Daily Papers

A cascaded voice agent must decide, throughout each silence of its user, whether the user has finished their turn: answering too early cuts the user off, answering too late makes the agent slow. It often needs related decisions too: will the user soon go on, is speech during its own turn an interruption or a backchannel, is a sound speech at all. With open detectors, each extra decision costs another model or another call, and on the 14-language eot-bench they cut users off in more than a quarter of mid-turn pauses. Because they are trained on segmented utterances, this mismatch is the largest single source of error. We present VoxTurn, a 0.8B speech LLM that casts these decisions as multiple-choice questions, scores the hidden state of each option, and answers any set of them in one forward pass with a block-isolating attention mask, exactly as if asked one at a time. Trained on real conversational pauses with labels derived from their timing, and re-queried while a silence lasts, VoxTurn reaches 11.5% false cut-offs in English at a 300 ms latency budget, comparable to two commercial detectors, and a 14-language mean of 19.9%, the lowest among open models, at 8.3 ms per decision. In Pipecat and LiveKit it cuts off 17% and 9% of turns, and it ranks fifth of 24 in TurnBench end-of-turn detection without any TurnBench training data. The model, code and framework plugins are released.

People overthink; language models over-sample, and the extra effort can talk both into a worse answer. Reasoning systems answer a hard question by sampling it many times (test-time scaling), and the more they draw, the more often a correct answer turns up somewhere, so coverage, the fraction of problems with at least one correct try, climbs and appears to be progress. But a deployed system must return one answer, and choosing it, not knowing which try is right, is selection; selection is capped, and past a point extra samples only make the model surer of a confident mistake, even as every draw adds cost. The gap between climbing coverage and stalled selection, the identifiability gap, is the answer a model can produce but not pick. So the real question is not whether to sample but how far, and the answer is: not far. For picking an answer, the vote has already settled within a few dozen draws, the modal ceiling; for scoring a benchmark, sooner still, the correlation ceiling. Beyond that, extra draws cost compute and add nothing, and can even make the answer worse. This paper turns the cutoff into a single number, the effective number of samples, that any sampling run already reveals. The bottleneck is recognizing a right answer, not generating one.

Released Large Language Models (LLMs) are often paired with a claimed knowledge cutoff date, or the dates at which training data was gathered. Such information is crucial for applications where the LLM must provide up to date information. However, this statement only scratches the surface: do all resources in the training data share the same knowledge cutoff date? Does the model's demonstrated knowledge for these subsets closely align to their cutoff dates? In this work, we define the notion of an effective cutoff. This is distinct from the LLM designer reported cutoff and applies separately to sub-resources and topics. We propose a simple approach to estimate effective cutoffs on the resource-level temporal alignment of an LLM by probing across versions of the data. Using this analysis, we find that effective cutoffs often differ from reported cutoffs. To understand the root cause of this observation, we conduct a direct large-scale analysis on open pre-training datasets. Our analysis reveals two reasons for these inconsistencies: (1) temporal biases of CommonCrawl data due to non-trivial amounts of old data in new dumps and (2) complications in LLM deduplication schemes involving semantic duplicates and lexical near-duplicates. Overall, our results show that knowledge cutoffs are not as simple as they have seemed and that care must be taken both by LLM dataset curators as well as practitioners who seek to use information from these models.

Ambient AI scribes draft clinical notes, and published audits find their dominant error is omission: information the encounter established that the note fails to record. The standard check is an LLM judge: a second model reads the note against the transcript and flags problems. We ask whether judges detect omissions. Public corpora cannot supply the answer key: their clinician reference notes and transcripts are materially discrepant. Our benchmark has 500 single-error note pairs from audited fact sheets, 298 with a named fact certainly absent and 202 added-or-altered controls. Across eight judge designs, paired discrimination (the flawed note below its clean twin, 0.5 a coin flip) reads 0.79-0.94 on added or altered content and 0.50-0.63 on omissions. On single notes, no design flags omissions reliably more often than perfect notes. Wording changes, voting and GEPA prompt optimisation move the operating point without creating usable detection. Restructuring the task recovers it: list the facts the transcript establishes, then check the note for each. Two methods reach it independently and trade off: a per-fact pipeline, and a GEPA-evolved prompt doing the same in one call. The pipeline's flags name the missing fact and its severity at 2.7% false alarms. The single call detects more (36.9% against 24.6%, p=0.002) at 6.2% false alarms and a tenth of the cost per note. A physician author validated 70 items and, where the two routes disagree, sided with the pipeline on 10 of 10 (p=0.002). A second clinician, not an author, graded the severity rubric blind and agrees to within a grade. On real vendor notes from a companion census no benchmark threshold transfers, but the re-calibrated single call detects more than the best of the eight at half its false-alarm rate. Omissions whose fact is restated elsewhere defeat both routes. We release the benchmark, prompts and judgements.

We study the problem of classification with a reject option for a fixed predictor, applicable in natural language processing. We introduce a new problem formulation for this scenario, and an algorithm minimizing a new surrogate loss function. We provide a complete theoretical analysis of the surrogate loss function with a strong H-consistency guarantee. For evaluation, we choose the decontextualization task, and provide a manually-labelled dataset of 2mathord,000 examples. Our algorithm significantly outperforms the baselines considered, with a sim!!25% improvement in coverage when halving the error rate, which is only sim!! 3 % away from the theoretical limit.

How many experimental studies would have come to different conclusions had they been run on larger samples? I show how to estimate the expected number of statistically significant results that a set of experiments would have reported had their sample sizes all been counterfactually increased. The proposed deconvolution estimator is asymptotically normal and adjusts for publication bias. Unlike related methods, this approach requires no assumptions of any kind about the distribution of true intervention treatment effects and allows for point masses. Simulations find good coverage even when the t-score is only approximately normal. An application to randomized trials (RCTs) published in economics journals finds that doubling every sample would increase the power of t-tests by 7.2 percentage points on average. This effect is smaller than for non-RCTs and comparable to systematic replications in laboratory psychology where previous studies enabled more accurate power calculations. This suggests that RCTs are on average relatively insensitive to sample size increases. Research funders who wish to raise power should generally consider sponsoring better-measured and higher quality experiments -- rather than only larger ones.

In natural language processing, a recently popular line of work explores how to best report the experimental results of neural networks. One exemplar publication, titled "Show Your Work: Improved Reporting of Experimental Results," advocates for reporting the expected validation effectiveness of the best-tuned model, with respect to the computational budget. In the present work, we critically examine this paper. As far as statistical generalizability is concerned, we find unspoken pitfalls and caveats with this approach. We analytically show that their estimator is biased and uses error-prone assumptions. We find that the estimator favors negative errors and yields poor bootstrapped confidence intervals. We derive an unbiased alternative and bolster our claims with empirical evidence from statistical simulation. Our codebase is at http://github.com/castorini/meanmax.

Hoffmann et al. (2022) propose three methods for estimating a compute-optimal scaling law. We attempt to replicate their third estimation procedure, which involves fitting a parametric loss function to a reconstruction of data from their plots. We find that the reported estimates are inconsistent with their first two estimation methods, fail at fitting the extracted data, and report implausibly narrow confidence intervals--intervals this narrow would require over 600,000 experiments, while they likely only ran fewer than 500. In contrast, our rederivation of the scaling law using the third approach yields results that are compatible with the findings from the first two estimation procedures described by Hoffmann et al.

Benchmark contamination is diagnosed with n-gram overlap, likelihood-based membership inference, or canary strings, and each needs something usually unavailable: the training corpus, a well-chosen test statistic, or foresight at release. A recent alternative reads it off a linear probe on internal activations. We show the natural way to do this does not work, specify one that survives measurement, then find that the correction making it work carries more variance than the null it is tested against. The protocol reports a zero-sum contrast on the depth profile of probe accuracy, recentred on a level-matched placebo baseline, tested against a label-permutation null, with the reference set twice the size of the suspect set. Each choice replaces a simpler alternative we rejected on measurement. Reporting the level of excess separability rather than its shape makes the false positive rate track the size of the analyst's own control set, 0.03 to 0.99 under a true null. Contrasting against a flat depth profile rejects a true null 0.72 of the time when surface decodability rises with depth, and loses all power when it falls. On real transformers the protocol fails a test the simulations did not pose. The recentring subtracts an estimate, and the permutation null holds it fixed. Re-estimated across split seeds on four audits of contaminated checkpoints, its standard deviation is 1.30 to 1.56 times the null's own in every arm: what is subtracted to remove a bias is more variable than what it corrects. The one nominally significant result, p = 0.0075, becomes 0.0745 once that variance is propagated, and no verdict is issued. The simulations missed this because their surface key is the covariate driving item variation; on real text it is a proxy, and degrading key quality in simulation reproduces it. We add a companion measurement and a widened null. No arm shows contamination.

This study evaluates the effect of counterfactual explanations on the interpretation of chest X-rays. We conduct a reader study with two radiologists assessing 240 chest X-ray predictions to rate their confidence that the model's prediction is correct using a 5 point scale. Half of the predictions are false positives. Each prediction is explained twice, once using traditional attribution methods and once with a counterfactual explanation. The overall results indicate that counterfactual explanations allow a radiologist to have more confidence in true positive predictions compared to traditional approaches (0.15pm0.95 with p=0.01) with only a small increase in false positive predictions (0.04pm1.06 with p=0.57). We observe the specific prediction tasks of Mass and Atelectasis appear to benefit the most compared to other tasks.

Safety alignment in open-weight language models is trivially removable: abliteration projects a refusal-mediating direction out of the weights in minutes, and no release-time defense we are aware of prevents it durably. What cannot be prevented can be deceived. Our defense, decoy hardening ("Fool's Gold"), concedes the refusal strip and poisons its payoff: once refusal is stripped, most answers to hazardous operational requests are confident, fluent decoys whose critical elements are falsified. Decoys are trained inside a differentiable simulation of the attack, expressing only in the attacked state; a refusal pin and benign leash hold clean-state behavior to the original. We instantiate it on seven models from five families (9B-122B, dense and mixture-of-experts). On the six models passing our pre-registered efficacy gate, 0.51-0.90 of attacked-state responses to held-out prompts are decoys, +0.27-0.84 attributable to the defense; all six stay within registered benign-behavior and capability budgets; the seventh (smaller) fails the gate (boundary case). Rates replicate on a frozen test split or untouched strata. The claim is epistemic: without independent ground truth, no observation surface we tested separates falsified answers from correct ones - on external red-team benchmarks' CBRNE-adjacent slice, the defended 122B is fatally wrong on 0.82-0.86 of matched-quality answers vs at most 0.10 undefended. Repeated sampling does not restore trust: element-wise consensus at K=64 reconstructs a fully usable procedure on 0.083-0.625 of prompts where the instrument validates, vs 0.58-0.96 undefended, with no label-free way to tell the regimes apart; on the weakest such model the claim is per-draw only. We evaluate chemical and biological hazards; the defense does not address in-context jailbreaks and protects only the initially released defended weights.

