Experimental design is commonly framed as choosing the experiment expected to provide the most information. Under partial identifiability however, persistent nuisance uncertainty can make the same observation carry different structural meanings. We introduce Resolution-Aware Experimental Design (RAED), which selects an experiment by the smallest expected nonempty structural candidate set achievable subject to false-exclusion control. We prove an exact cross-nuisance aliasing separation: an experiment can be preferred by structural and full-latent information gain, average classification, and nuisance-marginalized informativeness while having arbitrarily poorer valid structural resolution. RAED nevertheless preserves the expected ordering under a genuine composite Blackwell comparison. To make this criterion operational, we develop a learned score-based implementation with finite-sample nuisance-average and positive-tail calibration, and characterize a rare-tail sample-complexity obstruction. Under constrained sensing, two subsurface-flow benchmarks exhibit genuine RAED--expected-information-gain (EIG) experiment-selection disagreements, with the clearest and largest held-out resolution differences in WCA. In a fluvial benchmark, tail protection changes the selected physical experiment and replaces hard-region false exclusions primarily with explicit ambiguity. In a mechanistic methane-oxidation benchmark, a prospectively specified 5\% false-exclusion tolerance also yields a nontrivial finite-sample population guarantee for tail-sensitive nuisance risk, with 95\% joint confidence across all three structural families.
Multi-agent AI systems improve inference by spawning agents and synthesizing reports. But another agent is not another observation: apparently independent reports may descend from the same evidence, and genuinely independent evidence can produce nearly identical reports. We formalize this as an epistemic Sybil problem. A report Z is an epistemic Sybil extension relative to reports R when I(Theta; Z | R) = 0. No report-only aggregator can generally distinguish replication from independent corroboration: identical reports can warrant different posteriors under unobserved ancestry. A Gaussian shared-root model shows common ancestry does not imply complete redundancy. Repeated extraction adds information toward a source-level ceiling, and correlated extraction errors, which a shared base model can induce among independent agents, lower that ceiling further. We test these predictions with more than 20,000 controlled LLM-agent report and extraction calls on synthetic evidentiary documents. Holding one evidence root fixed while report multiplicity rises from 1 to 32 collapses naive posterior coverage from 0.940 to 0.263. Holding report count fixed while evidence-root multiplicity rises from 1 to 16 closes the gap, and the aggregators are statistically indistinguishable at k = 16. The agent's replicate extraction errors are correlated (gamma_cal = 0.719, estimated out of sample), and a correlated-extraction aggregator restores calibration accordingly. A controlled manipulation isolates representation similarity from evidential ancestry. It changes a report-space deduplication mechanism's mean inferred cluster count by 1.425 (95% CI [1.363, 1.485]), whereas a fourfold change in true ancestry changes it by only 0.040 ([-0.045, 0.120]). Collective inference should therefore track evidential ancestry and dependence, not agent or report multiplicity or similarity.
Multi-turn tool calling is a core evaluation scenario for large language model (LLM) agents. On public tool-calling benchmarks, open-weight models now approach or even surpass closed-source frontier models in aggregate accuracy. However, this metric averages over many different multi-turn situations and obscures whether progress is balanced across them. We propose an action-class-oriented diagnostic framework that decomposes multi-turn failures into two orthogonal modes: action-class miscalibration and action-execution failure. The framework operates over a four-class action space (TOOL_CALL/ASK/REFUSE/CONFIRM) and introduces a self-revealing upper bound Acc <= GAR (Gold Action Recall); the two modes show up as bound violation (Acc > GAR, exposing state-grader masking of miscalibration) and large bound slack (GAR >> Acc, localizing execution failure within TOOL_CALL). We validate it on a panel of tool-calling models across multiple multi-turn benchmarks. Across our panel, the diagnostic reveals action-class miscalibration as a substantial failure mode the state grader cannot see. This gap inflates standing for heavily tool-trained families, which our diagnostic separates from families with context-appropriate action choice. Calibration is reshapable through context-only perturbations, but the reshape is heterogeneous: a single perturbation moves accuracy in opposite directions across families (up to +11.5 vs -21.0 pp on the same scenario), and its effect further depends on the perturbation mechanism. We argue that multi-turn tool-calling evaluations should supplement aggregate accuracy with action-class diagnostics that expose what the model actually does in each scenario.
