The Ceiling Is in the Channel: Auditing Learner Gaps and Measurement Frontiers in Clinical Prediction

Clinical prediction can saturate for two different reasons: a fitted learner may fail to extract available information, or the recorded variables may impose a population frontier. We separate these quantities through the \emph{learner gap} and the \emph{measurement-channel ceiling}. Optimal balanced accuracy is characterized by total-variation separation, yielding architecture invariance, a sharp partial-identification result under replacement contamination, a cross-fitted ceiling estimator, and exact conditions for multimodal decision improvement. We add two finite-sample diagnostics, namely a label-permutation optimism floor and an underfit curve, and validate the audit on three real cohorts: UCI readmission ($n=99{,}343$), BRFSS diabetes ($n=253{,}680$), and NHANES HbA1c ($n=10{,}219$). Well-tuned gradient boosting nearly reaches the estimated frontier in UCI and BRFSS, whereas deliberately or practically deficient learners retain large gaps. NHANES yields a null difference between questionnaire and measured marginal frontiers but a significant joint complementarity gain, refining the simplistic claim that an objective modality must dominate. Across all cohorts, modest AUROC gains coexist with substantially larger Bayes decision-flip rates, and several architectures estimate similar frontiers while their achieved balanced accuracy differs sharply. A PRISMA-guided synthesis of 104 clinical tasks then shows that the same channel-level regularities recur across more than 18 disease categories: a broad but non-universal structured-clinical region, diminishing same-channel gains across model families, and higher performance when measurement channels change. The framework converts saturation from an empirical observation into an auditable decision: improve the learner when headroom remains; improve measurement when it does not.

Accurate in space, unreliable in time: how LLMs represent national cultural change

Assessments of cultural alignment have become an important part of the development and improvement of large language models (LLMs). However, the majority of the evaluations treat culture as a single snapshot, investigating only whether a model represents a society accurately at the current time. Research in cultural psychology shows that cultural values change at different rates and directions over time. Therefore, a "culturally aware" model should capture not only where a culture is today but also how it has changed over time. We examine this missing dimension of cultural awareness using more than two decades of the World Values Survey data. We compare the cultural trajectories of 40 countries with the trajectories produced by four state-of-the-art (SOTA) LLMs on the Inglehart-Welzel cultural map. Our findings show that while models generally place countries close to their most recent surveyed positions, these representations tend to lag several years behind that position. They also capture only part of the magnitude of the observed change, introduce movement where little occurred, and rarely reproduce reversals in countries' trajectories. These findings point to temporal flattening and suggest that snapshot accuracy can give an incomplete picture of cultural awareness in LLMs and have implications for model evaluation, representational harms, and the governance of culturally aware AI systems.

Are Near-Tied LLM Rankings Robust to Family-DIF-Guided Benchmark Recomposition?

Small leaderboard gaps are often interpreted as evidence that one language model is better than another, but their sign may depend on which benchmark items are included. We test this using item-level responses from five benchmarks and a family-label-free spectral approximation to multidimensional item-response theory (MIRT). In owner-disjoint folds, one owner half identifies items with low residual differential item functioning across model families (low-DIF); the resulting frozen, source- and easiness-balanced weights score models in the other half, while equally short matched-random subtests control for generic subtest variation. Full-benchmark and low-DIF rankings remain strongly correlated ($τ_b=.900$--$.948$). Yet in four of five benchmarks, 30.9--47.1\% of cross-family pairs initially within one percentage point reverse order, exceeding their matched-random medians by 16.9--28.6 percentage points (all $p=.001$). The fifth benchmark shows no reliable excess ($-0.9$ points, $p=.689$). The pattern survives all pre-specified population perturbations, and residual item--family signatures replicate across owner halves; however, no family shows a consistent advantage across benchmarks. Thus, globally stable rankings can still leave individual near-tie orderings sensitive to benchmark composition, and sub-one-point leaderboard gaps should be accompanied by evidence that the implied ordering is composition-robust.