When can an agent failure be caught? An audit is usually limited by the record rather than by the method. CatchBench therefore puts one auditor's question to three information states: the declared configuration before a run (PRE), a growing prefix of its trace (LIVE), and the finished trace (POST). Prior benchmarks fix one of these states or vary the telemetry; to our knowledge none scores all three under one task-method interface. Each state admits different questions, so seven task contracts carry their own labels and metrics rather than one leaderboard. Four are evidential; three are Gold-derived mechanism diagnostics. The release scores 72 entrants, from rule scanners and structural models to eleven LLM judges across nine model families (GPT, Claude, Gemini, Gemma, Llama, Qwen, DeepSeek, Mistral, Nova), over 1187 declared configurations and 1162 recorded runs. Most of the arena does not order: 56 of 138 registered contrasts separate, and the rest are published unresolved rather than ranked. The two sharpest results cut against our own data. One rule ignores every name and permission; it flags each capability declared after the first. On one of six configuration sources it reaches a perfect F1, so a score there measures how the corpus was built rather than how well a method reasons. Our admissibility bar then rejected one injected substrate and withheld evidential status from the other. A benchmark number is therefore not interpretable until the process behind its labels is published and tested for the shortcut it may leave. We report both, and regenerate every ordering from released predictions with no model call.

Recent research has generated hope that inference scaling could allow weaker language models to match or exceed the accuracy of stronger models, such as by repeatedly sampling solutions to a coding problem until it passes unit tests. The central thesis of this paper is that there is no free lunch for inference scaling: indefinite accuracy improvement through resampling can only be realized if the "verifier" (in this case, a set of unit tests) is perfect. When the verifier is imperfect, as it almost always is in domains such as reasoning or coding (for example, unit tests have imperfect coverage), there is a nonzero probability of false positives: incorrect solutions that pass the verifier. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling even with an infinite compute budget. We find that there is a very strong correlation between the model's single-sample accuracy (i.e. accuracy without unit tests) and its false positive rate on coding benchmarks HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model (Fig. 1a). When we consider that false positives have a negative utility compared to abstaining from producing a solution, it bends the inference scaling curve further downward. Empirically, we find that the optimal number of samples can be less than 10 under realistic assumptions (Fig. 1b). Finally, we show that beyond accuracy, false positives may have other undesirable qualities, such as poor adherence to coding style conventions.

Tool-call judges evaluate argument values proposed by LLM agents. We find that they accept a correct value less often when their input holds no reference for it, that is, no text that says what the value should be. This happens when the value is a default value from the function schema that the user did not mention, when it has to be inferred, or when the text it comes from was cut from the judge input. We call this drop in acceptance the derivation gap. It appears across all seven judges we evaluate, including System One models CLM and Jev. The gap is not only a truncation artifact: correct default values have the largest gap on a benchmark with no truncation. A controlled source-relocation test shows that judges follow the reference in their input: they favor a visible incorrect value over a correct value whose source is hidden. We introduce DeRef (Derived Reference): a separate model reads the full history without seeing the candidate and writes a reference for the judge. Across three benchmarks, DeRef improves overall argument AUROC for all judges tested. On the controlled test, it raises AUROC from chance or below to between 98.40 and 99.59.

Large Language Models (LLMs) are increasingly being used to summarize and verify claims in critical sources that inform consequential decisions. Such claims are full of plural definites. If three of five patients improved after taking drug A, the claim the patients improved after drug A is not true, since two did not, yet its negation, the patients didn't improve after drug A, is not true either, since three did. In formal semantics, this truth-value gap is known as the Homogeneity Gap: both the plural definite and its negation are neither true nor false when the property holds for only a fraction of the plural referent. A model that cannot represent this gap turns mixed evidence into a clean verdict, an error that can be costly in high-stakes use. We introduce HomogeneityNLI, a diagnostic dataset of 1,000 minimal pair items crossing determiner, polarity, and group composition across ten logical conditions. Evaluating seven open-weight models, we uncover a pervasive Bivalence Bias: models assign plural definites their homogeneous truth conditions. In the zero-shot setting, no model recognizes the gap for both polarities in more than 15% of scenarios, and negated definites are judged False in 84% of cases. Explicit linguistic rules, chain-of-thought, and enumeration prompts produce unstable, model-specific effects that often entrench the bias. Our findings indicate that current open-weight LLMs acquire the truth conditions of plural predication but not its trivalent logic, and cannot yet be trusted to summarize or verify claims whose evidence is mixed.

We study the problem of detecting the correlation between two Gaussian databases XinR^{ntimes d} and Y^{ntimes d}, each composed of n users with d features. This problem is relevant in the analysis of social media, computational biology, etc. We formulate this as a hypothesis testing problem: under the null hypothesis, these two databases are statistically independent. Under the alternative, however, there exists an unknown permutation sigma over the set of n users (or, row permutation), such that X is rho-correlated with Y^sigma, a permuted version of Y. We determine sharp thresholds at which optimal testing exhibits a phase transition, depending on the asymptotic regime of n and d. Specifically, we prove that if rho^2dto0, as dtoinfty, then weak detection (performing slightly better than random guessing) is statistically impossible, irrespectively of the value of n. This compliments the performance of a simple test that thresholds the sum all entries of X^TY. Furthermore, when d is fixed, we prove that strong detection (vanishing error probability) is impossible for any rho<rho^star, where rho^star is an explicit function of d, while weak detection is again impossible as long as rho^2dto0. These results close significant gaps in current recent related studies.

In the reduced basis method, the evaluation of the a posteriori estimator can become very sensitive to round-off errors. In this note, the origin of the loss of accuracy is revealed, and a solution to this problem is proposed and illustrated on a simple example.

Sharp bounds on two potential-outcome means need not combine into sharp bounds on their difference. We characterize exactly when composition is guaranteed across environments sharing a joint potential-outcome law. For binary treatment and outcome, fix strict selection-probability bounds and interior treatment propensities. Subtracting the full model's sharp mean intervals gives sharp average-treatment-effect bounds for every compatible observed law if and only if the absolute selection bounds coincide across environments. Under this alignment, the complete mean region is rectangular; its endpoints and attaining causal models require O(K) arithmetic operations for K environments. For every mismatch, failure occurs on a relatively open set of positive observed tables at the prescribed propensities and persists under small joint drift. We give computable perturbation certificates and an exact family whose gap is first order in the mismatch. A separate finite-confounding family shows when the lost joint information changes an identified causal sign. These results distinguish local marginal sharpness from joint attainability after combining environments.

Standard methods for detecting discontinuities in conditional means are not applicable to outcomes that are complex, non-Euclidean objects like distributions, networks, or covariance matrices. This article develops a nonparametric test for jumps in conditional means when outcomes lie in a non-Euclidean metric space. Using local Fr\'echet regressionx2014which generalizes standard regression to metric-space valued datax2014the method estimates a mean path on either side of a candidate cutoff, extending existing k-sample tests to a flexible regression setting. Key theoretical contributions include a central limit theorem for the local estimator of the conditional Fr\'echet variance and the asymptotic validity and consistency of the proposed test. Simulations confirm nominal size control and robust power in finite samples. Two applications demonstrate the method's value by revealing effects invisible to scalar-based tests. First, I detect a sharp change in work-from-home compositions at Washington State's income threshold for non-compete enforceability during COVID-19, highlighting remote work's role as a bargaining margin. Second, I find that countries restructure their input-output networks after losing preferential US trade access. These findings underscore that analyzing regression functions within their native metric spaces can reveal structural discontinuities that scalar summaries would miss.

Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 150 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). A prompt ablation shows the low coverage is not an under-prompting artifact: explicitly asking models to be thorough does not close the gap. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.

A common assumption holds that enough observational and interventional data, given to a strong enough predictor, suffices. We report a failure mode that contradicts it. Across hundreds of structural causal models, on identified quantities a strong predictor and a Bayesian baseline both succeed, but on unidentified quantities (the couplings between counterfactual worlds) the predictor collapses to a point, on 28% of models to one no valid model can produce, while the truth is an admissible interval more data never narrows. The gap is structural: prediction cannot represent uncertainty over counterfactual couplings. We cast a world model as a single positive semidefinite coupling kernel K(T,T') over admissible worlds, whose diagonal is the ordinary posterior (what a predictor recovers) and whose off-diagonal is the cross-world coupling it cannot, which every counterfactual reads. The paper is the theory of that off-diagonal. It is real: two states with identical posteriors differ on a cross-world query, and the off-diagonal is the coupling that fixes counterfactuals. It can be bounded: positive semidefiniteness is partial-identifying information the marginals lack, and enforcing it bounds counterfactuals in polynomial time where the exact response-type program is intractable. Logical structure sharpens it: ontology axioms tighten the bound by up to a third, propagating to couplings they never touch. It can be acquired: targeted scars, constraints learned from encountered infeasibilities, close the gap several times faster than untargeted ones. Its full reconstruction is approximate counting of the admissible worlds, tractable below the Sly-Sun threshold and inapproximable above; we do not claim to beat the worst case.

Noise plagues many numerical datasets, where the recorded values in the data may fail to match the true underlying values due to reasons including: erroneous sensors, data entry/processing mistakes, or imperfect human estimates. We consider general regression settings with covariates and a potentially corrupted response whose observed values may contain errors. By accounting for various uncertainties, we introduced veracity scores that distinguish between genuine errors and natural data fluctuations, conditioned on the available covariate information in the dataset. We propose a simple yet efficient filtering procedure for eliminating potential errors, and establish theoretical guarantees for our method. We also contribute a new error detection benchmark involving 5 regression datasets with real-world numerical errors (for which the true values are also known). In this benchmark and additional simulation studies, our method identifies incorrect values with better precision/recall than other approaches.