Language model agents increasingly propose actions, observe external feedback, and explain their own behavior. Their confidence and rationales are convenient monitoring signals, but convenience is not verification. We introduce an environment-grounded audit in which every intermediate proposal receives an exact outcome. A language model operates an evolutionary Contexto search whose feedback function assigns every valid guess an exact rank without human annotation. Across 200 runs spanning five configurations and three model families, four reporting configurations produce 12,249 self-reports. We test three assumptions: stated confidence is calibrated, inherited rationales affect later proposals, and fitness-based selection improves report quality. All three fail. Operators overstate top-100 success by factors of 4.8 to 9.3, while calibration and discrimination dissociate across model families. Controlled interventions on 754 inherited rationales bound any measured benefit of the genuine rationale to roughly 250 ranks. Neither fitness-based nor random selection produces a detectable selection differential or parent-to-offspring transmission in report accuracy, despite sharply different search behavior. Agent self-reports should therefore be treated as claims to verify against the environment, not as evidence of their own reliability.
Memory is widely viewed as an important unsolved problem for LLMs and VLMs, and current benchmarks typically evaluate it by testing accuracy over long text or video. However, accuracy alone misses properties that matter for real long-horizon tasks. We introduce ECCBench, a benchmark and evaluation protocol that measures memory beyond a system's capacity--its raw accuracy at a specific budget--via three axes we call ECC: efficiency--the computation, in FLOPs, needed to answer from memory; compression--whether compressible inputs are remembered more accurately or efficiently; and calibration--whether the system abstains in response to its own uncertainty and the cost of an error. We find that pretrained VLMs compress their memory over text but not video and are poorly calibrated on both. Among a broader set of memory backbones, several non-Transformer architectures achieve better compression-calibration tradeoffs than RoPE Transformers, suggesting they may be useful components for agents operating over long horizons.
Manish Kansana, Mohammed Yusuf Mujawar, Sudip Mittal +2cs.RO cs.CV
Robotic perception from a single viewpoint is often limited by self-occlusion and incomplete surface visibility. This paper presents DARP(Dual-Arm Robotic Perception) https://doi.org/10.21227/rmv3-be47, a calibrated dual-arm RGB-D-IR dataset for object-centered robotic perception using two independently moving eye-in-hand manipulators positioned on opposite sides of a shared tabletop workspace. Each arm carries an Intel RealSense sensor that continuously records RGB, depth, and stereo infrared data while synchronized robot joint states are logged for pose recovery. Objects are placed without fixed poses or marked locations, and the acquisition procedure performs automatic localization, cross-arm confirmation, adaptive viewpoint generation, and continuous multimodal recording. DARP contains ten unique tabletop objects and preserves the original sensor recordings, robot-state logs, object-level metadata, and calibration information required to reconstruct camera trajectories in a shared metric frame. To evaluate the geometric consistency of the acquisition, we implement a deterministic multi-view fusion pipeline that converts calibrated RGB-D observations into complementary partial point clouds and measured surface meshes without using learned or generative completion methods. Evaluation on 224 held-out RGB-D keyframes comprising 1,563,466 three-dimensional query points yields a median point-to-mesh distance of 2.13~mm and an RMSE of 4.04~mm, with 96.56\% of points within 10~mm of the measured-surface mesh. DARP is intended as a reusable resource for multi-view reconstruction, collaborative robotic perception, multimodal fusion, active perception, and future learning-based reasoning over partial object observations.
Conventional algebraic triangulation solves 3D human pose estimation (HPE) from multi-view 2D keypoints. The typical approach, decoding 2D keypoints from predicted heatmaps, is unreliable as heatmaps can be multimodal under occlusion, and collapsing them into single peaks discards their spatial distribution. We seek to use the entire heatmap to estimate 3D poses more accurately, which requires solving two problems: how to robustly fuse heatmaps across views, and how to assess the reliability of heatmaps. For the former, we introduce a novel objective, Multi-viewExpected-OKS Maximization (MEOM), that locates a 3D joint where the views agree in probability mass. For the latter, we adopt highest-density-region (HDR) calibration as a diagnostic of that mass, independently of distance-based metrics. The proposed framework covers two settings, with and without 3D supervision. Without 3D supervision, we optimize 3D poses from pretrained heatmap predictors by maximizing MEOM, achieving comparable performance with state-of-the-art methods that rely on larger backbones, temporal fusion, and simulated 3D data. On ambiguous Human3.6M (H36MA) and occluded CMU Panoptic frames, the advantage is substantial. When 3D labels are available, we train the model end-to-end with a combined MEOM and MSE loss, achieving 19.11 mm absolute MPJPE on Human3.6M outperforming the state-of-the-art volumetric approach on absolute MPJPE at half the inference cost.
