tasks. Rather than rejecting metrics altogether, this paper calls for a rebalancing of evaluation paradigms, advocating a more human-centered, context-aware, and fine-grained approach to assessing the visual models' outcomes.

Evaluation practices in machine learning are increasingly objects of scientific study in their own right: their assumptions, statistical properties, and failure modes determine which research conclusions can be trusted. We audit one such practice - the use of aggregate-subset Membership Inference Attacks (MIAs) for evaluating machine unlearning - across 8 dataset-architecture pairs, 4 unlearning methods, and over 500 experimental runs, identifying five failure modes: (1) limited and regime-dependent identifiability between baseline and retrained models, with KS-test indistinguishability ($p > 0.05$) in $\approx$42\% of experiment--task pairs; (2) regime-dependent signal-to-noise collapse; (3) monotonicity violations in 5 of 8 experiments under graded forgetting; (4) inter-task rank inconsistency, and (5) regime confounding by model training quality. To quantify these failures, we develop the Membership Inference Attack Unlearning Score (MIAU), a normalized diagnostic framework integrating three MIA comparisons as gap closure fractions between baseline and retrained references. We formalize four validity properties any unlearning metric should satisfy. Our findings indicate that aggregate-subset MIAs are reliable unlearning diagnostics only in narrow high-memorization regimes, and motivate complementary evaluation paradigms, alongside MIA-based protocols.

NIGHTS is a recent benchmark for human perceptual similarity over objects, scenes, and composition. DreamSim, the supervised metric introduced with NIGHTS, and later supervised metrics trained on the same data report large gains over self-supervised baselines. The standard NIGHTS evaluation uses the unanimous-agreement subset of a released 100,000-triplet corpus, leaving weak-majority and split-vote triplets outside the headline benchmark. We evaluate eleven perceptual similarity methods on the full released corpus, analyzing unanimous, weak-majority, and split-vote triplets separately. On the 64,420 non-tie triplets, cosine distance on a frozen self-supervised backbone reaches 74.1% accuracy without training, LoRA fine-tuning reaches 77–79%, and the DreamSim ensemble reaches 80.7%. A split-half reproducibility estimate from the triplet vote counts shows that majority labels reproduce only 70.1% of the time overall, and only 33–50% on weak-majority triplets. The extra accuracy comes from matching the collected votes, not from showing that the metric would better match a new set of human judgments. The evaluation target is also unstable across task formulations: NIGHTS’ 2AFC similarity labels and JND same/different labels disagree on 51–64% of shared stimuli. Fine-tuning changes the feature basis as well, increasing foreground-only fragility and making DreamSim the most fragile method under rotation, color, and blur. NIGHTS supports a narrower conclusion: cosine distance on self-supervised features captures the stable signal currently measurable in these labels. Additional claims of human alignment require a redesigned benchmark with more votes per triplet, axis-specific similarity questions, per-subject statistics, and ablation tests.

Human labels are important supervisory signals for preference-based reinforcement learning, but their collection is often constrained by practical considerations that standard training objectives tend to ignore. For instance, Direct Preference Optimization (DPO) usually treats observed preference labels as clean, even though real annotation pipelines rely on heterogeneous annotators whose reliability varies across individuals and with the difficulty of each comparison. This setting introduces three challenges: (1) Noisy labels introduce systematic bias into the preference optimization objective and need to be rigorously corrected. (2) The allocation of pairwise comparisons to different annotators is a design problem that directly affects the statistical efficiency of the training pipeline. (3) Non-random routing creates selection bias in annotator evaluation, complicating downstream decisions such as compensation. We address these challenges within a unified framework for DPO that combines annotator-aware posterior correction with information-guided annotator routing. Posterior correction leverages an instance-dependent noise model to infer latent ground-truth preferences from noisy labels, mitigating bias in the DPO objective. Annotator routing is formulated as an optimal experimental design problem that allocates comparisons to annotators according to their expected statistical information under a fixed annotation budget. To support reliable annotator evaluation under non-random routing, we introduce a doubly robust estimator that corrects for selection bias. Semi-synthetic experiments on the UltraFeedback dataset show that our approach improves ground-truth preference recovery and downstream AlpacaEval 2 performance over baselines, with consistent gains across multiple open-source LLM model architectures. Annotator reliability-estimation experiments further show that the doubly robust estimator substantially reduces mean squared error relative to other baselines, demonstrating its value for annotator evaluation.