Automated checking pipelines increasingly place one language model as the checker and another (or the same one) as the fixer. We ask whether that wiring changes what the checker reports. Measuring false alarms on human-verified-correct ProcessBench traces with the present task held byte-identical, we find that a completed audit -> repair episode already in the model's context lowers false alarms in 15 of 15 model x wording combinations, by 2.8 to 11.5 percentage points against a length-matched non-audit control, a 9 to 25% reduction relative to that control. The direction contradicts what the accumulated-message literature predicts: an episode whose audit reported an error lowers false alarms further still, at all five wordings on the model where that manipulation lands cleanly, though a negativity asymmetry predicts more flagging. Decomposing the episode finds repair content and audit verdict complementary: different components carry the effect on different model families. Signal-detection analysis locates the change in the threshold rather than in discrimination -- the criterion moves in 15 of 15 combinations and survives correction in 13 while d' survives in none, though the d' test is half as sensitive by construction -- and a hand audit of 50 false alarms finds 82% simply wrong, so at this operating point the shift need not be harmful. With reasoning enabled the effect keeps its relative size on both models tested, and the threshold reading holds there too.

Almost every large language model that reaches a broad audience is quantized: trained in full precision, then compressed for efficiency. This step is assumed harmless and its safety is rarely re-checked. We find its principal side effect is increased bias that standard safety evaluation misses. Holding the model, its training, and the prompts fixed, a quantized model still refuses harmful requests, still avoids over-refusing benign prompts, and still selects the unbiased multiple-choice answer. Yet asked an open-ended question, the same model volunteers stereotypes in all eight languages we probe, in roughly one in four open-ended answers under an independent judge (~24% to ~27% across the compression ladder): it passes every standard check and still reaches users measurably more biased. The selective gap is a robust finding; whether open-ended bias further increases with compression is less certain, sensitive to the judge that scores it. We address both with QuantiBias, a benchmark that pairs a generative, multilingual stereotype probe with the refusal and multiple-choice controls that isolate open-ended generation, contrasts each build with and without reasoning, and rates the content severity of what it generates. Across two backbone models (Qwen and Gemma), a five-family screen, and eight benchmarks, quantizers allocate their extra precision by capability data that carries no bias-prevention signal, and reasoning before answering roughly halves the effect on some families while doing nothing on others. A quantized build must be re-evaluated for open-ended bias, not only on the short-form safeguards it already passes.

Not all data in a typical training set help with generalization; some samples can be overly ambiguous or outrightly mislabeled. This paper introduces a new method to identify such samples and mitigate their impact when training neural networks. At the heart of our algorithm is the Area Under the Margin (AUM) statistic, which exploits differences in the training dynamics of clean and mislabeled samples. A simple procedure - adding an extra class populated with purposefully mislabeled threshold samples - learns a AUM upper bound that isolates mislabeled data. This approach consistently improves upon prior work on synthetic and real-world datasets. On the WebVision50 classification task our method removes 17% of training data, yielding a 1.6% (absolute) improvement in test error. On CIFAR100 removing 13% of the data leads to a 1.2% drop in error.

A training-data quality gate changes both which examples survive within regions of the input space and how training mass is distributed across those regions, so a gate-versus-random comparison does not isolate these mechanisms. We introduce Counterfactual Gate-Support Decomposition (CGSD), a four-arm, budget-matched design that crosses gate eligibility with support distribution and decomposes the observed gate effect into within-support selection, between-support redistribution, and their interaction. The practical constraint is finite-pool realizability: we characterize mass and token-assembly conditions under which the required arms can be constructed, together with a support-extinction case in which the ordinary design is undefined. On Korean-English and Korean-Japanese retrieval, the mean within-support selection component is the only one that exceeds its calibrated null band (+29.7 and +5.6 holdout items; 30/30 positive replicates), decisively in Korean-English and marginally in Korean-Japanese; redistribution and interaction remain unresolved at the available power. The same pattern holds for a second, unrelated gate family -- a model-confidence threshold rather than a deterministic verifier -- judged against its own null (+27.3 items, 9/9 positive), so the result is not a property of one gate's mechanism. On a mathematics gate, zero-eligibility supports leave the ordinary design undefined at every budget, while the extended design is constructible only inside narrow budget bands that an exhaustive scan locates and that the natural operating point misses -- the assembly condition separating a property of the pool from a limitation of the packer. CGSD therefore returns a decomposition, a constructible-but-underpowered design, or an explicit non-constructibility result, and says which; it prevents support redistribution from being silently interpreted as sample-level gate quality.

Instruction fine-tuning (IFT) can increase the informativeness of large language models (LLMs), but may reduce their truthfulness. This trade-off arises because IFT steers LLMs to generate responses containing long-tail knowledge that was not well covered during pre-training. As a result, models become more informative but less accurate when generalizing to unseen tasks. In this paper, we empirically demonstrate how unfamiliar knowledge in IFT datasets can negatively affect the truthfulness of LLMs, and we introduce two new IFT paradigms, UNIT_{cut} and UNIT_{ref}, to address this issue. UNIT_{cut} identifies and removes unfamiliar knowledge from IFT datasets to mitigate its impact on model truthfulness, whereas UNIT_{ref} trains LLMs to recognize their uncertainty and explicitly indicate it at the end of their responses. Our experiments show that UNIT_{cut} substantially improves LLM truthfulness, while UNIT_{ref} maintains high informativeness and reduces hallucinations by distinguishing between confident and uncertain statements.

The false discovery rate (FDR) measures the share of false positives in a set of statistical tests. I develop simple and intuitive bounds on the FDR in cross-sectional predictability publications. The simplest bound requires just a few lines of math and finds FDR le 25% based on summary statistics in eight out of nine previous studies. A more refined bound finds FDR le 9%. The FDR is small because randomly selecting accounting ratios produces statistically significant predictability far more often than would occur if there were no predictability. The bounds also reconcile the disparate FDR estimates in the literature.

Ask a language model the same question about the same document twenty times, and it sometimes returns two different answers. We built Probity, a benchmark of 60 tasks and 470 items from real venture-financing filings, to measure how often this happens. Then we audited our own corpus and found a defect any excerpt-built benchmark can carry: items whose evidence is missing from the window of text the model is shown. The audit flags 36 items and separates two failures a single flag would conflate: evidence genuinely absent from the window and answers that must be computed from numbers the window does supply. Flagged items change their answers far more often, wobbling at 0.255 against 0.087 on the 427 clean items, and excluding them cuts apparent cross-model agreement by about a fifth. Before testing whether the missing evidence explains the instability, we registered a prediction: re-cut each window to hold its evidence, and instability should fall below a set threshold. It failed: the repair moved wobble by 0.058, with an interval containing zero. We report the association as correlational. Almost all measurements sit where instability cannot show, which bounds what a corpus built for accuracy can say about stability. We release the corpus, all 112,800 raw responses, and the audit as a runnable check for any document benchmark.

Social scientists are now using large language models to create "silicon samples" - synthetic datasets intended to stand in for human respondents, aimed at revolutionising human subjects research. However, there are many analytic choices which must be made to produce these samples. Though many of these choices are defensible, their impact on sample quality is poorly understood. I map out these analytic choices and demonstrate how a very small number of decisions can dramatically change the correspondence between silicon samples and human data. Configurations (N = 252) varied substantially in their capacity to estimate (i) rank ordering of participants, (ii) response distributions, and (iii) between-scale correlations. Most critically, configurations were not consistent in quality: those that performed well on one dimension often performed poorly on another, implying that there is no "one-size-fits-all" configuration that optimises the accuracy of these samples. I call for greater attention to the threat of analytic flexibility in using silicon samples.

Isolation forest (iForest) has been emerging as arguably the most popular anomaly detector in recent years due to its general effectiveness across different benchmarks and strong scalability. Nevertheless, its linear axis-parallel isolation method often leads to (i) failure in detecting hard anomalies that are difficult to isolate in high-dimensional/non-linear-separable data space, and (ii) notorious algorithmic bias that assigns unexpectedly lower anomaly scores to artefact regions. These issues contribute to high false negative errors. Several iForest extensions are introduced, but they essentially still employ shallow, linear data partition, restricting their power in isolating true anomalies. Therefore, this paper proposes deep isolation forest. We introduce a new representation scheme that utilises casually initialised neural networks to map original data into random representation ensembles, where random axis-parallel cuts are subsequently applied to perform the data partition. This representation scheme facilitates high freedom of the partition in the original data space (equivalent to non-linear partition on subspaces of varying sizes), encouraging a unique synergy between random representations and random partition-based isolation. Extensive experiments show that our model achieves significant improvement over state-of-the-art isolation-based methods and deep detectors on tabular, graph and time series datasets; our model also inherits desired scalability from iForest.

Operator-learning benchmarks ship as full-precision arrays, and they have grown to terabyte scale. A curator who wants to distribute one has to decide how coarsely to store it. That decision is usually made by fixing a tolerance on the reconstruction error of the stored field. We show that this quantity is measured in the wrong place. It compares the stored field with the original, before any model is trained. What the curator is buying is the accuracy of a model trained on the compressed copy, and the two can disagree: two PDEBench families stored to identical field error differ threefold downstream. A solution operator smooths, so only part of the codec error ever reaches the model's output. That part can be measured. Push a compressed field through a model already trained at full precision, and compare its output against the output on the clean field. The measurement costs forward passes and no training on compressed data. It spans more than two orders of magnitude across six PDE families, and it orders downstream cost where field error does not, inverting 12 of 104 cross-dataset comparisons against field error's 36. Read as a budget, the same curve nominates a storage rate for each dataset. A back-check against trained models finds those rates wrong in both directions by up to a factor of two in stored bits.

Modern image editing models can satisfy a text instruction while breaking the physics of the edited scene. A new object may cast no shadow, a mirror may fail to reflect visible geometry, or an object may float above a surface that should support it. We study physical plausibility diagnosis, detecting whether an edited image violates scene physics, naming the violation type, localizing the affected region, and explaining the failure in language. We introduce a counterfactual benchmark whose controlled synthetic component uses Mitsuba~3 to generate 5,500 images from 500 scene families. Each family contains one clean image and ten matched violations involving shadows, reflection, support, surface response, and occlusion. The renderer pipeline provides category labels, affected-region masks and boxes, scene metadata, and explanation targets. We use LLaVA-1.5-7B, Qwen2.5-VL-7B, and InternVL3.5-8B as diagnostic baselines rather than proposed methods. On a 1,650-image synthetic test set, the adapted baselines reach 64.0--67.8\% category macro-F1 on standard held-out scenes. For LLaVA-1.5-7B, category macro-F1 falls from 64.0\% on the standard split to 40.8\% under intervention shift. This gap shows that high in-distribution accuracy partly reflects cues tied to rendering and counterfactual construction.