A Controlled Evaluation of Model Rankings and Input Reliance in Surface Water Segmentation

Performance evaluation for surface-water segmentation commonly uses an aggregate metric such as global intersection-over-union (IoU) to rank model configurations. However, a configuration ranking does not by itself establish why one system performs better, whether a close ordering is stable, or how strongly predictions rely on individual inputs. We examine these distinctions primarily on Sen1Floods11 through repeated configuration comparisons, paired test-chip analysis, fixed-checkpoint input stress tests, and geographic reweighting, with a targeted secondary evaluation of supervised input configurations on GEOID-Flood. The cross-modal student achieves the highest three-seed mean IoU on Sen1Floods11, but close orderings vary across seeds and geographic weighting, while ancillary-input rankings differ between Swin-UNet and U-Net. The GEOID-Flood evaluation shows substantial agreement in supervised ancillary-input effects, although the exact architecture ordering remains configuration dependent. Fixed-checkpoint tests further establish reliance on terrain and WorldCover without establishing a clean-input performance benefit, while target semantics and the later WorldCover prior restrict the evaluation to retrospective all-water segmentation. These results show that aggregate metrics remain useful for ranking complete configurations, but ranking stability, component attribution, input reliance, and deployment scope require distinct evidence. Performance evaluation should therefore match the evidence reported to the claim being made.

Evaluating Tiny Recursive Models Across Training for Code Generation

Code generation increasingly relies on large transformer models, whose capability advances with scale. Yet such a scale is costly, creating demand for small models, especially where data is limited. Recursive models address this by reusing a single block to add depth rather than stacking independent layers. Such models are typically evaluated by teacher-forced fit (next-token loss on ground-truth prefixes) or task accuracy, at a single checkpoint, whereas code is produced by free-running generation, where the model extends its own output. Whether a teacher-forced advantage survives free-running generation, and whether it holds across training, remains open. To study both, we compare a ~28M-parameter autoregressive Tiny Recursive Model (TRM-AR) on natural-language-to-Python code generation against parameter-matched and depth-matched controls, tracking fit and generation across 40 epochs and three seeds. The fit ranking between the recursive model and the depth-matched control reverses twice. Selecting each checkpoint by validation loss and examining the trajectory yields a consistent comparison. At equal parameters, TRM-AR fits, generates, and generalizes better than the parameter-matched control while recovering approximately 45% of the validation-loss gap and 57% of the generation-quality gap between the two controls, at roughly 175 times the per-step cost of the parameter-matched control. However, at equal effective depth, the larger transformer fits and generates better at its validation optimum, suggesting TRM-AR's advantage lies in resistance to overfitting, not greater capability. These findings suggest that recursive code generation models should be evaluated jointly on fit and generation across the training trajectory rather than at a single checkpoint.

When Linguistic and Internal Confidence Diverge in Large Language Models

Users often ask large language models (LLMs) to report how confident they are, but it is unclear whether such linguistic confidence tracks the model's internal confidence. We study this question across 8 classification tasks, 2 generation tasks and 30 models from three families. For classification, we compare linguistic confidence with logits-based confidence along three axes: association, magnitude agreement and calibration. For generation, we test whether linguistic confidence tracks semantic-entropy-based uncertainty. The axes frequently diverge. Instance-level association is weak on average, although it improves on easier items and for stronger base models. Instruction-tuned models often report higher confidence and sometimes show higher association, but they also have larger confidence gaps and worse calibration. Prompt design mostly changes the distribution of reported confidence. Attitude cues inflate confidence without improving alignment, while score exemplars can preserve rank-order signal when they avoid collapsed confidence values. Regression analyses show that distributional properties of confidence scores explain much of the observed alignment pattern, with model metadata playing a smaller role after controls. These results support a lossy-channel view of linguistic confidence. A more dispersed verbal confidence distribution can carry useful rank information, but it does not make the scores calibrated. Linguistic confidence should therefore be evaluated with multi-axis diagnostics before being used in downstream reliability pipelines.

A Statistical Audit of Physical AI Benchmark Redundancy

Physical AI models are evaluated on suites of benchmarks that differ across model reports, leaving the model-by-benchmark matrix sparse and the relationship between benchmarks unmeasured. We construct a matrix of 51 models on 12 physical AI benchmarks, selected from a registry of 51 benchmarks and 152 models by reporting density, combining scores from model cards and benchmark papers with our own evaluation runs under each benchmark's official protocol. We measure how much information the benchmarks share and show quantitative evidence of Redundancy. Redundancy affects reported rankings: collapsing the two substitute pairs into single columns moves 22 of 51 models by three or more places under an equally weighted average. We then select benchmarks greedily under a utility combining score dispersion with variance not explained by the already-selected set, and obtain a four-benchmark subset retaining 78.5\% of the utility of all 12, on which we fit a Bradley--Terry ranking. The procedure requires only benchmark-level scores with sufficient overlap and is not specific to physical AI.