Explainable AI (XAI) offers far-reaching consequences for human-AI collaboration, but the fundamental question of how best to evaluate these systems has proven frustratingly elusive. Methodologies typically revolve around either proxy metrics which tell us little about the utility of an explanation, or human evaluation which---while generally considered the gold standard---is expensive, difficult to scale, and challenging to reproduce. In this paper, we investigate if Multimodal Large Language Models (MLLMs) can act as surrogate human participants to automate the evaluation of explainable AI. To do this, we first propose a ``purpose-grounded'' evaluation framework with specific metrics applicable to both MLLMs and human users alike. Then, we offer some conceptual insights as to why and when we can trust these models to automate this process. Lastly, we conduct extensive empirical testing across five purpose-grounded areas and nine popular MLLM APIs. Results indicate that all models trend positively, but OpenAI models (particularly GPT-5 mini) are the best aligned at replicating user responses, and for complex visual reasoning tasks more advanced modes (like GPT-5) perform best, which would be the recommended choices in practice. We expect our insights not to replace user testing, but rather to help XAI researchers complement their current evaluation methodologies and move the field away from heavy reliance on proxy metrics for measuring explanation utility.

Adaptive immune receptor repertoire sequencing data has emerged as a promising potential modality for disease diagnosis, relying on computational methods to analyze T-cell receptor (TCR) sequences from an individual’s blood sample. Published methods rely on different cohorts, data preprocessing pipelines, and evaluation metrics, making direct comparison across methods challenging. We have developed BenchRep-T, a unified benchmark that standardizes multiple publicly available TCR repertoire datasets and evaluates nine computational approaches, spanning statistical enrichment of shared sequences, feature-engineered ensembles, deep learning, and graph-based clustering. BenchRep-T evaluates methods on four tasks: disease classification across conditions, performance scaling under restricted sequence-sampling depth, recovery of known antigen-specific driver sequences, and evaluation of sensitivity to demographic confounding. Under controlled evaluation, simple baselines prove competitive, with tree-based models trained on V- and J-gene usage and short sequence motifs approaching the classification performance of more complex methods. Our findings underscore the complexity of modeling TCR repertoire data, and show that no single method dominates across all tasks. BenchRep-T provides a framework for rigorous and reproducible evaluation of TCR repertoire classification methods to accelerate the development of immune repertoire-based diagnostics.

AI evaluation pipelines often produce conflicting judgments rather than clean labels. In pairwise LLM evaluation, this conflict is especially visible: disagreement can arise from ambiguous items, underspecified rubrics, heterogeneous or unstable human raters, or an LLM judge whose verdict changes when the response order is swapped. We propose BayesJudge, an online Bayesian meta-evaluation layer for conflicting human-LLM judgment streams. For each comparison, BayesJudge estimates a panel-relative posterior verdict distribution over the two responses, with a tie or ambiguity state when such labels are available. At the same time, it estimates rater-specific human confusion matrices and LLM presentation-order bias. The method uses tie-open labels to keep ambiguity observable and paired order-swapped judge calls to separate response quality from presentation effects. We formulate the exact online posterior recursion and use a scalable Rao--Blackwellized assumed-density SMC approximation for streaming inference. Controlled synthetic experiments demonstrate recovery of prespecified evaluator parameters and illustrate two protocol-level identifiability mechanisms: tie-open labels expose ambiguity mass, and order-swapped paired judgments separate item preference from position bias. On real-world SummEval dataset, BayesJudge successfully detects systematic presentation-order effects in LLM judge outputs, infers distinct expert and crowdworker behavior signatures without rater metadata, and produces posterior uncertainty estimates that correlate with human disagreement. Our code is available at https://anonymous.4open.science/r/BayesJudge-0879.

Studies of human reasoning have shown that people are typically stronger at evaluating reasoning than producing it from scratch. In contrast, large reasoning models (LRMs) are trained to excel at producing long chains of reasoning to solve complex problems. How then do LRMs perform at evaluating reasons? We investigate this with the Valid-Answer-Invalid-Reasoning (VAIR) dataset: math problems and solutions with trivial reasoning flaws but valid answers, designed to isolate reasoning evaluation from the confound of reasoning production. Unlike humans, who we find are only 6\% worse at grading than solving such problems, we find a substantial production-evaluation gap in LRMs: frontier models score as low as 48\% when evaluating VAIR solutions, despite near-perfect solution production. Why this enigma? Through chain-of-thought (CoT) analysis, we find evidence of an answer confirmation bias: LRMs often produce then check for the correct answer instead of carefully verifying each step, fabricating rationalizations even when noticing anomalous reasoning. Linear probes corroborate this, showing that while LRM activations encode some representation of valid reasoning, they fail to robustly represent VAIR solutions as invalid. Causal patching of the final answer's representations causes LRM verdicts and activations to flip, demonstrating that answer validity is responsible for models' confirmation biases. These findings indicate an outstanding limitation in dominant approaches to reasoning training, which incentivize LRMs to produce and confirm reasoning towards correct answers, but not to robustly evaluate the underlying reasons.