Verifying the final answer of a long reasoning trace may show that the trace is wrong but it does not distinguish three cases. The problem may lie beyond the model's capability, another attempt would succeed, or one step in an otherwise recoverable trace made the trace fail. Standard hidden-state probes are trained on final-answer correctness, a label that conflates these cases. Monte Carlo rollouts distinguish the cases and locate the failing step at additional inference cost. We test the hidden states of 4B and 8B reasoning models for the failure type and the failing step and compare each probe with the strongest baseline that uses no hidden states. On incorrect 8B traces, the confidence probe, a calibrated correctness probe at a midpoint cutoff, misses 81% of cases in which a single step makes the trace unrecoverable and 47% of cases in which another attempt would succeed. A probe with a different training design over the same states reverses the ordering of these two kinds of failure, so either ordering can be read from those states. At both scales, no single tested probe locates the failing step more accurately than a rule based only on trace length, and at 4B a probe trained to predict that step fails to outperform this rule in any of 10 runs on identical traces. The confidence probe therefore indicates that an answer is wrong, but its failure-type ordering depends on its training design.

Single-cell trajectory inference maps transcriptomic measurements onto developmental continua, yet configurations that fit the data equally well can assign conflicting cell fates. FateMultiplicity is a label-free framework that constructs a statistically admissible model set, or Rashomon set, without lineage labels, by evaluating model discrepancy on cross-fitted held-out genes under non-inferiority testing calibrated against random-seed variation. Multiplicity is large and depends more on the diversity of the model space than its size: twelve configurations of a second algorithm expose 20.0% of cells where twenty-four of the first expose 3.8%. Whether the per-cell certified fate margin FM yields more reliable assignments than the fitted model already provides is then tested, and it does not. On simulation ground truth, on the same cells, FM discriminates misassignment at AUC 0.682, against 0.965 for the baseline configuration's own decision margin (p = 0.003) and 0.854 for a seed-dispersion baseline. Informativeness is governed by the breadth of the admitted set, not its cardinality: at cardinality four, seed refits give 0.933 and hyperparameter-perturbed sets 0.701. Relaxing the infimum to a q-quantile recovers discrimination but converges toward the single model's own confidence; the supremum reaches 0.973 because theta*'s membership bounds it from below, while the infimum is unanchored. Multiplicity in trajectory inference is worth measuring and reporting, but per-cell certification over a label-free Rashomon set is not a route to more reliable fate calls. Two constructions survive: a margin-erosion ratio separates real from spurious branch points in simulation (AUC 0.890, untested on real data), and against clonally observed fate, uncertified cells disagree with their clone's outcome 16.4 percentage points more often than certified cells (p < 0.001).

Early stopping based on hold-out data is a popular regularization technique designed to mitigate overfitting and increase the predictive accuracy of neural networks. Models trained with early stopping often provide relatively accurate predictions, but they generally still lack precise statistical guarantees unless they are further calibrated using independent hold-out data. This paper addresses the above limitation with conformalized early stopping: a novel method that combines early stopping with conformal calibration while efficiently recycling the same hold-out data. This leads to models that are both accurate and able to provide exact predictive inferences without multiple data splits nor overly conservative adjustments. Practical implementations are developed for different learning tasks -- outlier detection, multi-class classification, regression -- and their competitive performance is demonstrated on real data.

With the growing importance of detecting misinformation, many studies have focused on verifying factual claims by retrieving evidence. However, canonical fact verification tasks do not apply to catching subtle differences in factually consistent claims, which might still bias the readers, especially on contentious political or economic issues. Our underlying assumption is that among the trusted sources, one's argument is not necessarily more true than the other, requiring comparison rather than verification. In this study, we propose ClaimDiff, a novel dataset that primarily focuses on comparing the nuance between claim pairs. In ClaimDiff, we provide 2,941 annotated claim pairs from 268 news articles. We observe that while humans are capable of detecting the nuances between claims, strong baselines struggle to detect them, showing over a 19% absolute gap with the humans. We hope this initial study could help readers to gain an unbiased grasp of contentious issues through machine-aided comparison.

RED-2400 is a public benchmark of algorithmically-rejected trading events from a live Solana decentralized-exchange filter stack. I logged the data continuously between 2026-04-10 and 2026-05-02. The benchmark contains 6,659 rejection events linked to 169,122 post-rejection price and liquidity observations and 1,836 graveyard-tracker snapshots. Outcome labels follow the five-tier classification of Kamat (2026c): saved (windowed), saved (early-death), missed, flat, and unclassifiable. Thresholds use the trough-to-reference and peak-to-reference price ratios within a 24-hour window. Most filter-design datasets cover the accept side only. That gap leaves reject-side outcomes unmeasured and biases filter validation. RED-2400 lets researchers replicate filter-precision claims directly. RED-2400 is the first window in a planned dataset series; subsequent windows will extend the time horizon and enable regime-stratified analysis.

This paper studies the prediction of a target z from a pair of random variables (x,y), where the ground-truth predictor is additive E[z mid x,y] = f_star(x) +g_{star}(y). We study the performance of empirical risk minimization (ERM) over functions f+g, f in F and g in G, fit on a given training distribution, but evaluated on a test distribution which exhibits covariate shift. We show that, when the class F is "simpler" than G (measured, e.g., in terms of its metric entropy), our predictor is more resilient to heterogenous covariate shifts in which the shift in x is much greater than that in y. These results rely on a novel H\"older style inequality for the Dudley integral which may be of independent interest. Moreover, we corroborate our theoretical findings with experiments demonstrating improved resilience to shifts in "simpler" features across numerous domains.

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is not. Three classifiers (a regex-only detector, a two-stage regex-plus-LLM pipeline, and an independent Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, respectively, with non-overlapping 95% confidence intervals. Per-model gaps range from 2.6 to 30.6 percentage points; all are statistically significant (McNemar's test, p < 0.001). The disagreements are systematic, not random: inter-classifier agreement measured by Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under the Sonnet judge; OLMo-3.1-32B moves in the opposite direction, from 9th to 3rd. The root cause is that different classifiers operationalize related faithfulness constructs at different levels of stringency (lexical mention versus epistemic dependence), and these constructs yield divergent measurements on the same behavior. These results demonstrate that published faithfulness numbers cannot be meaningfully compared across studies that use different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies rather than single point estimates.

Computer-aided detection (CADe) systems for colonoscopy promise to reduce clinical miss rates, yet reliable real-world deployment remains elusive. This translational gap stems in part from a structural flaw in model development: the reliance on curated datasets that under-represent the long negative stretches and procedure-related artifacts characteristic of routine examinations. Training and evaluating architectures strictly on these lesion-centric benchmarks creates an illusion of success, since such benchmarks cannot capture clinically crucial metrics. To expose this gap, we establish TRUE-Colon, a standardized benchmarking protocol that measures key deployment characteristics alongside localization accuracy, and evaluate four real-time architectures (Faster R-CNN, YOLOv8, YOLOv11, RT-DETR) across curated benchmarks (SUN, PICCOLO) and 60 unedited, full-length procedures (REAL-Colon). We observe a consistent transfer asymmetry: models trained strictly on curated clips suffer a severe performance collapse when evaluated on full procedures, whereas procedure-trained models substantially improve rejection of non-polyp content on REAL-Colon, and largely retain their accuracy on curated benchmarks. Beyond transferability, we find that the Transformer detector attains the strongest sensitivity and the earliest, most persistent detections, while the convolutional detectors stay competitive at a higher throughput. Together, these results indicate that both training and benchmarking for deployable CADe should shift from curated, lesion-centric clips toward full-procedure data and deployment-relevant operating points. Source code is available at https://github.com/sdoerrich97/true-colon.

Regional dropout strategies have been proposed to enhance the performance of convolutional neural network classifiers. They have proved to be effective for guiding the model to attend on less discriminative parts of objects (e.g. leg as opposed to head of a person), thereby letting the network generalize better and have better object localization capabilities. On the other hand, current methods for regional dropout remove informative pixels on training images by overlaying a patch of either black pixels or random noise. Such removal is not desirable because it leads to information loss and inefficiency during training. We therefore propose the CutMix augmentation strategy: patches are cut and pasted among training images where the ground truth labels are also mixed proportionally to the area of the patches. By making efficient use of training pixels and retaining the regularization effect of regional dropout, CutMix consistently outperforms the state-of-the-art augmentation strategies on CIFAR and ImageNet classification tasks, as well as on the ImageNet weakly-supervised localization task. Moreover, unlike previous augmentation methods, our CutMix-trained ImageNet classifier, when used as a pretrained model, results in consistent performance gains in Pascal detection and MS-COCO image captioning benchmarks. We also show that CutMix improves the model robustness against input corruptions and its out-of-distribution detection performances. Source code and pretrained models are available at https://github.com/clovaai/CutMix-PyTorch .

Models driven by spurious correlations often yield poor generalization performance. We propose the counterfactual (CF) alignment method to detect and quantify spurious correlations of black box classifiers. Our methodology is based on counterfactual images generated with respect to one classifier being input into other classifiers to see if they also induce changes in the outputs of these classifiers. The relationship between these responses can be quantified and used to identify specific instances where a spurious correlation exists. This is validated by observing intuitive trends in face-attribute and waterbird classifiers, as well as by fabricating spurious correlations and detecting their presence, both visually and quantitatively. Furthermore, utilizing the CF alignment method, we demonstrate that we can evaluate robust optimization methods (GroupDRO, JTT, and FLAC) by detecting a reduction in spurious correlations.

When a language model is fed a wrong answer, what happens inside the network? Current understanding treats truthfulness as a static property of individual-layer representations-a direction to be probed, a feature to be extracted. Less is known about the dynamics: how internal representations diverge across the full depth of the network when the model processes correct versus incorrect continuations. We introduce forced-completion probing, a method that presents identical queries with known correct and incorrect single-token continuations and tracks five geometric measurements across every layer of four decoder-only models(1.5B-13B parameters). We report three findings. First, correct and incorrect paths diverge through rotation, not rescaling: displacement vectors maintain near-identical magnitudes while their angular separation increases, meaning factual selection is encoded in direction on an approximate hypersphere. Second, the model does not passively fail on incorrect input-it actively suppresses the correct answer, driving internal probability away from the right token. Third, both phenomena are entirely absent below a parameter threshold and emerge at 1.6B, suggesting a phase transition in factual processing capability. These results show that factual constraint processing has a specific geometric character-rotational, not scalar; active, not passive-that is invisible to methods based on single-layer probes or magnitude comparisons.