What Do Medical Vision-Language Models Learn in Radiology? Transfer, Alignment, and Source-Proxy Leakage Under Distribution Shift

Medical vision-language models (VLMs) can appear reliable in-domain while failing when acquisition domain, paired supervision, or evaluation protocol changes. We study this failure mode as a representation-level blind spot relevant to epistemic intelligence, without claiming a formal estimator of epistemic uncertainty. Using NIH ChestXray14 and CheXpert, we first isolate source-only cross-dataset visual transfer from unsupervised domain-adaptation diagnostics. Using PadChest and OpenI, we then evaluate multimodal alignment under strict pair-index retrieval and quantify metadata-derived source-proxy information retained in frozen embeddings. Self-supervised visual initialization improves NIH-to-CheXpert transfer over supervised ImageNet initialization in matched ResNet-18 comparisons, whereas adversarial adaptation is useful only in a narrow regime and becomes unstable as adversarial pressure increases. Multimodal exact-pair retrieval remains low under external OpenI stress testing, and source-proxy information remains recoverable from learned representations. Qualitative nearest-neighbor and Grad-CAM analyses show clinically plausible cross-dataset structure and thoracic attention patterns in many cases, while device-heavy and false-positive cases remain ambiguous. Auxiliary architecture checks are task-dependent and do not support a universal backbone ranking. Overall, the study shows that apparent competence under a single protocol can conceal transfer, alignment, and shortcut-related failure modes, motivating stress-tested evaluation of medical VLMs under distribution shift.

Causal Analysis for Time Series Foundation Models

Transitioning from bespoke time series models towards time series foundation models changes the relationship of model and application from one-to-one to one-to-many. This shift introduces concentration risk as many, potentially high-risk, forecasting applications are exposed to the same biases and failure modes of a single time series foundation model. At the same time, this centralization allows for economies of scale in model development and validation. In this study we investigate how biases and failure modes of time series foundation models can be identified before deployment. We propose a causal analysis framework to investigate the ability of a time series foundation model to preserve time series patterns. To achieve this, we intervene on parameterized synthetic time series generators and measure the corresponding change in model output under ceteris paribus conditions. We apply our causal analysis framework to Chronos-2 and TimesFM-2.5 and test them across six distinct time series patterns. We find safe configurations for trend and harmonic oscillation patterns. The results also indicate a bias in both models towards overestimating persistence, sudden failures for both models against the regime switch pattern and failure for TimesFM-2.5 against the energy-release pattern. Our review of the original works for both models indicates that the findings might be explained by the data used for pretraining. We conclude our study with suggestions for further model development, recommendations for application-specific model selection, and a discussion of limitations and further research directions.

PhysicsBench: A Unified Leaderboard for Generative and Predictive Models in Engineering Design and Simulation

Generative and predictive artificial intelligence models are increasingly used to generate geometry and to predict physical fields and scalar quantities in engineering design and simulation. Yet these models are typically evaluated in isolation, on academic datasets at unconstrained scales, with inconsistent metrics and procedures. We present PhysicsBench, a unified benchmark and leaderboard that evaluates generative and predictive models under one standardized procedure. PhysicsBench spans seven generation and prediction tasks across 1D, 2D, and 3D domains and ranks 66 models on nine datasets, comprising industrial-scale CAD/CFD/FEA simulations and public references, expanded into 28 configurations. One procedure and ranking apply to both families, each ranked within its own tasks. Evaluation spans realistic, limited data scales from S to XL rather than the unlimited training sets common in academic benchmarks. A common metric suite captures geometric fidelity with distributional distances, physical-field and scalar accuracy, and engineering-specific field- and shape-validity. BenchRank debiases correlated metrics and ranks by PageRank over a head-to-head dominance graph, so every reported quality metric is also ranked, with computational cost in a separate efficiency view. Across tasks, an architecture's large-scale academic standing weakly predicts its small-data ranking. The top model changes with data scale in six of the seven tasks, and no model leads more than one task. PhysicsBench turns "state-of-the-art" from a self-reported claim into an openly published foundation for model selection.

STONIC: A Layered Measurement Contract for LLM Value Profiling

LLM value studies often merge questionnaire ratings, pairwise choices, and values inferred from generated text into one profile. That merge assumes that the three observations describe the same stable preference. STONIC tests this assumption on 5,144 situations from four banks and 35 fixed model configurations. It compares responses rated in isolation, choices made under counterbalanced conflict, spontaneous answers, and later choices between a model's own answer and authored alternatives. 10 of 17 configurations with usable behavioral data preserve the endorsement-choice relation across banks. Every one of the 17 eligible configurations prefers its own earlier answer (median effect 0.790), although option position changes the choice rate in every eligible configuration. Profile shape transfers most strongly from ratings to conflict choices and weakens for spontaneous text. Three-way annotation of 200 L3 responses provides a task-local check of the semantic audit: FULCRA agrees most closely with the human majority, while DeBERTa retains useful rank information after calibration. Hidden states encode the completed decision more clearly than the prompt alone. Thus the models show reproducible behavioral continuity, but the evidence does not support one scorer-independent value identity across interfaces.

