Papers for
ai model evaluators
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Generative models trade feature learning against mode collapse risks
The balance between feature learning and collapse in dynamical systems-based generative models
Abstract: Over the past decade, a growing body of work has established that implicit biases in learning dynamics fundamentally shape the solutions found by neural networks, governing their learned features and generalisation performance. However, despite these strides in deep learning theory, far less is known about representation learning in generative models based on dynamical systems, from normalising flows to diffusion. Importantly, empirical evidence suggests that continuous-time generative models can be prone to mode collapse, yet a learning theory explaining such biases remains elusive. To address this, we use tools from dynamical systems and operator theory to express the gradient flow of a general class of stochastic differential equations as a composition of linear operators, allowing us to relate collapse to low-rank biases in generative models. In variational inference and flow-matching tasks, we show that this bias can manifest in two different mechanisms: i) where collapsed models are stable fixed points of the gradient flow, or ii) in sequential feature learning, where smaller modes take progressively longer to be learned, producing an effective collapse for a fixed training budget. Together these findings point to a trade-off between feature learning and collapse, suggesting an operator view as a step towards a unified framework for understanding the learning dynamics of dynamical systems-based generative models.
Calibrated readings let users measure neurons in ai models reliably
Calibrated concept profiles: making the reading of a model's neurons a measurement
Abstract: Interpretability has largely moved from reading a model's neurons directly to reading a second model trained on the first: a probe, a sparse autoencoder, a crosscoder. The move answered real faults -- mixed meanings, thresholds without a unit, readings that did not transfer from one study to the next -- but gave up what a direct reading alone provides: nothing trained in between, the neurons themselves named, and a reading that can be repeated and compared. Here the direct reading is made a measurement. A neuron's activation varies with the prompt, so its reading is a distribution over a population of stimuli. We declare that population: every factor around a target construct is drawn at its rate in a reference corpus and logged. From the declared distribution we generate a battery of prompts, and draw it twice. From the battery we compute the calibrated concept profile, one number per neuron, stated in units of the neuron's own spread on ordinary input; the two draws give every number a noise unit. Before the profile is read, a fixed set of checks licenses the battery on the model, and the smallest change the profile can resolve across training checkpoints is measured. Realised for Python on OpenCoder-1.5B and followed across checkpoints spanning two trillion training tokens, the profile gives what the alternatives do not report: a unit on every number, a chance level for every comparison, and which neurons carry a construct at every checkpoint. The battery, the code and the data behind every result are released.
Large language models cannot reliably report their financial knowledge over time
LLMs Don't Know What They Know: Benchmarking Temporal Self-Knowledge for Financial Backtesting
Abstract: Financial backtests of large language models (LLMs) are meaningful only if the model did not already know the outcomes it is asked to predict. Practitioners guard against this by trusting a vendor's stated knowledge cutoff, by asking the model what it knows, or by instructing it to answer as of a past date. To verify whether any of these safeguards can be trusted, we introduce FinDendro, a benchmark of 2,226 dated U.S. facts from 1999 to 2026, covering monthly macroeconomic series, month-end stock prices and Federal Reserve decisions. Each fact yields four prompts: a value question, Recall and Predict questions that pair each candidate value with a claim of knowing or guessing without a declared date, and a Predict question with a historical declared date. Controls for persistence and dominant outcomes keep chance accuracy near 0.27. Results show that for most frontier models, self-reports understate knowledge. GPT-5 answers 76% of controlled facts correctly but claims to know 38%, and answers 61% of the facts it says it does not know. Knowledge claims can persist beyond the measured knowledge boundary, by up to seven quarters, and a declared date removes most of the claims of prior knowledge while leaving accuracy unchanged. Backtests inside a model's training window therefore cannot be certified leak-free by self-report, instruction or stated cutoff. These findings motivate behavioral measurement of temporal knowledge boundaries, and FinDendro provides a benchmark and procedure for selecting backtest periods based on demonstrated factual knowledge.
