# Uncertainty quantification LLMs

Methods for estimating and communicating the reliability of large language model outputs.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Realized versus Population Reliability: Anytime-Valid Self-Consistency Certification under Correlated Reasoning

DevFeed: [Realized versus Population Reliability: Anytime-Valid Self-Consistency Certification under Correlated Reasoning](<https://devfeed.tech/articles/realized-versus-population-reliability-anytime-valid-self-consistency-certification-under-correlated-reasoning-145197.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26592>)

Author: Terence D. Agbeyegbe, Toneukarin G. Agbeyegbe

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [aime](<https://devfeed.tech/tags/aime.md>), [betting](<https://devfeed.tech/tags/betting.md>), [certificate](<https://devfeed.tech/tags/certificate.md>), [certification](<https://devfeed.tech/tags/certification.md>), [claude-haiku-4-5](<https://devfeed.tech/tags/claude-haiku-4-5.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [correlation](<https://devfeed.tech/tags/correlation.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [error-control](<https://devfeed.tech/tags/error-control.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [simulations](<https://devfeed.tech/tags/simulations.md>), [stat-me](<https://devfeed.tech/tags/stat-me.md>)

### AI overview

Correlated language-model rollouts change what an anytime-valid self-consistency certificate establishes. Under a per-prompt latent-state model, the standard betting e-process controls error for the realized belief, but not necessarily for average population behavior; the two null hypotheses coincide only at zero correlation. Within-prompt rollout averages have effective sample size capped at 1/ρ, so adding rollouts cannot repair population validity. A prompt-level e-process restores it, and under a pure rollout budget the reported optimal allocation is one rollout per prompt. In Claude Haiku 4.5 experiments, the gap was dormant on GSM8K (estimated ρ=0.435, θ=+0.96), but on AIME (ρ=0.480, θ=−0.15) the single-unit rule crossed on 24% of prompts at nominal ε=0.05, while the prompt-level procedure declined (0.00).

### Source excerpt

arXiv:2611.26592v1 Announce Type: cross Abstract: Self-consistency certification treats repeated sampling from a language model as sequential evidence accumulation and stops when an anytime-valid e-process crosses a threshold, guaranteeing finite-sample Type I error control by Ville's inequality. Prior constructions assume sampled rollouts are conditionally independent. We show this assumption fails in practice: reasoning streams are positively correlated through shared reasoning prefixes or across-prompt heterogeneity, changing not merely certificate efficiency but the hypothesis certified. Modeling dependence via a per-unit latent state (the random-effects / intraclass-correlation structure) induces exchangeable, equi-correlated increments with correlation $\rho$. Correlation splits the single null into a realized null about the drawn belief and a population null about average model behavior, which coincide only when $\rho=0$. We show: (i) the standard betting e-process remains anytime-valid for the realized null, not the population null; (ii) read as a population-null test, a suitable mixture has asymptotic crossing probability at least $\pi=P(\mu(Z)>0)$, potentially exceeding nominal $\epsilon$; (iii) the within-unit mean's effective sample size saturates at $1/\rho$, so extra within-unit rollouts cannot restore population validity; and (iv) a prompt-level e-process restores population validity, with optimal allocation at one rollout per prompt under a pure rollout budget. Simulations and Claude Haiku 4.5 rollouts confirm this regime dependence: on GSM8K ($\hat{\rho}=0.435$, $\hat{\theta}=+0.96$) the gap is dormant, whereas on AIME ($\hat{\rho}=0.480$, $\hat{\theta}=-0.15$) the single-unit rule crosses on 24% of prompts at nominal $\epsilon=0.05$, while the prompt-level procedure declines (0.00). This correctness result identifies which reliability claim an anytime-valid certificate actually supports under dependence.