A Comparative Study of Label-free Representation Quality Metrics in Deep Learning

We present a comparative study of label-free metrics for assessing the quality of representations in deep neural networks to understand their reliability under a wide variety of configurations. We group existing label-free metrics into three families based on their construction and analytically establish connections between metrics within the same family. We then characterise the sensitivity of spectral metrics through controlled synthetic experiments. Finally, all label-free metrics are evaluated against downstream task accuracy across a diverse set of 260 vision models on six datasets spanning generic object classification, fine-grained object classification, scene recognition and geospatial task, stratifying results by architecture class and training objective. We find that intrinsic dimensionality (ID) is the most reliable predictor among the metrics considered. However, the reliability of all metrics, including ID, is moderated by architecture class and training objective. Our results provide a clearer understanding of what label-free representation quality metrics measure, when they are reliable, and how to interpret them in practice.

DefaultShift: Auditing Semantic Default Shift in Accelerated Text-to-Image Models

Few-step text-to-image models increasingly replace slower generators, yet acceleration can silently change distributions over unspecified attributes even when individual outputs remain plausible and aligned. We call these distributions semantic defaults and their change under replacement semantic default shift. Existing quality, preference, and diversity evaluations do not test whether a replacement preserves its reference model's semantic defaults. We introduce DefaultShift, a paired audit that labels repeated samples with closed semantic vocabularies, measures probability-mass movement, and separates interpretable ranking from confirmatory cross-fit inference. Across 14 reference and replacement pairs, adjusted color discrepancies range from 0.054 to 0.303 with recipe-specific directions. A 1,000-image human audit reproduces the ordering. We further introduce DefaultShift-Select, an offline calibration method that reduces human-measured shift by 10.3 percent to 35.1 percent across Turbo, DMD2, and FLUX without material quality loss. Under balanced evaluation, selected data recover 4.3 accuracy points and 7.5 worst-group points over uncalibrated replacement data. DefaultShift makes semantic preservation under acceleration measurable and actionable.

What Aggregate Scores Miss: Measuring Item-Level Regressions in Commercial LLM API Migrations

Context: Software systems that depend on commercial large language model APIs must migrate to successor versions when vendors deprecate older models. Migration decisions typically rely on aggregate benchmark scores, which compress heterogeneous item-level behaviour into a single net figure. Objective: We measure what that compression conceals. Method: On three pairwise upgrades in the GPT-5.4 to GPT-5.6 Sol product sequence, we query 900 public benchmark items (graduate-level knowledge, olympiad mathematics, instruction following) 50 times per item per model, classify each item as reliably improved, reliably regressed, practically equivalent, or inconclusive under false-discovery-rate control and a practical-significance threshold, and calibrate the results against a label-permutation null. Results: Across all nine migration-benchmark cells, reliable improvements and reliable regressions coexist. Edges with aggregate gains of up to 7.3 percentage points contain up to 8.3% reliably regressed items; edges with aggregate losses contain up to 10.7% reliably improved items. On the instruction-following benchmark, the gap between strict and loose scoring widens by 3.9 percentage points on the latest migration: a 3.9-point regression under strict scoring shrinks to 0.04 points under loose scoring. Conclusion: Migration decisions based on aggregate scores alone miss substantial bidirectional item-level change. The complete response-level archive and per-item scoring outputs are released.

Beyond Accuracy: Assessing Calibration of Geospatial Foundation Models and Their Sensitivity to Distribution Shifts

Geospatial Foundation Models (GeoFMs) are most commonly ranked and selected by accuracy on standard benchmark conditions via averaged ranks. We show that this protocol is too narrow: the promised deployment in critical EO tasks requires further angles of analysis, mainly calibration, the agreement between a model's confidence and its correctness. Across 16 frozen encoders, four classification and five segmentation datasets, and two orthogonal stress axes, every encoder degrades as corruption intensifies, and the ranking changes as well. Across the four classification benchmarks, EO-pretrained and ImageNet-pretrained encoders are indistinguishable on clean accuracy and clean calibration, and EO pretraining provides no more stability under shift than ImageNet pretraining. Under shift the GeoFMs drift further into overconfidence than the ImageNet-pretrained encoders, at every grade and in every corruption family. A centered kernel alignment (CKA) analysis ties this to representational rigidity: EO-pretrained embeddings move less under corruption while losing just as much task information and remaining overconfident. We apply three commonly explored uncertainty quantification methods and find that temperature scaling and deep ensembles cannot counteract the degradation, while a Gaussian-process probe roughly halves ECE under severe cloud only by tripling it on clean data. In selective prediction experiments, we find that confidence-based abstention cannot defer around confidently wrong predictions, and advocate that benchmark rankings and evaluations should therefore operate across a multitude of conditions and metrics to more holistically evaluate model development progress and close the gap to real world deployment scenarios.