Changing the judge model alters AI chatbot rankings significantly
Swapping the Judge Changes the Ranking: A Preregistered Study of Cross-Provider Replication in LLM-Judged System Comparisons
Abstract: Practitioners who compare systems with an LLM judge and have no human labels often check a conclusion by repeating it with a judge from another provider. We test how far that check can be trusted. In a preregistered study (with a disclosed post-registration reduction of the judge set), five judge models from three providers re-judged 4,000 Chatbot Arena battles (nine systems, 33 system pairs) that carry human votes, each in both response orders. Agreement with the human vote was low at the verdict level (median Cohen's kappa = 0.22), and system rankings did not survive a change of judge: Kendall's tau with the human ranking ranged from 0.17 to 0.89 (median 0.67), so the preregistered hypothesis that rankings remain stable was not supported. Conclusions replicated by two low-cost judges from different providers matched the human conclusion more often than single-judge conclusions (+0.10 precision), but the gain came from filtering out one poorly aligned judge, and replication involving that judge left a conclusion for only 35% of comparisons. Judges that agreed with humans reached a stable conclusion with no more verdicts than human votes required; the others failed to converge on 26-70% of system pairs at the sample sizes tested (up to 100 verdicts per pair). Disagreements with the human vote co-occurred across judges far more often than independence predicts, although a single human vote per battle cannot separate shared judge bias from voter noise.
Checking test state identity improves vision language action model comparisons
Verifying Test-State Identity in Paired Evaluation of Vision-Language-Action Models
Abstract: Simulation-based VLA evaluation often treats matching task, seed, and episode identifiers as sufficient evidence that two policies faced identical test conditions. Under specific vectorized reset and autoreset conditions, different termination timing can change subsequent initial-state assignment, so the same episode index need not mean the same test state. We present an evaluation-integrity framework for paired comparison - invariants I1-I5, provenance logging, optional scene fingerprints, and an automated checker - and a controlled LeRobot 0.6.1 LIBERO harness study with emulated controllers. Four evidence levels are kept separate: (A) mixed-horizon Sync wave-1 IDs [20, 5, 22, 7] vs [4, 5, 6, 7]; (B) homogeneous separate-job Sync control (16 cells, 0 ID/fingerprint mismatches); (C) Async+Freeze batch-size map differences (12 I3 mismatches; bs=8 incomplete; mechanism unisolated); (D) a synthetic McNemar illustration only. Controllers were emulated and no real-VLA shared-vector experiment was completed. We recommend checking declared/resolved init-state identity and provenance before paired statistics. We do not claim that published VLA leaderboards are generally invalid.
Audio visual benchmark tests evidence based reasoning in omni language models
EvaOmni: An Evidence-Grounded Benchmark for Audio-Visual Reasoning
Abstract: Omni LLMs can naturally locate key audio and visual evidence and integrate information across modalities for audiovisual reasoning. Existing audio-visual reasoning benchmarks that mainly rely on multiple-choice questions usually score models solely based on answer correctness, making it difficult to assess whether a model truly forms a reasoning chain from evidence to the correct answer. In addition, textual or unimodal shortcuts may allow models to answer correctly without analyzing the necessary audio-visual evidence, further weakening the reliability of evaluation benchmarks. To address these issues, we introduce EvaOmni, an evidence-grounded benchmark containing 392 videos and 1,027 purely human-annotated question-answer pairs, covering 15 capabilities and three levels: retrieval and localization, basic understanding, and complex reasoning. We design a series of rigorous annotation procedures in which experts annotate audio and visual evidence and construct video-based distractors; humans review whether questions meet strict inspection criteria, and finally models filter textual and unimodal shortcuts. Beyond high-quality question-answer pairs, we require models to simultaneously return the answer and supporting evidence intervals in order to evaluate the complete reasoning chain. We propose the Coverage Grounded Score (C.G.) to measure how much annotated evidence a model identifies, and the Bottleneck Grounded Score (B.G.) to penalize poor temporal localization in either modality; both are incorporated into answer correctness. Across 7 open-weight and 17 proprietary Omni LLMs, Gemini3.7-Flash achieved the highest overall accuracy (79.4%), C.G. (70.8%), and B.G. (35.1%). At Level 3, it reached 79.1% accuracy, but its B.G. was only 12.3%. These results show that even the most advanced models still have room for development in grounded audio-visual reasoning.