Human players in cooperative games often rely on conventions: reusable patterns of play that coordinate well when used by the whole team, but can fail otherwise. Evaluating agents in this setting is difficult because human demonstration datasets are usually unlabeled: trajectories do not identify which player, group, or convention produced them. As a result, standard proxy-based evaluations can miss important population diversity. In particular, proxy bots produced by regularised self-play around a pooled behavioural cloning policy can remain concentrated around a single dominant mode of play. We introduce a pipeline for recovering and evaluating this hidden convention diversity from unlabeled demonstrations. Our method first learns latent convention structure from the dataset, then selects a compact set of distinct but plausible conventions, and finally reconstructs each selected convention into a high-performing playable proxy using KL-regularised PPO. A key practical challenge is that KL-regularised PPO is sensitive to the KL penalty, which is often tuned using online interaction. We provide an empirical offline calibration procedure for choosing this KL penalty in the PPO+KL reconstruction settings studied here. On Hanabi, we find that the BC-derived proxy suite is highly concentrated, and that the relative ordering of agents changes when evaluation uses the latent proxy suite instead.

Modern deep learning commonly relies on AdamW with prescribed learning rate schedules, but recent works challenge both components: Schedule-Free optimization removes explicit schedules via iterate averaging, and Muon improves the update geometry by orthogonalizing momentum for matrix parameters. Despite Muon's strong empirical performance, its underlying mechanism remains partially understood. We study Muon through the river-valley loss landscape, where useful training progress occurs along a flat, low-curvature bulk subspace (the river), while high-curvature dominant directions form steep valley walls that induce oscillations. We empirically show that while Muon's orthogonalization accelerates river progress by increasing the bulk component, it also amplifies dominant-direction noise, causing oscillatory trajectories. Building on this, we propose Anytime MUon with Stable gradient Evaluation (AMUSE), which integrates Muon's rapid bulk progress with the stabilizing effect of Schedule-Free averaging. AMUSE uses a time-varying interpolation coefficient that initially evaluates gradients near the fast Muon sequence for rapid adaptation, then gradually shifts toward the stable averaged sequence to suppress valley-wall oscillations. As a result, AMUSE requires no learning rate schedules and supports anytime training. Across vision tasks and large language model pretraining, AMUSE consistently improves the performance-iteration Pareto frontier over (Schedule-Free) AdamW and Muon.

Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes overfitting and performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators imposes a ceiling on search performance. We introduce AIRA$_2$, which addresses these bottlenecks through three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug interactively. On MLE-bench-30, AIRA$_2$ achieves a mean Percentile Rank of 81.5% at 24 hours and 83.1% at 72 hours, outperforming the strongest baseline, which achieves 72.7%. On AIRS-Bench, AIRA$_2$ exceeds human state-of-the-art on 6 out of 20 diverse research tasks. Ablations confirm that each architectural component is necessary, that performance follows a predictable scaling law that transfers across LLM backbones, and that the "overfitting" reported in prior work was driven by evaluation noise rather than true data memorization.

There are now thousands of measurement instruments used to evaluate language models, and the resulting scores are routinely used to support consequential claims about capabilities and risks. Yet, the evidence linking benchmark scores to constructs (i.e., abstract concepts such as reasoning ability or safety) is often implicit, incomplete, or absent. We introduce AI Construct Lexis, a versioned ontology and interactive artifact that extracts and organizes relationships among constructs, behaviors, and measurement instruments as claimed in the literature for LLM evaluation. Drawing on over 8,337 published papers from selected top AI venues, we use language models to extract constructs, definitions, behavioral indicators, and measurement instruments, yielding 977 distinct constructs, 1,694 measurement instruments, 830 behavioral indicators, and 5,900 relationships. The Lexis captures relationships claimed in the published literature, not necessarily validated relationships; valid score interpretation still requires empirical evidence that the instruments capture the constructs they are used to support. By exposing the field’s implicit measurement assumptions, the Lexis enables their inspection, testing, falsification, and refinement into evidence-based relationships. On expert-annotated samples, the extraction pipeline achieves high agreement on extracted concepts while revealing substantial ambiguity in the boundaries of constructs and behaviors. By making these relationships explicit, AI Construct Lexis provides infrastructure for a more rigorous and cumulative science of AI evaluation.