Accurate uncertainty estimation is essential for machine learning systems de- ployed in high-stakes domains such as medicine. Traditional approaches primarily rely on probability outputs from trained models (point predictions), which provide no formal guarantees on prediction coverage and often require additional calibra- tion techniques to improve reliability. In contrast, conformal prediction (region prediction) offers a principled alternative by generating prediction sets with finite- sample validity guarantees, ensuring that the ground truth is contained within the set at a specified confidence level. In this study, we explore the impact of pre-training approach, dataset scale and domain on both point and region-level uncertainty quantification, by studying domain-specific vision medical foundation models vs. general domain vision foundation models. We conduct a comprehensive evaluation across foundation models trained on retinal, histopathological, and Chest X-Rays data, applying various calibration techniques. Our results demonstrate that (1) pre-training on higher-quality domain-specific datasets along with self-supervised learning leads to better-calibrated point predictions than general domain pre-training, (2) stan- dard re-calibration methods alone cannot fully mitigate uncertainty discrepancies across models trained on different data sources, (3) domain-specific foundation model can lead to more efficient conformal prediction. These findings highlight the importance of careful model selection and the inte- gration of both point and region prediction to enhance the reliability and trust- worthiness of medical AI systems. Our work underscores the need for a holistic approach to uncertainty quantification in recent development of medical vision foundation model, ensuring robust and interpretable AI-driven decision-making.
Test-time prompt tuning (TPT) has emerged as a powerful paradigm, refining prompts for each test sample via entropy minimization (EM) over multiple augmented views. However, we identify a limitation in the standard EM-based adaptation: it inherently drives the model toward overconfident predictions disregarding sample-specific uncertainty, leading to significant calibration degradation. To address these limitations, we propose a new objective that replaces the conventional EM loss by aligning the original-view prediction with a target distribution derived from augmented views via cross-entropy, while adversarially incorporating the entropy of the target distribution to capture sample-specific uncertainty. Furthermore, to better construct this target distribution, we apply confidence-aware temperature scaling to each augmented-view prediction according to its confidence, sharpening confident predictions while softening uncertain ones. This formulation allows the model to increase confidence only when the target distribution is reliable, while preserving uncertainty when it reflects ambiguous or conflicting augmented-view predictions. Extensive experiments across diverse benchmarks demonstrate that our approach not only achieves state-of-the-art accuracy but also significantly improves model calibration.
Sunwhi Kim, Sunyul Kim, Meounggun Jo +1cs.HC cs.CV
AI image generators now create face portraits that are hard to tell from real photographs. Vision-language models (VLMs) are increasingly proposed to flag such images. We benchmarked 19 VLMs on the same 198 face portraits -- real photographs and identity-matched ChatGPT-4o and Imagen 3 versions -- under the same task as our earlier study of 1,667 adults (85% correct overall; accuracy fell steeply with age). The June-2026 cohort of 14 models only matched adults in their 20s-30s. Four weeks later the ceiling broke. Among five July-2026 releases under the identical protocol, gpt-5.6-sol reached 92.8% balanced accuracy (five-draw mean 92.1%), clearly above adults in their 20s (88.5%), and claude-fable-5 detected every AI image while averaging 91.9%. Model sensitivity now exceeds young adults decisively (d' up to 3.4 versus ~ 2.4). What has not been overtaken is human calibration. Model criteria spread from c = -1.10 to +1.45 while humans sit near zero at every age; both new leaders are biased (+0.44, -0.97), and only a few mid-ranked models approach the human balance. Changing the labelled examples still flipped about one answer in four. The best machines now out-see young adults here, without matching the human balance between suspicion and trust.
Detecting hallucinations in Large Vision-Language Models (LVLMs) requires both accurate span localization and well-calibrated confidence scores. Fine-tuned generative VLMs excel at identifying hallucinated text spans but suffer from overconfidence and high inference latency. Discriminative sequence taggers offer deterministic speed and superior calibration but exhibit conservative span recall. We present SpanCalib-VLM, a hybrid dual-system for the SHROOM-Visions Shared Task that combines a multimodal sequence tagger, consisting of XLM-RoBERTa-Large fused with a SigLIP vision encoder via cross-attention, with our fine-tuned generative VLM (Qwen3.5-4B-SHROOM-SFT). Through a Union-Calibrated Fusion strategy, candidate spans from the generative model are re-scored with calibrated probabilities from the sequence tagger. On the SHROOM-Visions English evaluation split, our ensemble achieves a Pearson calibration correlation of 0.41 and an overall IoU of 0.39, with a clean-response IoU of 0.91} and overall detection accuracy of 70.7%. We make our model weights and code publicly available.