Teacher advantage in on-policy distillation lies in targeted corrections
Demystifying On-Policy Distillation: Where Does the Teacher's Advantage Live?
Abstract: On-policy distillation (OPD) is increasingly studied as a cheaper alternative to reinforcement learning, yet it remains a form of knowledge distillation (KD) that relies on teacher supervision. We revisit OPD from the KD perspective by asking where the teacher's advantage lies along student-generated trajectories. Controlled token flips at teacher-student disagreements directly measure the accuracy gained from teacher corrections. Single-token flips yield a modest average gain on selected failed trajectories, whereas repeated flips raise accuracy from 28.4% to 74.5%, close to the teacher's 76.9%. Gains continue to accumulate with more flips, but mass- and KL-based selection outperform random flips by 15-18 percentage points at counts of similar magnitude. Selective intervention remains useful after OPD, but greater argmax agreement does not consistently improve accuracy across model pairs. We advocate training schedules that learn more valuable teacher corrections while maintaining distributional stability. Controlled flips and fork-resolution analysis evaluate OPD beyond accuracy by measuring which teacher preferences the student acquires and what further correction can recover.
Hallucination detection labels differ based on faithfulness versus factual correctness
The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation
Abstract: In recent years, several methods for detecting when large language models (LLMs) hallucinate have been developed. These methods are often benchmarked with open-domain question answering (QA) datasets containing questions and corresponding short reference answers. First, an LLM is used to generate answers to questions within the QA dataset. Then, some automated labeling strategy is used to label these answers as hallucinated or not by comparing them with the reference answers in the dataset. This evaluation setting creates a methodological ambiguity between two criteria: reference faithfulness (whether the answer is fully supported by the reference) and factual correctness (whether the answer is free from contradictions and factually false specific claims). In practice, automated labelers may apply the former criterion even when the intended target is the latter. We study this potential criterion mismatch using 900 human-labeled question-answer pairs spanning three commonly used QA datasets and three generator models, with labels targeting answer-level factual correctness. We evaluate lexical similarity metrics, a reference-entailment NLI baseline, and seven LLM judges under controlled prompt variants as automated labelers. Our experiments reveal substantial disagreement both among automated labeling strategies and between these labels and human annotations. Many strategies also exhibit strong directional error biases, and for most judge-generator pairs, replacing a faithfulness-oriented prompt with a factual-correctness prompt improves agreement with human annotations and reduces false-positive dominance, indicating that automated hallucination labels depend strongly on how the target criterion is specified. Label-source choice should therefore be considered a fundamental part of benchmark design and made explicit, validated, and matched with the benchmark goal.
Open contrastive decision model performs near chance on judge benchmarks
CLM-as-a-Judge: Evaluating an Open Contrastive Decision Model on Public Judge Benchmarks
Abstract: An open contrastive decision model is near chance as a judge on the hard public benchmarks: Contrastive-LM/CLM-v0.1-8B scores between 0.351 (best- of-four, chance 0.250) and 0.593 (pairwise, chance 0.500), is statistically indistinguishable from coin flipping on RM-Bench and JudgeBench, and answers every HaluEval item with one constant label, matching the trivial always-first baseline at 0.581. Judges with the same parameter count score far higher everywhere: a reward model reaches 0.764 to 0.976 and a generative judge 0.611 to 0.778, and every gap to CLM is significant after Benjamini-Hochberg correction. Two properties do work. Raw confidences are overconfident by up to +0.401, yet one pooled temperature fit on held-out calibration items repairs expected calibration error to at most 0.062, and the repaired confidence ranks the model's own errors above chance on three of six benchmarks. The decision order-flip rate is 0.0002 against 0.2188 for the generative judge, and the length-preference shift is -0.023 against -0.217. The confidence-gated cascade, however, escalates between 0.923 and 1.000 of items to the strong judge at the preregistered 0.97 retention bar: calibrated confidence about a near-chance judge has almost nothing to keep. The design: five public preference benchmarks and one hallucination benchmark with real labels, scored under a preregistration frozen before any test item was seen, against generative, reward-model, and trivial baselines, with per-item predictions released.