When deploying machine learning models in high-stakes real-world environments such as health care, it is crucial to accurately assess the uncertainty concerning a model's prediction on abnormal inputs. However, there is a scarcity of literature analyzing this problem on medical data, especially on mixed-type tabular data such as Electronic Health Records. We close this gap by presenting a series of tests including a large variety of contemporary uncertainty estimation techniques, in order to determine whether they are able to identify out-of-distribution (OOD) patients. In contrast to previous work, we design tests on realistic and clinically relevant OOD groups, and run experiments on real-world medical data. We find that almost all techniques fail to achieve convincing results, partly disagreeing with earlier findings.

Early warning signals have been proposed to forecast the possibility of a critical transition, such as the eutrophication of a lake, the collapse of a coral reef, or the end of a glacial period. Because such transitions often unfold on temporal and spatial scales that can be difficult to approach by experimental manipulation, research has often relied on historical observations as a source of natural experiments. Here we examine a critical difference between selecting systems for study based on the fact that we have observed a critical transition and those systems for which we wish to forecast the approach of a transition. This difference arises by conditionally selecting systems known to experience a transition of some sort and failing to account for the bias this introduces -- a statistical error often known as the Prosecutor's Fallacy. By analysing simulated systems that have experienced transitions purely by chance, we reveal an elevated rate of false positives in common warning signal statistics. We further demonstrate a model-based approach that is less subject to this bias than these more commonly used summary statistics. We note that experimental studies with replicates avoid this pitfall entirely.

Machine learning is increasingly used to guide branch-and-cut (B&C) for mixed-integer linear programming by learning score-based policies for selecting branching variables and cutting planes. Many approaches train on local signals from lookahead heuristics such as strong branching, and linear programming (LP) bound improvement for cut selection. Training and evaluation of the learned models often focus on local score accuracy. We show that such local score-based methods can lead to search trees exponentially larger than optimal tree sizes, by identifying two sources of this gap. The first is that these widely used expert signals can be misaligned with overall tree size. LP bound improvement can select a root cut set that yields an exponentially larger strong branching tree than selecting cuts by a simple proxy score, and strong branching itself can be exponentially suboptimal (Dey et al., 2024). The second is that small discrepancies can be amplified by the branch-and-bound recursion. An arbitrarily small perturbation of the right-hand sides in a root cut set can change the minimum tree size from a single node to exponentially many. For branching, arbitrarily small score discrepancies, and differences only in tie-breaking, can produce trees of exponentially different sizes, and even a small number of decision differences along a trajectory can incur exponential growth. These results show that branch-and-cut policies trained and learned using local expert scores do not guarantee small trees, thus motivating the study of data-driven methods that produce policies better aligned with tree size rather than only accuracy on expert scores.

Decision APIs that return a probability for each option are routinely used as if those probabilities were properties of the options themselves. That reading is valid only if the model satisfies independence of irrelevant alternatives (IIA): adding an option that is neither the right answer nor its strongest rival should leave the odds between those two unchanged. We test this in a pre-registered study of jev-1.13.0, a hosted decision model, using 299,463 calls on intent classification (BANKING77, CLINC150) and topic classification (AG News). Every added option moves the log-odds between the gold label and the strongest wrong label: the root-mean-square systematic shift is 0.20-0.52 nats, 1.2 to 3 times the run-to-run variation, even when the added option takes at most 0.1% of the probability mass. The shifts are directional. An option similar to the wrong answer draws probability from it (+0.10 nats toward the gold label). A paraphrase of the wrong answer produces an intermediate pattern: the gold label's probability is almost conserved, but the paraphrase receives only about a third of the wrong answer's share rather than the half that strict substitution implies. The distortion grows steadily with the logarithm of the number of options, reaching about 1 nat at 255 options, and off-domain padding pushes the odds against the gold label. In contrast, questions packed into one call do not interact (bias < 0.001). Scoring each option with an independent yes/no question removes the option-set dependence but picks the right answer 8.7 points less often. Jev sits between two local baselines, an option-independent scorer and a prompt readout. Its probabilities are best read as relative to the option list they were computed over.

The growth rate of Av(1324) is the last unknown Stanley-Wilf limit of a length-four pattern. Since 2020 the best rigorous lower bound has been the 10.271012 of Bevan, Brignall, Elvey Price and Pantone; we raise it to 10.412263. Their construction lays 1324-avoiders along an infinite staircase of two-cell blocks alternating with single connecting cells, and forbids a 1324 by a local condition on how the points of a block interleave with the components of the cell beside it. Certain points of a block, its leaves, may be exempted from that condition. They exempt them on one axis only, and report that they could not count the possibilities when both are exempted. This paper exempts them on both axes and counts the result. The second relaxation is their own argument read in the other axis, and the Harris correlation inequality bounds the joint count below by the product of the two counts taken separately, at the cost of one factor of the component generating function.

We might hope that when faced with unexpected inputs, well-designed software systems would fire off warnings. Machine learning (ML) systems, however, which depend strongly on properties of their inputs (e.g. the i.i.d. assumption), tend to fail silently. This paper explores the problem of building ML systems that fail loudly, investigating methods for detecting dataset shift, identifying exemplars that most typify the shift, and quantifying shift malignancy. We focus on several datasets and various perturbations to both covariates and label distributions with varying magnitudes and fractions of data affected. Interestingly, we show that across the dataset shifts that we explore, a two-sample-testing-based approach, using pre-trained classifiers for dimensionality reduction, performs best. Moreover, we demonstrate that domain-discriminating approaches tend to be helpful for characterizing shifts qualitatively and determining if they are harmful.

Convolutional neural networks are capable of learning powerful representational spaces, which are necessary for tackling complex learning tasks. However, due to the model capacity required to capture such representations, they are often susceptible to overfitting and therefore require proper regularization in order to generalize well. In this paper, we show that the simple regularization technique of randomly masking out square regions of input during training, which we call cutout, can be used to improve the robustness and overall performance of convolutional neural networks. Not only is this method extremely easy to implement, but we also demonstrate that it can be used in conjunction with existing forms of data augmentation and other regularizers to further improve model performance. We evaluate this method by applying it to current state-of-the-art architectures on the CIFAR-10, CIFAR-100, and SVHN datasets, yielding new state-of-the-art results of 2.56%, 15.20%, and 1.30% test error respectively. Code is available at https://github.com/uoguelph-mlrg/Cutout

Large Language Models (LLMs) have impressive capabilities, but are prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we use high-quality datasets of simple true/false statements to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that at sufficient scale, LLMs linearly represent the truth or falsehood of factual statements. We also show that simple difference-in-mean probes generalize as well as other probing techniques while identifying directions which are more causally implicated in model outputs.

Data-centric curation pipelines frequently rely on model confidence scores to flag and filter noisy or mislabeled training instances. Evaluating this filtering convention on a large-scale consumer lending sample (LendingClub, N = 1,344,936) uncovers an underlying demographic asymmetry: high-income defaulters are disproportionately classified as label noise relative to low-income defaulters (Cramer's V approximately 0.03-0.07). Re-examining this behavior through the lens of equal opportunity [Hardt et al., 2016] reveals a far more severe discrepancy: a 16.86 percentage point gap in true positive rate (recall) between high- and low-income borrowers who ultimately defaulted. Implementing a sequential feature-blinding methodology allows us to isolate the drivers of this disparity across three distinct mechanisms: (1) direct reliance on self-reported applicant income; (2) algorithmic absorption of upstream institutional bias encoded within origination interest rates; and (3) a residual disparity (3.55 percentage points in cross-validation; 2.56 percentage points on a held-out test partition, Z = -4.04, p < 0.0001) that remains even after purging both income and interest rates from the model. Out-of-sample signed SHAP valuations demonstrate that this residual gap is maintained by structural proxies, most notably loan amount and home ownership status. These empirical findings show that simply blinding an algorithm to sensitive attributes fails to ensure fairness when institutional pricing decisions and behavioral proxy variables collectively reconstruct the omitted signals. We outline the practical implications of these findings for auditing data-centric AI workflows within regulated financial institutions.

Breakthroughs in machine learning are rapidly changing science and society, yet our fundamental understanding of this technology has lagged far behind. Indeed, one of the central tenets of the field, the bias-variance trade-off, appears to be at odds with the observed behavior of methods used in the modern machine learning practice. The bias-variance trade-off implies that a model should balance under-fitting and over-fitting: rich enough to express underlying structure in data, simple enough to avoid fitting spurious patterns. However, in the modern practice, very rich models such as neural networks are trained to exactly fit (i.e., interpolate) the data. Classically, such models would be considered over-fit, and yet they often obtain high accuracy on test data. This apparent contradiction has raised questions about the mathematical foundations of machine learning and their relevance to practitioners. In this paper, we reconcile the classical understanding and the modern practice within a unified performance curve. This "double descent" curve subsumes the textbook U-shaped bias-variance trade-off curve by showing how increasing model capacity beyond the point of interpolation results in improved performance. We provide evidence for the existence and ubiquity of double descent for a wide spectrum of models and datasets, and we posit a mechanism for its emergence. This connection between the performance and the structure of machine learning models delineates the limits of classical analyses, and has implications for both the theory and practice of machine learning.

Sharp, multidimensional changepoints-abrupt shifts in a regression surface whose locations and magnitudes are unknown-arise in settings as varied as gene-expression profiling, financial covariance breaks, climate-regime detection, and urban socioeconomic mapping. Despite their prevalence, there are no current approaches that jointly estimate the location and size of the discontinuity set in a one-shot approach with statistical guarantees. We therefore introduce Free Discontinuity Regression (FDR), a fully nonparametric estimator that simultaneously (i) smooths a regression surface, (ii) segments it into contiguous regions, and (iii) provably recovers the precise locations and sizes of its jumps. By extending a convex relaxation of the Mumford-Shah functional to random spatial sampling and correlated noise, FDR overcomes the fixed-grid and i.i.d. noise assumptions of classical image-segmentation approaches, thus enabling its application to real-world data of any dimension. This yields the first identification and uniform consistency results for multivariate jump surfaces: under mild SBV regularity, the estimated function, its discontinuity set, and all jump sizes converge to their true population counterparts. Hyperparameters are selected automatically from the data using Stein's Unbiased Risk Estimate, and large-scale simulations up to three dimensions validate the theoretical results and demonstrate good finite-sample performance. Applying FDR to an internet shutdown in India reveals a 25-35% reduction in economic activity around the estimated shutdown boundaries-much larger than previous estimates. By unifying smoothing, segmentation, and effect-size recovery in a general statistical setting, FDR turns free-discontinuity ideas into a practical tool with formal guarantees for modern multivariate data.