Geometric eye trackers can provide the spatial accuracy required for gaze-based interaction and multimodal studies, but their measurements remain sensitive to residual session-specific calibration error. Research on correcting this error is difficult to compare because methods are typically evaluated with different devices, target layouts, and error definitions. We present a calibration-focused dataset containing 163 trials from 12 participants, with separate 18-point fitting and 32-point test grids, and use it to evaluate global, local, and composite correction functions under a common spatial-extrapolation protocol. We further introduce a lightweight neural refiner that combines ranked predictions from complementary calibrators. On this controlled dataset, post-vendor correction reduces the mean angular error from $1.53^\circ$ to $1.03^\circ$ with the strongest classical composite and to $0.96^\circ$ with the refiner. In a closed-loop gaze task, lower residual error is associated with higher performance across four online correction conditions. These results provide a reproducible data-quality benchmark for using gaze as a behavioral signal in interactive modeling.
Machine learning is usually evaluated after an evidence interface has been fixed. A dataset, sensor suite, query language, action set, or experimental protocol determines which observations can be obtained, and learning is judged by what it extracts from them. We study a complementary capability. A learner's state can determine which evidence-generating experiments it can reliably realise under bounded resources, even when primitive affordances remain fixed. We call this learner-relative experiment family its effective epistemic reach, and use Wide Learning for task-relevant learning-induced changes in that family.We formalise effective reach relative to learner state, deployment budget, reliability threshold, and evaluation distribution. In a controlled construction, two hidden worlds have exactly the same public observation law. An informative diagnostic exists in a fixed five-primitive substrate. Before calibration, one address attempt realises it with probability at most $2^{-10} = 1/1024$, below a pre-specified 0.95 threshold; after calibration, held-out realisation is 1. Public-channel total variation is 0, whereas the realised diagnostic has total variation 1, and sealed binary risk moves from approximately 1/2 to 0. The construction establishes that learning can change effective epistemic reach even when primitive affordances and deployment resources are held fixed. It opens a complementary evaluation question for learning systems: not only what they infer from available evidence, but what informative evidence experience teaches them to bring within reach.
Gaussian process (GP) regression with a single global GP (GP-glo) incurs cubic computational cost, limiting scalability to large datasets. Product-of-experts GP models (GP-pro), which combine local GP models to capture global correlations, alleviate this computational burden. However, training local experts on disjoint data subsets can lead to overestimated posterior variances. We propose GP-pro-c, a product-of-experts GP model that calibrates these variances using an information-based method. The method exploits the monotonicity and submodularity of information gain in GPs to define a calibration ratio that reduces the posterior variance of individual local GP models. We evaluate GP-pro-c using negative log-likelihood (NLL), root mean squared error (RMSE), and expected normalised calibration error (ENCE). Experiments on four synthetic functions and six regression datasets show that GP-pro-c achieves average reductions of 2.3% in NLL and 12.0% in ENCE compared with the uncalibrated GP-pro model. The proposed method mitigates posterior variance overestimation while maintaining predictive accuracy and reducing computational complexity. GP-pro-c provides a promising approach for uncertainty estimation in scalable GP models and may serve as a useful surrogate model for Bayesian optimisation with high-dimensional and large-scale data.
Benjamin Turtel, Paul Wilczewski, Kris Skotheim +2cs.LG cs.AI
This paper evaluates how reward function choice shapes the performance and behavior of LLM forecasters. We compare five proper scoring rules as training objectives for binary forecasts of resolved real-world events. Although the rules share the same theoretical incentive for truthful probability reporting, the resulting models differ in calibration, probability use, and estimated profiles of bias, information, and noise, with smaller differences in aggregate accuracy and discrimination. The Brier-trained model has the lowest observed Brier score and highest AUC-ROC, while the log-trained model has the highest observed log score and lowest calibration error. Models with similar aggregate performance also reach that performance through different combinations of bias, information, and noise. Proper scoring rules therefore need not behave interchangeably as training objectives. Reward choice may shape not only how well an LLM forecasts, but how its forecasting errors are structured. Each condition uses a single seed, so some differences may reflect training stochasticity.
We study risk-averse decision making, in which an agent selects actions while being uncertain about the true system state. The risk is measured via optimized certainty equivalent (OCE) metrics, which generalize popular criteria such as mean-variance risk and conditional value-at-risk (CVaR). We characterize the optimal policy under known distributions, and show that it reduces to a prediction set-based solution for the CVaR. This provides an operational interpretation of conformal prediction-type prediction sets. For unknown distributions, we develop a data-driven calibration strategy, based on a synthetic model for the likelihood and held-out calibration data, yielding high-probability control of the OCE risk. The approach is evaluated on two wireless beamforming settings.