Decorrelation Is Not Complementarity: Skill, Not Lineage, Governs Trusted-Monitor Ensembles

Trusted monitoring has a cheap, trusted model score a stronger untrusted model's actions, and a diverse ensemble of them beats a single stronger monitor at matched cost. They are built by minimising average pairwise correlation, and that paper's twelve monitors shared one base model, leaving open what supplies the diversity. We study 24 open-weight monitors spanning nine pretraining lineages and a 29x range of detection skill (pAUC at 10 percent FPR, 0.028 to 0.803) on backdoored code. The metric used to build panels does not predict what a panel is for, and we can say why. Agreement on attack items splits into a shared-detectability signal component and an idiosyncratic error component, which predict ensemble gain with opposite sign (Spearman -0.25 and +0.26), so their sum, the metric actually used, predicts it barely at all (+0.05); the cancellation holds in 7 of 8 evaluations. Skill acts on signal (+0.53) while error stays flat (-0.01), which is why a monitor's own skill predicts its agreement with the pool (Spearman 0.84, n = 24, permutation p below 0.0001). Pretraining lineage is the obvious way to buy decorrelation, and it does not pay. At matched member capability, cross-lineage panels detect no better (permutation p = 0.13), and lineage barely moves the metric either (+0.064, p = 0.18). We report that against ourselves: on our own 22-monitor pool the same test read +0.104 at p = 0.037 until two monitors were added. An earlier pool topping out at pAUC 0.23 had already invalidated another analysis. Such a quantity is a property of the pool assembled. Panel gain over the best member falls monotonically with panel skill (-0.66 at k = 2, -0.70 at k = 3), and no correlation-weighted selection beats picking the single best monitor out of sample. Across six attacker models the gain result holds in all six, the agreement and cancellation results in five of six.

The Price of Thinking: Reasoning Effort as a Model-Specific API Contract

API buyers purchase a dated contract, not a model name alone: the contract includes the requested and served model, reasoning-effort term or its omission, output rail, service product, prompt, and price schedule. We study the reasoning-effort term through a registered paired contrast of Sonnet 5 with explicit high effort against the same model with effort omitted, using 30 AIME 2026 items and five calls per item. Every paid attempt was assigned one frozen terminal category, and inference resampled items while retaining their repeated calls. Mean delivered cost was \$0.01031 per call higher under the explicit-high contract than under the omitted contract [+\$0.00204, +\$0.01974]. The corresponding accuracy contrast was +0.0133 [-0.0267, +0.0467]; we did not detect an accuracy difference, and the interval permits a gain of up to 4.67 percentage points that this design cannot rule out. Cost per correct answer was \$0.08665 under the high-effort contract and \$0.07662 under the omitted contract, as registered point estimates. A dated contract census, Models-API metadata, and preregistered raw-response probes further documented model-specific omission semantics, including within a provider; claims remained at documentation grade when raw structure was indeterminate. The request registry, parser, terminal taxonomy, statistical plan, and analysis pipeline were frozen before outcomes were examined; the resulting claims are bounded to the model, task, and collection date studied.

Demographic Injection in Medical Language Models under Diversity, Equity, and Inclusion Prompts

Clinical-AI guidance increasingly recommends prompting language models to reason with attention to diversity, equity, and inclusion (DEI). We measure a side effect that misrepresents patients: a one-sentence DEI prompt appended to a medical question leads models to add patient demographic attributes (race, socioeconomic status, sex) the question never stated, in effect rewriting who the patient is. We call this demographic injection. Across 47 models, four medical benchmarks, and 376,000 responses scored by a validated model-judge pipeline, a single DEI prompt raises the injection rate from 0.7% to 33.1% (47x) in all 47 of 47 models, attributable to the equity content rather than to added length (18x above a length-matched control; p=1.4x10^-14). Most added content is a general population statement that leaves the answer unchanged, but a smaller subset attaches an attribute to the specific patient or changes the selected option (0.25-2.4% of responses, 99.8% toward the incorrect option), where the invented demographic changes the answer the model recommends. Phrasing scales the effect from 14% to 56%. DEI prompts are just one example of a more general mechanism. Any instruction that nudges how a model reasons can make it add unrequested details, including details about the patient. Flagged outputs are treated as model errors under study, not clinical guidance.