Explaining an inferred fact under rule evaluation can require identifying the inclusion-minimal input sets that suffice for the inference and the deletions that make the fact disappear. For a fixed union of conjunctive queries, every minimal support is bounded by the query body. For recursive rules, the same answer may depend on large supports, and the number of minimal supports may be exponential in the input. We study the gap through deletion-based explanation, using inclusion-minimal endogenous input facts that entail the atom together with fixed background facts. We organize these supports as a hypergraph and prove that it determines actual causes, counterfactual causes, responsibility, and deletion robustness. The resulting view separates nonrecursive queries from recursive Datalog at the level of minimal input explanations. For positive-length reachability, minimal supports are exactly simple directed paths, and deletion robustness is the minimum directed edge cut. We also prove invariance under fixed-goal equivalent positive Datalog programs and an NP-hardness calibration for the robustness threshold problem.

The term "hallucination" converge different failure modes with specific geometric signatures in embedding space. We propose a taxonomy identifying three types: unfaithfulness (Type I: ignoring provided context), confabulation (Type II: inventing semantically foreign content), and factual error (Type III: wrong details within correct conceptual frames). We introduce two detection methods grounded in this taxonomy: the Semantic Grounding Index (SGI) for Type I, which measures whether a response moves toward provided context on the unit hypersphere, and the Directional Grounding Index (DGI) for Type II, which measures displacement geometry in context-free settings. DGI achieves AUROC=0.958 on human-crafted confabulations with 3.8% cross-domain degradation. External validation on three independently collected human-annotated benchmarks -WikiBio GPT-3, FELM, and ExpertQA- yields domain-specific AUROC 0.581-0.695, with DGI outperforming an NLI CrossEncoder baseline on expert-domain data, where surface entailment operates at chance. On LLM-generated benchmarks, detection is domain-local. We examine the Type III boundary through TruthfulQA, where apparent classifier signal (Logistic Regression with AUROC 0.731) is traced to a stylistic annotation confound: false answers are geometrically closer to queries than truthful ones, a pattern incompatible with factual-error detection. This identifies a theoretical constraint from a methodological limitation.

Probes are the workhorse of interpretability. If a model's hidden states predict a variable, the model is said to represent it. But a probe score has no fixed meaning. An R^2 of 0.6 may only reflect what the input already gives away, and the same score can mean different things on different data. We propose reading every probe score against two reference points: a floor, what a declared set of simple inputs already predicts, and a ceiling, what the full input can predict. The gap between them, the headroom, is the range in which a probe can show that a model computes something beyond the simple inputs. We prove that headroom vanishes in two ways: the target stops depending on a hidden variable the model must infer, or the input stops revealing it. We test this on transformers trained for in-context meta-analysis, which must infer the hidden heterogeneity between studies to weight them correctly, and where both reference points are known. Under distribution shift, probe scores fall and prediction error rises 12--15times, yet the model recovers a similar share of the headroom, indicating that the data lost information, not the representation. We then analyze the real models. The single-cell foundation model scGPT encodes biological variability only partially. We also revisit four influential LLM probing studies, which claim that models represent geography, the state of an Othello board, truth, and the demographics of their users. Against a floor computed from the input text alone, some of these claims hold, while others are largely explained by the text itself.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

LLM-as-a-Judge frameworks are increasingly trusted to automate evaluation in place of human experts, yet their reliability in high-stakes medical contexts remains unproven. We stress-test this assumption for detecting incomplete patient-facing medical responses, evaluating three rubric granularities (General-Likert, Analytical-Rubric, Dynamic-Checklist) and three backbone models across two clinician-annotated datasets, including HealthBench, the largest publicly available benchmark for medical response evaluation. LLM Judges discriminate complete from incomplete responses at and slightly above near chance (AUC 0.49--0.66); at the threshold required to recall 90% of incomplete responses, clinicians must still review the vast majority of the dataset, offering no triage utility. Even when model and clinician verdicts agree, they rarely cite the same explanation; and when they diverge, false positives stem from over-flagging non-essential gaps while false negatives reflect outright detection failures. These results reveal that LLM Judges and clinicians apply fundamentally different completeness standards; a finding that undermines their use as autonomous evaluators or triage filters in clinical settings.

We study the fine-grained uniform convergence behavior of halfspaces beyond worst-case VC bounds. For inhomogeneous halfspaces in R^d with dge 2, we show that standard first-order VC bounds are essentially tight: even consistent hypotheses can incur population error Θ(dln(n/d)/n), and in the agnostic setting the deviation scales as τln(1/τ) at true error τ. In contrast, homogeneous halfspaces in R^2 exhibit a markedly different behavior. In the realizable case, every hypothesis consistent with the sample has error O(1/n). In the agnostic case, we prove a bandwise, log-free deviation bound on each dyadic risk band via a critical-wedge localization argument. Unioning over bands incurs only a lnln n overhead, and we establish a matching lower bound showing this overhead is unavoidable. Together, these results give a fine-grained and nearly complete picture of uniform convergence for halfspaces, revealing sharp dimensional and structural thresholds.

Claims that language models homogenise are usually measured against human judgements collected for the study, which makes the human side an artifact of the design: a crowdworker given the model's instruction is running the model's prompt. We measure convergence against a human reference nobody built for the purpose -- 2,523 reader mark sets across 120 web documents, produced by people highlighting for their own reasons on a platform where the overlay of others' marks is off by default. Agreement is the overlap between two size-matched sentence sets minus the overlap expected when each is resampled within its own depth-and-length bands. The null's calibration is demonstrated, not asserted: every pair involving a random baseline lands within 0.006 of zero. On the median document each party names 14 sentences of 70; two readers share 4.1 and two models 8.7. Across 18 model arms spanning 11 vendors, 3 countries and both weight regimes, the median of 153 model pairs is +0.093 against a human yardstick of +0.040, and 99 sit entirely above the human interval. Two frontier models from rival labs reach +0.203, twice what GPT-4o agrees with itself on a second call. The effect is not determinism, prompt wording, procedure, vendor or routing, and it is graded: the smallest models agree at the human level. No model agrees with readers detectably more than a reader does, and at equal depth and length no surface feature separates their choices. The multiples are procedure-dependent and the ordering is not: models are cut to their sharpest set while a reader's is a random draw from what they marked, and blunting the models alike halves the gap without closing it. Tested out of sample on four models released after this analysis, against predictions fixed beforehand, none clears the human interval. A population simulated from several models is not several populations.

Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.

Large vision language models, such as CLIPs, have revolutionized modern machine learning. CLIPs have demonstrated great generalizability under distribution shifts, supported by an increasing body of literature. However, the evaluation datasets for CLIPs are variations primarily designed for ImageNet benchmarks, which may not fully reflect the extent to which CLIPs, e.g., pre-trained on LAION, robust to spurious correlations. To bridge the gap, we collect a real-world dataset called CounterAnimal that contains realistic spurious features found in animal photos. CounterAnimal consists of a) the common group: comprising animals on common backgrounds, and b) the counter group: including animals on unusual backgrounds. The performance drops from the common to counter groups quantify the reliance of models on spurious features (i.e., backgrounds) to predict the animals. We find that CLIPs trained on either LAION or the OpenAI data exhibit notable performance drops on the counter group. Surprisingly, we observe that single-modal models trained on ImageNet are more robust than CLIPs. We provide both theoretical and empirical explanations for why CLIPs still learn spurious features. Our findings suggest that distribution shifts remain an open problem for CLIPs, and one needs to be cautious about test setups when evaluating foundation models pre-trained on a significantly different scale and distribution.

Ask a pretrained biomedical language model whether "cortisol 28 ug/dL" and "stock-market volatility" are related, and it returns a cosine similarity of 0.83 on a scale where 1.0 means identical. The two share no mechanism. This is not a corner case: every off-the-shelf biomedical encoder we tested (BioBERT, PubMedBERT, BioM-ELECTRA) scores unrelated cross-domain pairs between 0.76 and 0.92 when the answer should be near zero. Accuracy on cross-domain discrimination is 0%. Retrieval systems survive this, because a language model downstream filters the noise. A Large Behavioural Model (LBM), a foundation model whose subject is a person rather than a sentence, does not: it reasons over a graph of a user's life and treats embedding proximity as evidence that two events are causally linked. False proximity writes a false causal edge, and everything downstream inherits the error. Here, embedding geometry is not a tuning knob; it is correctness. We report the fix. A contrastive pass over 72,034 pairs raises PubMedBERT BIOSSES correlation from 0.633 to 0.828 and within-vs-across-domain separation from 1.05x to 1.63x. A second pass, BODHI, mines hard negatives from edges absent in a biomedical knowledge graph and lifts separation to 2.30x and the discrimination gap to +0.392, at a 4.5% BIOSSES cost. On an Intel Xeon 6737P with AMX, OpenVINO cuts single-query latency from 1367 ms to 10 ms (133x) and reaches 555 sentences/sec. One finding contradicts standard advice: FP16 beats INT8 on this silicon at every serving batch size, and we explain why. The same model on a no-AMX Ice Lake instance runs 13-27x slower. We release the benchmark suite, training corpora, the BODHI generator, and the OpenVINO scripts.

While there are numerous benchmarks comparing the performance of modern language models (LMs), end-task evaluations often conflate notions of *factual accuracy* ("truth") and *reasoning ability* ("rationality", or "honesty" in the sense of correctly reporting implications of beliefs). Our goal is a dataset that clearly distinguishes these two notions. Our approach is to leverage and extend a collection of human-annotated *entailment trees*, engineered to express both good and bad chains of reasoning, and using a mixture of true and false facts, in particular including counterfactual examples, to avoid belief bias (also known as the "content effect"). The resulting dataset, called BaRDa, contains 3000 entailments (1787 valid, 1213 invalid), using 6681 true and 2319 false statements. Testing on four GPT-series models, GPT3(curie)/GPT3(davinici)/3.5/4, we find factual accuracy (truth) scores of 74.1/80.6/82.6/87.1 and reasoning accuracy scores of 63.1/78.0/71.8/79.2. This shows the clear progression of models towards improved factual accuracy and entailment reasoning, and the dataset provides a new benchmark that more cleanly separates and quantifies these two notions.