Multi-organ ultrasound classifiers increasingly combine attention, mixture-of-experts routing, uncertainty gating, and evidential deep learning (EDL) objectives to address heterogeneous anatomy and acquisition. Yet a plausible design rationale does not by itself establish that an added component improves the trained system. We contribute a controlled complexity-audit framework, applied to the deployment decision between the maximal evidential candidate Full-EDL and simpler alternatives. Six candidates were evaluated on the primary dataset and three in an internal replication, using ten matched seeds, frozen image-level partitions, capacity- and optimisation-aware comparisons, symmetric temperature scaling, paired decision rules, and a separate out-of-distribution (OOD) veto. Retaining Full-EDL did not establish a reliable macro-F1 gain on either dataset, while the simplified alternatives remained inconclusive under the non-inferiority margin. Simple cross-entropy with temperature scaling (Simple-CE+TS) met the calibrated negative log-likelihood criterion on both datasets and showed favourable selective-risk ordering. The raw calibration advantage of evidential training disappeared after temperature scaling and did not recur on the second dataset. The gate had negligible observable influence at the audited checkpoints, and deleting the Full-only chain revealed no stable task or calibrated-loss benefit. Simple-CE nevertheless triggered the OOD veto against the fetal probe but not the lung probe, precluding an unconditional OOD-safety claim. We therefore selected Simple-CE+TS for the evaluated in-distribution objective while retaining Full-EDL as the maximal reference. Components should earn retention through functional and retraining-based evidence, and calibration and distribution-shift reliability should be evaluated separately.
An LLM agent shown a professional-looking market panel commits to a directional call on a provably unpredictable question far more often than one asked the bare question: across 12 frontier models, commitment rises from 6.5% to 54.0% as evidence is escalated. It commits just as readily when every number on the panel is invented: fabricating the entire display, so nothing the model can see is true except the question itself, still lifts commitment from 24.5% to 36.8%, statistically indistinguishable from the 37.6% produced by genuine market data. What unlocks confident action is not information but the authority of its packaging. The failure is narrow and locatable. Incapacity is not the answer: on matched answerable questions attached to the same panels, the same models answer essentially always, at near-perfect accuracy. Nor is it belief - stated probabilities barely move across the gradient that swings action by 48 points, and score worse than a climatological baseline. Missing judgment isn't it either: asked to classify a question's knowability before acting, models call it irreducible 90% of the time and then commit on just 0.4% of those. The act/don't-act gate is what fails, and the effect is concentrated in a few models rather than universal. Because the gate is separable, it can be trained. Supervised fine-tuning of a 3B model on 540 synthetic cases, predominantly dice, coins, jars and timers, drives commitment to 0.0% on the original cases and transfers to three unseen domains. It does not survive everything: the gate holds exactly when the response format leaves room to reason, and rigid formats that remove that room leave the model confident and wrong on questions it otherwise answers correctly. The gate is trainable and context-fragile, and deployment needs both halves of that sentence.
Reliable evaluation of vision-language models (VLMs) and medical vision-language models (Medical-VLMs) requires calibrated confidence, particularly under realistic clinical conditions. However, existing efforts mainly focused on improving accuracy, leaving calibration in the medical domain underexplored. To this end, we propose MVC-Bench, a calibration-centric benchmark for medical image classification with VLMs and Medical-VLMs. MVC-Bench assesses the calibration across three axes: (i) robustness to modality, backbone, and domain shift (ii) effectiveness of calibration strategies and prompt-tuning methods (iii) stability under prompt-template and random-seed variations. The benchmark covers eight different backbones, three medical modalities, including fundus imaging, histopathology, and chest X-ray under in-domain and domain shift settings. It compares post-hoc calibration, train-time calibration, and zero-shot inference methods, together with six prompt-tuning methods. Across more than 1638 controlled experiments, we report accuracy and Expected Calibration Error (ECE) as primary metrics, and further report results with complementary calibration measures, including Maximum Calibration Error (MCE) and Adaptive Calibration Error (ACE). We further investigate the underlying causes of miscalibration in VLMs and Medical-VLMs and propose a simple train-time calibration method, Multi-Class Margin (MCM) regularization, which achieves lowest ECE on 10 out of 12 settings in in-domain and remains competitive under domain shifts. Collectively, MVC-Bench provides a structured evaluation framework and actionable guidance for improving calibration in safety-critical medical workflows.