SOL measures how closely text generation matches data distributions
SOL: Measuring Gaps between Text Distributions by Double Sliced Wasserstein Metrics
Abstract: Evaluating text generation requires measuring how well the generated distribution matches the data distribution. For autoregressive models, this is done by the perplexity. Diffusion and flow-based language models can only provide a likelihood bound, whose tightness differs between model families. Sample-based substitutes such as generative perplexity with entropy do not consider the distribution fit. We propose SOL, a distance between text distributions. Each sequence is represented by the empirical measure of its hidden states under a fixed transformer and the distributions of these measures are compared by the double sliced Wasserstein distance. We prove that SOL is a metric if the transformer is injective. Experiments show that SOL detects distributional failures, recovers expected model trends, and provides stable sample-based estimates. We put forward SOL to fill the gap in the current evaluation protocol used for non auto-regressive models. As a first step we use SOL to re-evaluate a variety of models trained on OpenWebText.
Instruction following model scores vary widely with wording changes
When Verifiable Counts Depend on Wording: Auditing Wording Robustness in Instruction Following
Abstract: Verifiable instruction-following benchmarks often express each constraint through one fixed template. We test whether scores remain stable when the operational requirement is unchanged but its wording varies. We introduce WISE, a matched evaluation suite and reporting protocol instantiated on exact word count, keyword inclusion exactly once, and an inclusive 8--12 word range. Across 100 matched tasks, up to thirteen models from seven providers, and repeated generations scored over the complete visible output, wording alone produces substantial compliance shifts. In an avoidance-family panel, five avoidance and exclusion forms fall below the positive baseline, while constructional controls also shift compliance substantially: in the nine-model control panel, compliance is 54.9% for the original positive form, 48.2% for a longer positive form, 36.7% when the target appears later, and 33.8% for AVOID1. A strict JSON-structure probe shows wording sensitivity beyond counting, with a different direction of effect. Effect sizes, failure directions, weakest forms, and model rankings vary across realizations. Under the most disruptive exclusion form, the top-ranked model changes and 24.1% of strictly ordered model pairs reverse. Human validation further shows that unanimous agreement on an exact-count interpretation can coexist with substantially different model behavior. WISE supplements conventional scores with mean and worst-form compliance, wording gaps, failure profiles, and ranking stability.
Function level scores measure flag rate more than vulnerability detection quality
A Function-Level Vulnerability Score Measures Flag Rate More Than the Model: Protocol Effects on Paired Benchmarks
Abstract: Language models are increasingly evaluated as vulnerability detectors, and scores reported for similar models differ widely between papers. We measured how much of that difference evaluation protocol accounts for, with model outputs held fixed. In a paired test, a model must flag a vulnerable function and clear its version after a fixing commit. Three choices that published evaluations make differently were varied one at a time: metric, verdict extraction and output budget. Seven frontier and large open models were evaluated on five released pair benchmarks and a set pooled for this work under one protocol, and 61 open models of 1.5B to 36B parameters on the pooled set. Function-level F1 follows how often a model flags both functions of a pair (Spearman $+0.86$ over 42 combinations) and is nearly unrelated to pair-level correctness ($+0.16$). On the pair score, extraction changes a model's number by $+0.001$ at the median and budget by $+0.02$ with an interval through zero, whereas the model changes a benchmark's number by up to 0.18 and the benchmark a model's by up to 0.16; a function-level score therefore measures flag rate more than model. For 37 of 68 models the difference between correct and reversed pairs is within its 95% interval of zero, the value for a null model that flags each function at a fixed rate, while both-flagged and both-cleared rates exceed that null by 0.055 on median, and for 64 of 68 both functions of a pair receive one answer more often than independence predicts: verdicts are determined by the text common to both functions. On length-matched pairs a linear probe on activations separates 0.78 by within-pair ranking, against 0.64 for a tf-idf baseline and 0.5 for length; the generated verdict is near chance for three of six models and at 0.55 to 0.57 for the other three, and a prompted logit is at chance for all six.