Validation-Frontier Representation Selection under Constrained Observation

AI systems deployed outside clean benchmark settings often rely on observations that are incomplete, unstable, costly, or degraded by monitoring failures. This paper studies representation selection under constrained observation: choosing a state representation when raw accuracy is not the only operational criterion. We propose a validation-frontier selector that combines balanced accuracy with penalties for feature cost, overfit gap, and validation-test instability. In a focused public-tabular benchmark using three scikit-learn datasets, five observation regimes, 45 matched task cells, 720 candidate actions, and 405 representation rows, the adaptive selector improves frontier score over full trace features by 0.025801 while reducing mean feature count by 22.733. Balanced-accuracy difference is small and not statistically significant. A broader offline stress test gives mixed results. The supported claim is therefore bounded: adaptive representation selection can improve a constrained-observation robustness-efficiency frontier in matched benchmark settings, but does not universally dominate trace baselines.

Attributing Preprocessing Invariance in Spectral Foundation Models

Preprocessing invariance is an appealing goal for spectral foundation models: a frozen model should remain useful when laboratories preprocess spectra differently. It is usually measured by training a classifier under one preprocessing pipeline and testing it under another, with preserved accuracy read as evidence of learning. We revisit that reading, using a Raman foundation model as a case study. Such models normalize their inputs before any learned parameter is applied. If that normalization maps two differently preprocessed spectra to the same vector, the encoder receives identical inputs, so the invariance cannot be attributed to learning. For a normalization that uses each spectrum's own statistics, this happens exactly when one spectrum is a positive multiple of the other plus a constant. Several standard preprocessing operations take that form. The encoder should therefore be measured against the normalization alone, which has no learned parameters. On six Raman evaluation datasets, the model does not measurably outperform its own normalization. It improves on raw spectra, but so does the normalization alone. Training does improve the encoder over random initialization, and a controlled experiment shows that it learns to ignore a transformation only when that transformation reaches it. A numerical test settles which transformations a given normalization removes. Across released systems in five modalities, most normalizations already remove transformations of that form, and several of those systems claim that invariance as learned. Replicating the comparison on two of them shows no gain either.

Certifying What Helps Customer-Return Timing: A Screen-and-Confirm Test for Conditioning Signals, and Why Decay Is Nearly Enough

Practitioners enrich customer-return models with ever more signals (lifetime value, category, recency/frequency, calendar, geography), and the temporal-point-process (TPP) literature follows suit with covariate- and external-covariate-conditioned intensities. But does any of it improve the timing, and how would you know? A null ("feature X doesn't help") is only meaningful if the model could have found a signal. We make two contributions--a method and a measurement--to answer this credibly. (i) A screen-and-confirm protocol that certifies whether a candidate signal improves a TPP's event-timing likelihood: a positive control plants a coupling of known strength and confirms the model recovers it, so a real-data null can be read as "no signal" rather than "weak method." The control is validated for categorical and continuous encodings, and on a real clock-driven dataset (NYC taxi hour-of-day). (ii) A model-free ceiling quantifying how little of customer-return timing is point-predictable at all (a single-digit percentage of gap variance from any covariate; returns are near-memoryless). With these we certify a clean result on three public benchmarks (Amazon, Taobao, RetailRocket) and a real marketplace (Thumbtack): the inter-event clock--continuous-time decay, long known to beat frozen-intensity models--is nearly sufficient, and the conditioning the field keeps adding is redundant or harmful on top of it (statistically null on the public benchmarks, at most 0.06 NLL; null to mildly harmful on the marketplace). We do not claim to discover that decay helps; our contribution is the tools that turn "conditioning doesn't help" into a checkable, certified statement--plus an honest-evaluation account of the read-out/leakage pitfalls we hit and retracted.