Open-ended Theory-of-Mind (ToM) trackers emit valid beliefs absent from finite references. A finite-reference-plus-matcher pipeline marks unmatched outputs false, creating proxy labels that can reverse proper-score model selection on fixed outputs. Holding 259 beliefs and paired scores fixed, reference recoding lowers weighted prevalence from 0.783 to 0.295 and reverses strictly proper Brier risk: a frozen source-prior rule leads native confidence by 0.227 under reference labels and trails by 0.152 under blinded adjudication, in all six authored scenarios. A reference-only Platt recalibrator reverses further. An ICE-specific reversal appears in a released 301-question NQ-open DPR-BERT pipeline: its average-confidence baseline improves instance-level calibration error by 0.045 under exact match but worsens it by 0.074 under human correctness, with both intervals excluding zero. On independently authored OpenToM narratives, 90-96% of audited unmatched beliefs are literally true and the paired direction again reverses. An exact decomposition attributes the distortion to omitted truths, and a closed-form criterion correctly classifies comparisons from twelve released systems. Frozen-audit retrospective replay shows 50 attempted annotations recover ranking direction with probability at least 0.996. TriSource-Restore anchors full-frame reference labels and frozen automatic judgments to a probability-sampled human pilot, maintains at least nominal coverage, narrows intervals, and repairs confidence subject to a base-rate deployment gate.
Probability calibration aligns model confidence with predictive accuracy, enabling clinicians to identify unreliable segmentation regions. This alignment breaks down under domain shift, where artifacts and unseen protocols produce confident errors. Existing post-hoc methods adapt the correction at test time, conditioning on predictive entropy, the logit pattern, or augmentation response, but each proxy is read from the terminal prediction, the very quantity that shift corrupts. This motivates reliability evidence beyond the terminal prediction, which categorical diffusion provides in two ways. First, a generative shape prior keeps a capacity-limited reference intact when appearance is corrupted, so its disagreement with the primary segmentor highlights primary-model errors. Second, every reverse step yields a class distribution, separating persistent disagreement from transient discrepancy. Aggregated over the trajectory, this disagreement correlates with Dice at 0.788, against 0.521 for a matched discriminative control. We therefore propose CARD (Calibration via Agreement in Reverse Diffusion), which maps the temporal aggregate of this disagreement to a temperature field applied per pixel across all classes, so that confidence changes while the segmentation does not. Across cardiac, prostate and brain MRI shifts, CARD lowers calibration error in 45 of 49 comparisons against the strongest baseline in each setting.
Jai Kumar Sharma, Peeyush Tapadiyacs.CV cs.AI q-bio.QM
Frozen hematology foundation-model (FM) embeddings reach near-saturated in-domain white-blood-cell (WBC) accuracy, but clinical deployment demands reliability across scanners, sites, stains and preparation pipelines. We audit 15 frozen encoders (hematology, pathology, and general vision) across four public single-cell acquisition domains along two axes: accuracy robustness and calibration. In-domain linear-probe macro-F1 is saturated (0.98-0.997), yet cross-dataset macro-F1 drops 34-72% and rankings re-order: DinoBloom-L, the in-domain best, falls to 10th of 15 on the most-shifted target (MLL23) at the benchmark's shared 224-px input, behind RedDino and several general and pathology encoders. Rank transfer is probe-dependent: 1-NN retrieval is more stable on average than a source-fitted linear head (median $ρ$ 0.65 vs 0.45), but neither probe universally predicts target robustness. Calibration also collapses: source-trained probes are nearly calibrated in-domain (expected calibration error, ECE, 0.004) but confidently wrong off-domain (ECE 0.35), and source-fitted temperature scaling transfers poorly. We further audit pretraining exposure and identify MLL23 as DinoBloom's internal cohort; because DinoBloom's only held-out dataset is also our source domain, this benchmark cannot isolate exposure from scanner-associated shift. Label-free adaptation and marginal-entropy-based model selection appear safe under balanced evaluation but fail under realistic WBC class-prior shift. Class-Balanced Re-standardization (CBR), a training-free pseudo-label-balanced feature normalization, improves all evaluated target-prior scenario means and partially improves calibration, although encoder-level exceptions and residual miscalibration remain. Hematology FM benchmarks must therefore jointly audit accuracy, calibration, exposure, and class-prior robustness.
Agentic systems increasingly gate actions on a model's own stated confidence, which assumes confidence tracks correctness at the moment of acting. We test this in a hidden-information chess variant where royal status can be secretly, repeatedly relocated between pieces, and where an agent's stated probability distribution over the opponent's hidden royal piece -- elicited every turn, separately from the move it chooses -- is scored against ground truth recoverable after the game. Across two independent batches, captures made at high stated confidence ($\geq 0.5$) about the hidden piece's location were correct in 1 of 62 cases. The calibration deficit is concentrated almost entirely in these events: 99.3% of it in the original batch, 98.7% in the replication. The same pattern, in weaker form, orders consistently (point estimates only; most pairwise gaps are not statistically distinguishable at this sample size) across four further model configurations spanning a second provider -- reported as scope for the finding, not as evidence that capability predicts calibration: a same-model comparison at a fixed external leaderboard score shows a deliberation-budget change alone moves the metric by nearly as much as a large cross-model gap. In a separate seat, conventional evaluation axes -- legality, cost, latency, completion rate -- can dissociate entirely from belief quality, with the configuration winning on every conventional axis producing the worst belief quality tested. A model exhibiting this pattern can still win the game its belief was about, which is why outcome-only evaluation would not detect it.