A growing literature argues that progress in human longevity is slowing, and that this slowdown may signal an approaching biological limit to lifespan. The argument rests on visible flattening of cumulative life expectancy records, evaluated through linear regressions or through comparisons of average gains across decades. These summaries, however, provide only weak evidence for slowing improvements, because they cannot distinguish a true from an apparent slowdown. An apparent slowdown may arise due to periods of low variance around unchanged mean annual gains. Applying a Bayesian change-point framework to annual gains, both at the best-practice frontier and separately for each record-holding country, we find no evidence that the expected annual gain has changed across nearly two centuries, but strong evidence that its variance has moved through distinct historical regimes, with progress most regular in recent decades. The recent flattening is consistent with a low-variance regime and provides little evidence for a decline in the underlying mean pace of improvement. Claims that the best-practice frontier signals a biological ceiling should therefore not rest on the recent bend in the cumulative record alone. The frontier is better read as an empirical benchmark, the highest country-specific survival achieved under observed historical conditions.

We set out to reproduce a recent hallucination-mitigation result on the POPE benchmark. The LLaVA-1.5-7B baseline we measured, 82.21 F1 on the adversarial split, sat 1.21 to 8.67 points above the adversarial baselines we planned to compare against (81.00, 79.30, and 73.54). We looked at the evaluation code and found a readout choice that moves F1 on its own. A two-token logit readout decides yes versus no by comparing two vocabulary IDs, 3582 ("yes") and 1217 ("no"). Across 9,000 greedy-decode questions spanning all three splits, neither ID is generated even once. The model emits " Yes" (token 3869) and " No" (token 1939), because SentencePiece prepends a space byte after the prompt suffix "ASSISTANT:". Read with an eight-token lookup that covers the space-prefixed surface forms, LLaVA-1.5-7B scores F1 0.8221, 0.8498, and 0.8713 on the adversarial, popular, and random splits. That 6.13-point adversarial shift exceeds the adversarial gain reported by five of the six papers we surveyed. None of them states its readout, and our reference ran under 4-bit NF4 quantization, so we cannot say how much of any reported gain a matched evaluation would keep. That comparison stays unresolved until readout and precision are matched. On the fixed baseline we run nine inference-time corrections of our own (none beat it) and dissect why the adversarial split resists logit-level intervention. In a 3,000-question probe, wrong predictions draw more image attention than correct ones (Cohen's d=-0.46, p=1.4times10^{-24}), the opposite of what grounding-based corrections assume. A four-readout audit across four further VLMs from three tokenizer families confirms the confound is protocol-level, not model-specific: porting the LLaVA-1.5 IDs verbatim drops F1 to 0.00 on the two disjoint-vocabulary models, whereas a tokenizer-derived readout holds between 0.80 and 0.92.

We investigate how looped transformers encode human preference, training lightweight evaluator heads on frozen Ouro-2.6B loop-iteration states on Anthropic HH-RLHF. v2: an erratum is prepended; the original manuscript is unchanged. A post-publication audit found the three headline results inflated by two independent evaluation errors. The 95.2% pairwise evaluator accuracy is a canonical-ordering artifact: the data were correctly split, but the evaluator learned to prefer the first-presented argument; its strict antisymmetrized accuracy on the full 8,552-pair test set is 63.9%. The 84.5% pairwise probe and the below-chance 21.75% pointwise probe were source-item leaks (orientation rows and pair partners crossing the train/test split); corrected pair-disjoint values are 56.5% and 54.2% -- above chance, so the "inverted polarity" finding is withdrawn. The central finding survives at much smaller magnitude: preference is decoded more accurately relationally than pointwise (paired +2.3 points, 95% CI [+1.3, +3.3]), and the antisymmetrized evaluator still beats the linear probe, but no corrected readout rivals end-to-end reward models. The methodological findings stand (constant-output degeneracy, flip test, swap-protocol metric deflation), with one correction: antisymmetrized accuracy, not antisymmetry correlation, certifies relational discrimination. The two errors are mutually invisible -- a split audit cannot see an ordering prior, antisymmetrization cannot see a leak -- so both checks are required. Full audit in the follow-up work (Kirin, 2026, in preparation).

Directional ablation removes an aligned language model's ability to refuse by projecting a single "refusal direction" out of the weights that write the residual stream. It needs no gradient-based training and no optimization, only a few hundred contrastive prompts, which makes it the canonical white-box attack on open-weight alignment. However, it has been established only on dense models up to roughly 70B parameters. We study whether it survives the shift to frontier mixture-of-experts (MoE) models whose residual streams are no longer a single tensor and whose weights ship quantized. We apply it to GLM-5.3-Flash (320B parameters, 288 routed experts, a four-wide hyper-connection residual, block-FP8). The attack survives the architecture, but what it reaches is no longer where a reader of the original recipe would look for it. Editing the attention, dense and routed-expert writers on their own removes 0.039, 0.016 and 0.148 of refusal respectively; editing all three together removes 0.776. As a result, 74% of the effect exists only under the joint intervention. The part the conventional recipe reaches by module-name matching accounts for 0.066 of that 0.776, which is why it fails silently on an MoE. The effect does not follow from removing just any direction: ablating a random direction orthogonal to it leaves refusal unchanged. A category-concentrated residue survives every edit we tried: subspaces fitted on violence, sexual content and hate leave measurable refusal at every rank from 1 to 12. We report the method, the 41-89 percentage-point reductions it achieves across seven harmful benchmarks with no detected change in capability, and the boundary where it stops.

We study the degree of reliability of extrapolation of complex electromagnetic permittivity functions based on their analyticity properties. Given two analytic functions, representing extrapolants of the same experimental data, we examine how much they can differ at an extrapolation point outside of the experimentally accessible frequency band. We give a sharp upper bound on the worst case extrapolation error, in terms of a solution of an integral equation of Fredholm type. We conjecture and give numerical evidence that this bound exhibits a power law precision deterioration as one moves further away from the frequency band containing measurement data.

Understanding a narrative means knowing how every event relates in time to every other, yet large language models (LLMs) are mostly tested on isolated event pairs or coarse labels. We ask five LLMs to reconstruct the complete network of Allen's 13 interval relations for 360 English narratives whose events are placed relative to one another rather than by dates, and evaluate the outputs under a pre-registered protocol for agreement with our gold annotations, converse consistency and logical satisfiability. End to end, the best model clearly beats trivial baselines on macro-F1, our primary metric, yet under our pre-registered parser no model matches the accuracy of always answering "before", which the skewed labels inflate. Most of that accuracy gap is a parsing artifact: formatting errors in large matrices cost whole rows, and a post hoc parser that recovers the determined cells of malformed rows raises coverage from 0.60 to 0.84 and closes the gap to the baseline for the best model. On the cells they do answer, models recover precedence and containment but rarely overlap, and they fail most on flashbacks, where both the narration and the event list present the later event first. Logical consistency varies widely yet says little about agreement with gold: enforcing it raises accuracy mainly by filling empty cells, and is beaten there by the trivial control of guessing the majority label. Reliable structured output and robust handling of relative time are prerequisites for temporal reasoning over whole narratives.

False negatives remain a critical limitation of computer-aided diagnosis (CAD) systems for breast cancer screening due to delayed detection and treatment. To address this issue, we propose a counterfactual data augmentation strategy that generates healthy mammograms by "erasing" lesions from anomalous images, thereby enriching the training distribution. We train a Denoising Diffusion Probabilistic Model on BI-RADS 1 (healthy) mammograms and use a RePaint-based sampling strategy to inpaint realistic normal tissue within annotated lesion bounding boxes. The resulting healthy counterfactuals replace annotated lesion regions with realistic healthy tissue while preserving patient-specific anatomical structure, as supported by similarity metrics between real and generated images. Image realism was further assessed by radiologists and found to be consistent with the original dataset quality. We evaluate counterfactual augmentation across four representative classifier architectures: a convolutional neural network (ConvNeXt), a vision transformer (ViT), a vision-language model pre-trained on mammogram-report pairs (Mammo-CLIP) and a multi-scale attention-based multiple-instance learning framework (FPN-MIL). Experiments conducted on the VinDr-Mammo dataset show improvements in sensitivity across all architectures, particularly at 80\% fixed specificity, contributing towards more reliable CAD systems for breast cancer. Code is available at: https://github.com/ines03garcia/diffusion-based-counterfactual-generation.

Strong empirical evidence that one machine-learning algorithm A outperforms another one B ideally calls for multiple trials optimizing the learning pipeline over sources of variation such as data sampling, data augmentation, parameter initialization, and hyperparameters choices. This is prohibitively expensive, and corners are cut to reach conclusions. We model the whole benchmarking process, revealing that variance due to data sampling, parameter initialization and hyperparameter choice impact markedly the results. We analyze the predominant comparison methods used today in the light of this variance. We show a counter-intuitive result that adding more sources of variation to an imperfect estimator approaches better the ideal estimator at a 51 times reduction in compute cost. Building on these results, we study the error rate of detecting improvements, on five different deep-learning tasks/architectures. This study leads us to propose recommendations for performance comparisons.

Statistical data fusion combines two panels that share a block of covariates but observe disjoint outcome blocks, and in its traditional form no row observes both outcomes at once. That rules out the discriminative criterion one would rather train a Deep Boltzmann Machine with, since multi-prediction training needs ground truth for whatever it holds out. We propose observed-block multi-prediction, which restricts the multi-prediction objective to targets drawn from what each row actually observes. It is well defined for any missingness pattern and reduces to the original criterion when rows are complete. Having a discriminative criterion that survives the setting lets us ask whether the joint model is needed at all, by separating what it contributes into a representation part and an inference part. On two consumer panels, on grids over sample size and covariate width spanning 35 cells and 875 runs, the fine-tuned DBM is the best of fifteen methods in every cell; but almost none of that advantage comes from generative pre-training, which is confined to the smallest sample size on one dataset and absent on the other. It comes from conditioning on one outcome block when predicting the other. This term amounts to +0.19 and +0.07 percentage points, is positive in all 35 cells, and, unlike every other contribution we measure, neither decays as the panels grow, nor requires a second hidden layer, nor requires more inference. Permuting one outcome block to destroy its association with the other removes the gain entirely, which is what the account predicts. The margins are small. But a small effect that does not decay is a different object from one that does, because it rests on evidence that no model mapping covariates to outcomes can accept.