## LLMs Are Already Confidence-Aware Zero-Shot Classifiers

DevFeed: [LLMs Are Already Confidence-Aware Zero-Shot Classifiers](<https://devfeed.tech/articles/llms-are-already-confidence-aware-zero-shot-classifiers-145035.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25289>)

Author: Owen Cook, Xingyi Song

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [Calibrated uncertainty](<https://devfeed.tech/topics/calibrated-uncertainty.md>)

Tags: [black-box](<https://devfeed.tech/tags/black-box.md>), [brier-score](<https://devfeed.tech/tags/brier-score.md>), [classification](<https://devfeed.tech/tags/classification.md>), [classifier](<https://devfeed.tech/tags/classifier.md>), [classifiers](<https://devfeed.tech/tags/classifiers.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [decision](<https://devfeed.tech/tags/decision.md>), [decode](<https://devfeed.tech/tags/decode.md>), [discrimination](<https://devfeed.tech/tags/discrimination.md>), [f1](<https://devfeed.tech/tags/f1.md>), [gemma](<https://devfeed.tech/tags/gemma.md>), [gemma-4](<https://devfeed.tech/tags/gemma-4.md>), [gemma-4-31b](<https://devfeed.tech/tags/gemma-4-31b.md>), [generation](<https://devfeed.tech/tags/generation.md>), [jev](<https://devfeed.tech/tags/jev.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [llms](<https://devfeed.tech/tags/llms.md>), [zero-shot](<https://devfeed.tech/tags/zero-shot.md>)

### AI overview

Across eleven open-weight LLMs and five classification datasets, confidence method choice rarely changed accuracy significantly, but token-probability logit scoring discriminated correct from incorrect predictions better than verbalised confidence or sampling (AUROC 0.74-0.77 vs. 0.57-0.66). Replacing labels with single-token identifiers preserved discrimination but reduced accuracy by 3-6 points on average. Decision-only scoring, which scores only positions where labels diverge in a prefix trie, cut decode steps by 88-96% on the three largest label sets while performing comparably to full-label scoring. The best tested open-weight configurations, Gemma 4 31B and Qwen 3.5 9B, were reported as competitive with Jev 1.13.0. These findings are from the supplied abstract.

### Source excerpt

arXiv:2611.25289v1 Announce Type: cross Abstract: Zero-shot large language models (LLMs) are increasingly used as text classifiers, but their labels are only useful when we know to what extent we can trust them. We compare black-box methods (verbalised confidence and sampling) with white-box methods that read token probabilities (P(True), and multiple logit-scoring methods using constrained generation). We evaluate eleven open-weight LLMs (1B-32B parameters) on five classification datasets with 3 to 151 labels, measuring classification performance (accuracy, F1-macro), discrimination (AUROC), and calibration (ECE and Brier score). The choice of confidence method rarely impacts accuracy significantly, but logit-based scoring methods lead to much better discrimination (AUROC 0.74-0.77) than black-box methods (AUROC 0.57-0.66). Replacing labels with single-token identifiers preserves discrimination but costs 3-6 accuracy points on average. We also introduce decision-only scoring, which only scores positions where labels diverge in the prefix tree (trie). It reduces decode steps by 88-96% on the three largest label sets and performs comparably to full-label scoring. The best open-weight model-method configurations (Gemma 4 31B and Qwen 3.5 9B) are competitive with the commercial Jev 1.13.0 classifier.

## Cross-Question Reliable Reinforcement Learning

DevFeed: [Cross-Question Reliable Reinforcement Learning](<https://devfeed.tech/articles/cross-question-reliable-reinforcement-learning-145036.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.25292>)

Author: Hector G. Rodriguez, Marcus Rohrbach

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [post-training](<https://devfeed.tech/topics/post-training.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [error](<https://devfeed.tech/tags/error.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [learning](<https://devfeed.tech/tags/learning.md>), [math](<https://devfeed.tech/tags/math.md>), [math-500](<https://devfeed.tech/tags/math-500.md>), [qa](<https://devfeed.tech/tags/qa.md>), [questions](<https://devfeed.tech/tags/questions.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [reward](<https://devfeed.tech/tags/reward.md>), [risk](<https://devfeed.tech/tags/risk.md>), [solved](<https://devfeed.tech/tags/solved.md>), [utility](<https://devfeed.tech/tags/utility.md>), [wikipedia](<https://devfeed.tech/tags/wikipedia.md>)

### AI overview

CARE (Cross-question Abstention Reward Estimation) trains LLMs to rank confidence across a set of question-answer pairs. For several risk tolerances, the set's confidences determine an abstention threshold, and rewards favor confidence scores on the correct side of that threshold. This contrasts with pointwise rewards that push each answer toward confidence 1 if correct and 0 if incorrect, which the authors say loses useful distinctions for selective answering. In math benchmarks MATH-500, GSM8K, and Big-Math-Digits, CARE raises coverage at 5% risk (C@5) from 0.8% to 42.4% versus existing pointwise rewards. Training on HotpotQA also raises C@5 from 12.3% to 23.6% on out-of-distribution general-knowledge, reasoning, and math benchmarks, including CommonsenseQA, GPQA, SimpleQA, and TriviaQA. These findings are reported in the supplied abstract.