Visual grounding is typically evaluated as a one-shot mapping from an informative referring expression to a visual target. This formulation misses a central property of real-world reference: target information is often incomplete, ambiguous, and established through interaction. We introduce a controlled evaluation framework for interactive visual grounding in large vision-language models (LVLMs), varying how much target information is provided upfront and how much must be acquired through dialogue. Across four human-grounded visual contexts and four interaction protocols, current LVLMs perform significantly below task-level human baselines. Interaction can help when follow-up questions refine or repair an initial target description. Performance is lowest when no initial description is provided and target information must be acquired through questions, indicating that proactive question-driven grounding remains difficult. LVLMs are also poorly calibrated, often reporting confidence that exceeds their empirical accuracy. Follow-up studies confirm these patterns across varied description sources (human versus AI), reasoning efforts, repeated interactions, description providers, and visual contexts. Overall, interactive visual grounding remains an important challenge, requiring visual matching, information seeking and synthesis.
State space models, especially Visual State Space Duality (VSSD), have emerged as efficient linear-time alternatives to Transformers for dense visual tasks. However, we observe that VSSD compresses spatial context into a global aggregation that suppresses high-frequency responses, causing excessive boundary smoothing in remote sensing semantic segmentation. To address this, we propose CRISP, a calibration framework with two components. Its core, the Duality Calibration Operator (DCO), restores local contrast and boundary responses through residual injection and frequency calibration within the VSSD backbone, without altering its linear complexity. To retain the recovered detail, an Orthogonal Multi-Prototype (OMP) head assigns multiple orthogonally constrained prototypes per class to model large intra-class variance. Extensive experiments on Potsdam, Vaihingen, and LoveDA show that, with approximately 30M parameters, CRISP achieves consistent gains in mean F1 (mF) and mIoU while remaining competitive with state-of-the-art methods. Code is available at https://github.com/crazylifeha/CRISP.
Split conformal prediction, not the pruning rule, supplies finite-sample marginal coverage once a pruned model is fixed independently of the conformal calibration split. We study the separate efficiency problem: can pruning preserve score geometry well enough to obtain smaller valid prediction sets? Calibration-Preserving Pruning (CPP) augments a base pruning score with nonconformity-gradient saliency and uses disjoint pruning, validation-selection, conformal-calibration, and test splits. Bounded score perturbations imply bounded conformal-quantile shifts and controlled set inflation, but do not make the generic coverage theorem CPP-specific. Final five-seed Qwen2.5-1.5B results at 50\% sparsity show the largest gains on large-label tasks. On DBpedia-14, CPP-SparseGPT reduces mean set size from \(10.1\) to \(8.6\) while changing accuracy from \(0.347\) to \(0.366\); CPP-Wanda reduces \(11.2\) to \(9.0\) with an accuracy trade-off from \(0.310\) to \(0.295\). Across 15 dataset--sparsity cells, CPP-SparseGPT produces smaller sets in 13 and higher accuracy in 11. Matched controls show that generic supervised gradients explain much of the gain: true-label CPP is not statistically resolved from matched Wanda+SNIP, whereas threshold-aware candidate-label CPP reaches \(7.8\) mean set size at explicit accuracy and offline-compute costs. RoBERTa-base and Llama-3-8B diagnostics support transfer, but our claims remain limited to reliability-sensitive classification.
Aligning deployed language models requires knowing when their outputs can be trusted, yet on-device models now ship to hundreds of millions of devices with no server-side moderation, and the configuration developers can actually deploy is rarely audited independently. We present a reproducible reliability audit of the developer-accessible on-device foundation model, framed as an oversight question: can a user or a resource-constrained developer tell when the model is wrong? Red-teaming it on calibration, confident confabulation on false-premise questions, and over-refusal of benign prompts, we find a \emph{task-asymmetric miscalibration}: its guardrails fail in opposite directions across tasks (confabulating on 69\% of false premises while refusing 18\% of entirely benign inputs), atop a self-reported confidence that is saturated and non-discriminative (AUROC 0.47; ECE 70, worst among comparable small models). Crucially, confident-correct and confident-wrong outputs are \emph{surface-indistinguishable}: a classifier over 15 user-visible features separates them at AUROC only 0.55 (equivalence-confirmed), leaving no signal for oversight at inference time. No cheap single-generation signal flags these failures ($\le$0.68 AUROC), whereas a black-box consistency wrapper requiring no model access recovers reliability (confident confabulation 75\%$\to$3\%; selective accuracy 43\%$\to$83\%) at a tunable cost. We contribute a model-agnostic audit protocol, a surface-indistinguishability test, and released code and frozen evaluation items as reusable infrastructure for auditing deployed models.