When Is a Steerable Concept Representation Real? Measurement Confounds in a Cross-Family Audit of Neuroscience Parallels in LLMs

Large language models (LLMs) are increasingly reported to exhibit human-like neural and cognitive signatures, including concept cells, mental number lines, and cognitive maps. These claims often rely on linear probing and activation steering applied to a single model, yet both methods are highly sensitive to measurement choices. A reported parallel may therefore reflect the model, the measurement procedure, or both. We audit four representative neuroscience-inspired paradigms across 17 models from five families, spanning $0.6$B to $72$B parameters. Our main experiment examines the causal steerability of concept directions. With raw activation units and a fixed layer and coefficient, steerability appears to increase with model scale, resembling an emergent capability. However, this pattern is produced by an uncalibrated pipeline rather than by a claim established in the steering literature. The trend depends jointly on raw units, the readout metric, and the operating point; correcting any one of these removes it. With residual-norm-comparable interventions and held-out operating-point selection, concept steering remains significant at every scale, but shows no significant trend across the Qwen3 series, although the confidence interval does not rule out a moderate positive slope. The remaining results are mixed. A linear geographic world map is consistently decodable in every tested checkpoint up to $72$B. Number magnitude is strongly encoded, but whether individual neurons appear bell-shaped or monotonic depends on the selection criterion. Language-specific structure is localizable, but the direction of the cross-lingual asymmetry reverses under a different attribution method. These results suggest that the main constraint on AI neuroscience is not a lack of phenomena, but a lack of comparable measurements and adequate controls. We release the protocol, stimuli, and code.

Latent-Regime Bias Auditing for Volatility Forecasting

Volatility forecasts are commonly evaluated with aggregate accuracy metrics such as RMSE and MAE, but these metrics can hide conditional failures that matter for risk management. This paper proposes a model-agnostic audit framework for evaluating whether volatility forecasts remain reliable across latent market regimes. We learn time-series representations of market-state windows, cluster them into regimes using only training information, assign regimes out of sample, and compare aggregate forecast behavior with regime-conditional bias, tail-underprediction, and underprediction-sensitive economic losses. Applied to daily volatility forecasting across cryptocurrency and ETF assets, the audit shows that models with competitive aggregate accuracy can still exhibit substantial regime-specific bias and severe tail underprediction. The results suggest that volatility forecasting should be evaluated not only by average error, but also by where and how forecasts become unreliable. Our framework shifts forecast evaluation from asking which model is most accurate on average to identifying the market regimes in which apparently accurate forecasts fail conditionally. Reproducibility: https://github.com/arthurchagas1/Latent-Regime-Bias-Auditing-for-Volatility-Forecasting

Query Timing Produces Opposite Positional Biases Between LLMs and Humans

Positional biases such as recency and primacy effects have been documented in large language models (LLMs), yet the underlying mechanism by which these models make their evaluations remains poorly understood. Both primacy and recency biases have been observed in human judgments in response to evidence, but recent work suggest that \emph{when} the listener updates their beliefs -- during the presentation of evidence or only at the end -- influences the presence of such effects. We investigate whether a similar phenomenon holds for LLMs, finding divergence from human behavior. These biases are more exacerbated in newer models compared to their predecessors.

Position: It's Time to Optimize LLMs for Self-Consistency

Despite ever-increasing sophistication in language model (LM) pre- and post-training pipelines, many important failures persist: models overcondition on user framing ("sycophancy"), exhibit incomplete logical generalization, and produce confident but incorrect responses. We argue that these failures arise from a modeling assumption permeating all aspects of the pipeline: that behavior can be specified and evaluated independently on single-output pairs. Many model failures are difficult, if not impossible, to detect without reasoning about relationships between a model's responses across inputs. In this position paper, we propose self-consistency as a framework for understanding these failures. We first observe that a wide variety of techniques designed to improve specific aspects of LM behavior-targeting properties as diverse as adversarial robustness and factual coherence-can be understood as special cases of a common "consistency optimization" procedure and addressed with a standard set of optimization tools. We next outline a set of new model properties that could be achieved by optimizing for consistency, and conclude with a discussion of what it would mean to develop generally consistent LMs, including the capabilities they would enable and the objections they raise.

Language Models Agree With Each Other, Not With Readers

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.

What Is Missing in Surgical Risk Stratification and Outcome Prediction: A Scoping Review of End-to-End Machine Learning Approaches

Postoperative adverse events, including mortality and morbidity, remain a major global burden, many of which are preventable through early identification of high-risk patients and targeted perioperative care. Accurate risk stratification is therefore essential. With the growing availability of large-scale electronic health records (EHRs), machine learning (ML) provides a data-driven approach to model complex clinical patterns. However, existing studies vary widely in design, and methodological practices remain fragmented. This scoping review characterizes ML pipelines for surgical risk stratification and outcome prediction using EHR data. We reviewed 190 studies covering the ML workflow, including data preprocessing, algorithm selection, model evaluation, and explainability. Most studies relied on single-center private datasets with limited data modalities, while the scarcity of open-access surgical datasets constrained reproducibility and generalizability. Reporting of key preprocessing steps, including missing data handling, feature selection, and class imbalance, was often incomplete. Conventional ML models and simple neural networks predominated, whereas deep learning and multimodal approaches remained uncommon. Benchmark datasets and standardized evaluation protocols were largely absent, hindering cross-study comparisons. Only about one-third of studies incorporated explainability methods. This review identifies methodological gaps limiting clinically robust postoperative ML tools and provides a structured reference to support more rigorous, reproducible, and clinically meaningful ML development for perioperative care.