Explainability is increasingly required for perception models in intelligent vehicles, yet whether explanations remain faithful under driving domain shift is still poorly understood. This work audits post-hoc explanations of a fixed YOLOv8s pedestrian detector across PIE and JAAD using ROI-based D-Deletion, frozen confidence terciles, rank-based tests, bootstrap intervals, and Holm correction. The audit shows that deletion-based faithfulness is strongly coupled to detection strength at explanation time, with Spearman correlations between 0.70 and 0.82 for D-RISE, making naive confidence-stratified comparisons unreliable. After controlling for detection strength within fixed f0 bins, D-RISE faithfulness remains domain-dependent in the central f0 range, with PIE showing higher D-Deletion than JAAD and Holm-adjusted significance. A non-perturbative EigenCAM baseline is less faithful than D-RISE but also exhibits score coupling, suggesting that the effect is not specific to D-RISE and is related to the deletion-based evaluation setup. These results motivate confidence-controlled XAI audits for safety-critical perception under domain shift.

We introduce a rank-statistic approximation of f-divergences that avoids explicit density-ratio estimation by working directly with the distribution of ranks. For a resolution parameter K, we map the mismatch between two univariate distributions μ and ν to a rank histogram on { 0, ldots, K} and measure its deviation from uniformity via a discrete f-divergence, yielding a rank-statistic divergence estimator. We prove that the resulting estimator of the divergence is monotone in K, is always a lower bound of the true f-divergence, and we establish quantitative convergence rates for Ktoinfty under mild regularity of the quantile-domain density ratio. To handle high-dimensional data, we define the sliced rank-statistic f-divergence by averaging the univariate construction over random projections, and we provide convergence results for the sliced limit as well. We also derive finite-sample deviation bounds along with asymptotic normality results for the estimator. Finally, we empirically validate the approach by benchmarking against neural baselines and illustrating its use as a learning objective in generative modeling experiments.

AI models might use deceptive strategies as part of scheming or misaligned behaviour. Monitoring outputs alone is insufficient, since the AI might produce seemingly benign outputs while their internal reasoning is misaligned. We thus evaluate if linear probes can robustly detect deception by monitoring model activations. We test two probe-training datasets, one with contrasting instructions to be honest or deceptive (following Zou et al., 2023) and one of responses to simple roleplaying scenarios. We test whether these probes generalize to realistic settings where Llama-3.3-70B-Instruct behaves deceptively, such as concealing insider trading (Scheurer et al., 2023) and purposely underperforming on safety evaluations (Benton et al., 2024). We find that our probe distinguishes honest and deceptive responses with AUROCs between 0.96 and 0.999 on our evaluation datasets. If we set the decision threshold to have a 1% false positive rate on chat data not related to deception, our probe catches 95-99% of the deceptive responses. Overall we think white-box probes are promising for future monitoring systems, but current performance is insufficient as a robust defence against deception. Our probes' outputs can be viewed at data.apolloresearch.ai/dd and our code at github.com/ApolloResearch/deception-detection.

Magnetic Resonance Imaging (MRI) is often acquired with anisotropic resolution to reduce scan time, producing stair-step artifacts along the through-plane direction. In through-plane MRI super-resolution, an efficiency-fidelity trade-off arises: feed-forward regressors are fast but oversmooth at large slice-thicknesses, while sampling-based methods improve fidelity at high inference cost. We propose DRIFT, a two-stage thickness-conditioned rectified flow framework for through-plane MRI super-resolution with continuous input slice-thickness. Stage 1 employs an Anatomical Projection Network (APN) to map low-resolution patches to a coarse high-resolution manifold, providing a deterministic anatomical initialization that shortens the residual transport of Stage 2 and stabilizes slice-wise refinement. Stage 2 refines details via rectified flow and introduces a Physics-Aware Difficulty (PAD) metric derived from slice-thickness induced through-plane bandwidth deficit to guide an Adaptive Integration Scheduler (AIS), allocating ODE steps by thickness. A Consistent Endpoint Trajectory Alignment (CETA) loss enforces thickness-consistent reconstructions. Experiments show that DRIFT outperforms super-resolution baselines while reducing inference cost. Code, models, and interactive demos are available at https://yoonseokchoi-ai.github.io/drift-eccv2026/.

Context can change whether a request is harmful without changing its topic or surface form. We ask whether residual-stream probes distinguish harmful requests from surface-matched benign controls at a useful operating point. Across three 7-8B model families, an activation sensor blocks 95.5-97.7 percent of judge-classified compliant attacks in a taxonomy-selected set. It also blocks 59.6-68.4 percent of XSTest prompts. A fully disjoint audit reconstructs near-ceiling source-contrast AUROC (0.996-0.999), but fixed transfer to matched pairs is weaker: 0.656-0.819 on the guard-selected Twin-n70 subset and 0.590-0.690 on the full Twin-n163 cohort. We test ten axes on the reference family and seven across all families with leakage, hold-out, and permutation controls. On Twin-n163, no axis evaluated without direct pair-boundary fitting reaches the specified numerical threshold. Requiring persistence on that full cohort was added at analysis time. A separately specified 24B/32B extension gives the same result. Pair-trained classifiers weaken under category and generation-batch hold-out and false-block 79.6-100 percent of XSTest at 95 percent in-corpus TPR. At the tested read points, these activation scores behave as broad-risk detectors rather than standalone context adjudicators.

On-policy distillation (OPD) is widely used for LLM post-training. When pushed with a reward-extrapolation coefficient lambda > 1, the student can lift past the teacher in domain, but past a threshold lambda* the same step violates the output contract on structured-output tasks. In a single-position Bernoulli reduction, we derive a closed-form base-relative clip-safety threshold lambda*(p,b,c) determined by three measurable quantities: the teacher modal probability, the warm-start mass, and the importance-sampling clip strength. Above lambda*, the extrapolated fixed point exits the clip-safe region, changing training from format-preserving to format-collapsing. We extend the rule to calibrated K-ary listwise JSON tasks where a single binding equivalence class dominates the output contract and SFT retains parse headroom. On Amazon Fashion, three pre-registered tests--a fine-grid cliff interval, a budget-extension test, and a small-clip cross-prediction--fall within their locked prediction windows, with the small-clip value matching the closed-form prediction below grid resolution. Operating just below lambda*, ListOPD brings a 1.7B Qwen3 student to in-domain parity with an 8B-SFT baseline at one-fifth the parameters. The gain is driven primarily by format adherence: NDCG@1 on parsed outputs remains flat across lambda, while parse validity sharply changes at the predicted boundary. The cliff diagnostic is rubric-independent, whereas the parity claim uses a Gemini-graded rubric and inherits that evaluator's exposure.

Recent advances in volumetric super-resolution (SR) have demonstrated strong performance in medical and scientific imaging, with transformer- and CNN-based approaches achieving impressive results even at extreme scaling factors. In this work, we show that much of this performance stems from training on downsampled data rather than real low-resolution scans. This reliance on downsampling is partly driven by the scarcity of paired high- and low-resolution 3D datasets. To address this, we introduce VoDaSuRe, a large-scale volumetric dataset containing paired high- and low-resolution scans. When training models on VoDaSuRe, we reveal a significant discrepancy: SR models trained on downsampled data produce substantially sharper predictions than those trained on real low-resolution scans, which smooth fine structures. Conversely, applying models trained on downsampled data to real scans preserves more structure but is inaccurate. Our findings suggest that current SR methods are overstated - when applied to real data, they do not recover structures lost in low-resolution scans and instead predict a smoothed average. We argue that progress in deep learning-based volumetric SR requires datasets with paired real scans of high complexity, such as VoDaSuRe. Our dataset and code are publicly available through: https://augusthoeg.github.io/VoDaSuRe/

Fine-tuning a large language model (LLM) on harmful data in a narrow domain, such as bad medical advice, can make it broadly misaligned. How realistic this risk is remains unclear, since models are very rarely fine-tuned on explicitly harmful data in practice. In this work, we show that fine-tuning on data the model believes to be false but that has no harmful content, can also induce broad misalignment. Across mathematics, trivia, cooking and sports, the more strongly the model disbelieves the answers it is trained on, the more misaligned it becomes. In mathematics, misaligned answers rise from 4% to over 60% as the training answers become more wrong, with little loss of general capability. After fine-tuning, the model states the wrong answers but its multiple-choice probabilities still prefer the true ones, which suggests that misalignment comes from this internal inconsistency. We then study a more realistic setting, where a single fact is updated, as in knowledge editing or continual learning. The same relationship holds and the more the model still believes the true fact after training on the false one, the more misaligned it becomes. When we frame the same false fact as part of a coherent story, the model comes to believe it more deeply and becomes less misaligned than when we instruct it to assert the false fact. Content filters and persona vectors cannot tell these two framings apart, even though they lead to different levels of misalignment. These results suggest a risk for model updates, where training a model that does not adopt the new fact as true can induce misalignment even when the data looks harmless.

Minimizing inconsistencies across successive versions of an AI system is as crucial as reducing the overall error. In image classification, such inconsistencies manifest as negative flips, where an updated model misclassifies test samples that were previously classified correctly. This issue becomes increasingly pronounced as the number of training classes grows over time, since adding new categories reduces the margin of each class and may introduce conflicting patterns that undermine their learning process, thereby degrading performance on the original subset. To mitigate negative flips, we propose a novel approach that preserves the margins of the original model while learning an improved one. Our method encourages a larger relative margin between the previously learned and newly introduced classes by introducing an explicit margin-calibration term on the logits. However, overly constraining the logit margin for the new classes can significantly degrade their accuracy compared to a new independently trained model. To address this, we integrate a double-source focal distillation loss with the previous model and a new independently trained model, learning an appropriate decision margin from both old and new data, even under a logit margin calibration. Extensive experiments on image classification benchmarks demonstrate that our approach consistently reduces the negative flip rate with high overall accuracy.

Counterfactual generation seeks to sample outcomes under a hypothetical intervention or decision using observational data collected under the factual assignment mechanism. We develop a flow-matching approach that combines a sample-split, doubly robust training objective with a learned coupling between observed source outcomes and target outcomes drawn from a fitted conditional outcome model. To enable finite-step generation, we leverage a score-corrected stochastic sampler based on a Gaussian-smoothed interpolation. Our main theoretical contribution is a coupling-sensitive KL bound for constant-step Euler discretization: the error is controlled by moments of the source--target displacement under the chosen coupling, rather than by global uniform regularity of the velocity field, and has near-linear dependence on the ambient dimension. We also establish finite-sample non-parametric guarantees for the learned velocity and score fields when both the conditional outcome model and the source-target coupling are estimated from data. These bounds separate approximation, coupling-replacement, nuisance-estimation, generalization, and Monte Carlo errors and, combined with the sampler analysis, yield an end-to-end guarantee for counterfactual generation. Experiments on synthetic and semi-synthetic image benchmarks support the coupling-dependent theory and show that, at finite discretization budgets, the stochastic sampler can outperform the corresponding deterministic ODE sampler.