Amit Kumar, Elnur Adl Zarabi, Suranjana Trivedy +4cs.LG cs.AI stat.ME
Large language models (LLMs) are increasingly used to provide prior causal knowledge for structural causal discovery, yet whether their direct-edge judgments and confidence can be trusted remains unclear. We systematically evaluate 12 instruction-tuned open-weight models across six benchmark causal graphs, five prompting strategies, and four confidence sources: verbalized, logit-based, cross-prompt agreement, and cross-model agreement. Under our language-only pairwise protocol, our evaluation yields three key findings. (i) LLM-based causal judgments are strongly recall-dominant: models predict overly dense graphs with many false-positive edges, while prompting mainly shifts the precision-recall trade-off rather than resolving overprediction. Gains from model scale diminish on the largest graphs and do not eliminate miscalibration. (ii) LLMs often capture causal relatedness without reliably identifying directness or orientation. Relative to published reference graphs, models misclassify 40.0% of indirect and 36.0% of reversed non-edges as direct edges, versus 28.2% of other non-edges. Moreover, 80.8% and 84.6% of these false positives receive verbalized confidence of at least 80%, revealing substantial overconfidence in structurally incorrect predictions. (iii) Conventional confidence estimates are unreliable, whereas agreement offers a more promising signal. Logit-based confidence frequently collapses near 1.0 regardless of correctness, while cross-prompt and cross-model agreement achieve better mean calibration and discrimination, though their advantages are not statistically significant after Holm correction. A benchmark-familiarity audit further identifies potential familiarity in five model-dataset pairs, all involving AsiaM. Overall, our results suggest LLMs are better viewed as sources of externally validated soft causal priors than as direct evidence of causal structure.
Large language models (LLMs) often produce fluent but incorrect answers with unwarranted confidence. A central limitation is that standard LLMs represent uncertainty through a single predictive distribution, conflating epistemic ignorance with genuine ambiguity. We introduce Credal Large Language Models (CLLMs): an ensemble of LoRA adapters induces a credal set whose lower and upper probabilities expose the spread of plausible predictive distributions rather than collapsing to a single softmax output. From this representation we derive two complementary commitment scores. Credal Token Commitment (CTC) is a token-space score that combines lower-bound support, credal width, and intersection entropy, computed without additional generation. Semantic Commitment Consistency (SCC) extends commitment to semantic space using sampled completions, with SCC-Gap measuring the mismatch between token-level and semantic-level support. We evaluate hallucination detection, calibration, selective prediction, and reasoning on Gemma-2-9B, Llama-3.1-8B, and Qwen2.5-7B across OpenBookQA, CoQA, TriviaQA, and ARC-Challenge. CLLM is the best method on QA accuracy at competitive expected calibration error, and CTC tracks the best hallucination AUROC within 1.5 pp on most settings without additional generation. On selective prediction at 80% coverage, CLLM with SCC reaches 99.0% accuracy on OpenBookQA, and on ARC-Challenge CLLM with Csem confidence achieves <= 0.6% ECE across the three backbones.
Marios Papamichalis, Regina Ruane, Theofanis Papamichalisstat.ML cs.LG
Two analysts who calibrate the same predictive model on independent samples will deploy different prediction sets every time, because the calibration threshold inherits the randomness of the data. Wherever deployments must be audited, cached, or approved across sites, this instability is costly: no one can verify that two calibrations produced the same object. We ask two questions: when can independent calibrations yield the identical classifier, and what must that agreement cost? Perfect agreement is impossible, since a procedure that almost always returns one fixed answer cannot remain valid for every distribution, and exact agreement through shared randomness forces the procedure to ignore its data. Sharing a single random seed and rounding the calibrated threshold up to a coarse shared grid resolves the tension: the deployed classifier becomes identical across analysts with any desired probability, coverage guarantees survive, and the price is a quantified increase in set size and calibration data. Matching lower bounds show that no threshold method can pay less, and the method's one tuning constant vanishes asymptotically. Without any shared seed, a fixed grid still confines all analysts to two adjacent classifiers, and no method does better. Replicability also blocks gaming: selecting the most favorable of many recalibrations barely moves a replicable classifier, while the same selection silently undercovers standard conformal prediction. Experiments on real ImageNet outputs, a four-hospital site split, and four language-model families match the theory, including the measured sample-cost frontier.