Performance of large language models in the optical diagnosis of colorectal polyps

Background and Study Aims: Accurate optical diagnosis of colorectal polyps guides resection strategy and surveillance, with multimodal large language models (MLLMs) showing potential for image-based diagnosis. We aimed to evaluate the diagnostic accuracy of MLLMs in classifying colorectal polyps and predicting histology. Methods: We conducted a retrospective diagnostic performance study using the PRIME dataset, a curated set of white light and narrow-band imaging (NBI) images. We evaluated Claude Opus 4, Google Gemini 2.5 Pro, GPT-o3, GPT-4o, and GPT-5. For Paris, Narrow-band Imaging Colorectal Endoscopic (NICE), and predicted histology, we calculated F1 scores, percent correct scores, and accuracy of each MLLM compared to expert responses for 132 cases. Cochran's Q and McNemar's Test were used to determine differences between predicted values of each MLLM. Results: The F1 scores among MLLMs were >0.9 for all models for neoplastic vs. non-neoplastic polyps. Gemini 2.5 Pro demonstrated the highest F1 scores for invasive vs. non-invasive polyps and low- vs. high-grade adenoma, at 0.560 and 0.492 respectively. Claude Opus 4 and GPT-5 had statistically significantly higher percent correct scores than other MLLMs at 41.7%, using Paris classification. Conclusions: Claude Opus 4 and Gemini 2.5 Pro showed the highest accuracy in differentiating polyp subtypes, performing closest to expert consensus. Sensitivity and specificity, however, did not meet ESGE standards, highlighting the need for prospective multicenter trials and the design of human-in-the-loop workflows before clinical deployment.

BayesAME: Bayesian Active Model Evaluation

Evaluating large generative models across benchmarks is time-consuming and computationally expensive. This drives the need for methods that can estimate full benchmark performance by evaluating models on only a subset of items, known as a coreset. Current literature mostly requires the practitioner to input a coreset size. However, when reliable performance estimation takes priority over efficiency, an evaluation method should also be capable of automatically determining a coreset size that reflects this priority. We introduce BayesAME, a sequential Bayesian framework specifically targeting automatic determination of the coreset size. BayesAME models performance as a random variable by defining a latent ability for each group of items sharing the same historical model performances, with a joint prior distribution encoding the belief that the target model behaves similarly to these historical models. The posterior distribution over these abilities is used to derive performance estimators, quantify performance uncertainty, and select items to add to the coreset via an information-gain criterion. The coreset is iteratively augmented until the performance estimate fluctuation and the performance uncertainty fall below their respective user-defined thresholds. We propose a multi-target extension that captures performance correlations across multiple target models to further reduce the coreset size. Through extensive experiments across diverse benchmarks, we demonstrate that BayesAME consistently outperforms sequential adaptations of existing methods. Crucially, our comprehensive analysis addresses recent skepticism in the literature, establishing that non-random coreset selection is advantageous over random selection. Finally, we highlight that leveraging continuous response log-likelihoods over traditional binary scores significantly enhances estimation accuracy.

Can Deep Generative Models Reproduce Non-Stationary Gaussian Random Fields?

Deep generative models (DGMs) are widely used for complex high-dimensional data and increasingly applied to spatial and spatio-temporal modeling. Their generated samples implicitly represent the learned data distribution and associated uncertainty. However, for real-world data, assessing whether DGMs have learned the underlying process is difficult because the ground truth is unknown and evaluation often relies on observations alone. We evaluate representative DGMs, flow matching (FM), DDPM, score-SDE, and VAE, on a known non-stationary Gaussian random field. This paper provides comprehensive metrics to assess recovery of the ground-truth mean and covariance structures, with oracle samples and a stationary control as references. All four models recover the mean surface, while their covariance recovery differs across model families: DDPM and score-SDE recover the covariance structure reasonably well, FM exhibits mildly attenuated non-stationarity and slight variance under-dispersion, and VAE has difficulty recovering the covariance structure. An experiment on ERA5 temperature anomalies further demonstrates how the framework can support the validation and development of DGMs for complex real-world spatio-temporal data.