Emoji rating bias hides true differences among language models
Two Emojis of Difference: What Multilingual Affective Generation Benchmarks Actually Measure
Abstract: We audit a multilingual affective generation benchmark eight instruction-tuned LLMs producing emoji summaries for 17,100 Bangla, English and Hindi sentences, with 6,960 human judgements and find its headline conclusions to be artefacts of the measurement instrument rather than properties of the systems. Treating annotators as a random rather than a fixed factor, no system differs significantly from any other ($F(7,14)=0.59$, $p=0.76$), although the conventional analysis declares 19 of 28 pairwise differences significant. Annotator identity explains far more rating variance than system identity, and the winning system changes whenever any single annotator is removed. The ordering that does emerge tracks output length: mean emoji count explains 78.7\% of between-system variance, and a within-item length-matched comparison over 2,599 pairs reverses the leaderboard. We further show that cross-provider anisotropy differences vanish under mean-centring, that per-language token costs change sign with the normalising unit, and that multi-view row-wise splits inflate macro-F1 by $3.1$ points and change the top-ranked system. In place of preference scoring we propose **emoji-affect decodability**, a reference-based probe whose rankings are stable to $\pm0.003$ macro-F1 across seeds.
Llm judge saves cost by escalating unsure cases for review
JEV-as-a-Judge: Accept When Confident, Escalate When Unsure
Abstract: LLM-as-a-judge enables evaluation across diverse tasks, but inference cost and confidence reliability become critical at scale. We study whether a decision-only judge can provide an economical first pass and identify when stronger evaluation is needed. Comparing jev-as-a-judge with sixteen generative and reward-model judges, with blinded human adjudication, we find it within three percentage points of a state-of-the-art LLM judge, our strongest comparator, on ordinary preference and evidence-grounded factuality at 0.36% of the comparator's fee. Larger gaps arise when judgments require checking a derivation or resisting an elaborately written wrong answer. On several benchmarks, JEV's gap to this comparator is concentrated in low-confidence decisions. A frozen cascade that accepts confident verdicts and escalates uncertain ones retains 99% of the comparator's accuracy at lower cost.
De-identified résumés still reveal ethnicity beyond language details
Beyond the Name: Demographic Leakage in De-Identified Résumés and Evaluation Artifacts in LLM Bias Audits
Abstract: De-identified résumé screening assumes that redacting explicit fields prevents ethnocultural inference; however, recent audits attribute residual leakage to declared languages. We investigate whether eliminating language fields resolves this leakage across nine open-weight models and 620 counterfactual résumés. By holding language attributes strictly identical, we isolate unstructured prose across five ethnocultural conditions and three cue-salience tiers. Target-group recovery averages 0.757 overall and saturates at 1.000 under high salience, demonstrating that non-language prose sustains demographic inference. Crucially, models diverge only under faint cues (0.086-0.690), establishing salience as an essential evaluation axis. Furthermore, pairwise LLM-as-a-judge outcomes are highly sensitive to evaluation design: forbidding ties yields an apparent selection-rate ratio of 0.39 alongside strong position and content effects, whereas permitting ties produces near-universal ties for most models ($\ge94\%$). Downstream scoring shows only very small between-condition differences, highlighting the need to distinguish demographic signals recoverable from résumé content from effects introduced by the evaluation protocol.