### Source excerpt

arXiv:2611.25292v1 Announce Type: new Abstract: AI systems operating near the limits of their capabilities can maximize utility while minimizing the risk of error by abstaining or deferring questions they are not confident about to a human or a stronger model. Recently, large language models (LLMs) have been post-trained to estimate confidence in their own answers. To achieve this, methods reinforce confidence scores based on the correctness signal to each question-answer pair individually, with rewards that encourage assigning confidence 1 to correct answers and 0 to incorrect answers. However, estimating confidence scores is useful in that it enables selecting only the most confident answers, to maximize the share of solved questions while accommodating the user's risk tolerance. Assigning confidence 1 to correct answers and 0 to incorrect ones is theoretically optimal, but unattainable in practice and detrimental for selective prediction with a non-zero risk tolerance: confidences pushed to 0 and 1 carry no information about how answers compare to one another. We therefore introduce CARE (Cross-question Abstention Reward Estimation), which computes rewards over a set of question-answer pairs: At each of several risk tolerances, the confidences across the set induce an abstention threshold, and each confidence is rewarded for lying on the correct side of it. Since the threshold itself comes from the set, every verbalized confidence is reinforced relative to other questions' confidence, not against its own binary correctness label. On math, CARE increases the coverage at 5% risk (C@5) across MATH-500, GSM8K, and Big-Math-Digits from 0.8% to 42.4%, compared to existing pointwise correctness-matching rewards. When trained on multi-hop Wikipedia QA from HotpotQA, CARE improves C@5 in out-of-distribution general-knowledge, reasoning, and math benchmarks (including CommonsenseQA, GPQA, SimpleQA, and TriviaQA) from 12.3% to 23.6%.

## Tell Me How Sure You Are: Communicating Uncertainty in Auxiliary Evidence to Foundation Models

DevFeed: [Tell Me How Sure You Are: Communicating Uncertainty in Auxiliary Evidence to Foundation Models](<https://devfeed.tech/articles/tell-me-how-sure-you-are-communicating-uncertainty-in-auxiliary-evidence-to-foundation-models-143512.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24351>)

Author: Soumyajit Chatterjee, Baturalp Buyukates, Abhirup Ghosh

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [AI Foundation Models](<https://devfeed.tech/topics/ai-foundation-models.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [AI Models](<https://devfeed.tech/topics/ai-models.md>)

Tags: [communication](<https://devfeed.tech/tags/communication.md>), [context](<https://devfeed.tech/tags/context.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [decision-making](<https://devfeed.tech/tags/decision-making.md>), [f1](<https://devfeed.tech/tags/f1.md>), [foundation-models](<https://devfeed.tech/tags/foundation-models.md>), [generative](<https://devfeed.tech/tags/generative.md>), [humans](<https://devfeed.tech/tags/humans.md>), [image-classification](<https://devfeed.tech/tags/image-classification.md>), [improvements](<https://devfeed.tech/tags/improvements.md>), [inference](<https://devfeed.tech/tags/inference.md>), [llm](<https://devfeed.tech/tags/llm.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>)

### AI overview

The study finds that including confidence in auxiliary evidence can improve zero-shot foundation-model image classification. Across five downstream models and three datasets, communicating uncertainty from human-observed or learned-model-predicted text characteristics improved LLM macro-F1 by as much as 26.18 percentage points compared with omitting uncertainty. The authors also report that uncertainty carries predictive information about whether downstream decisions are correct. These findings are limited to the evaluated image-classification setup.

### Source excerpt

arXiv:2611.24351v1 Announce Type: new Abstract: In many queries, the context information provided to generative foundation models, such as LLMs and VLMs, is inherently uncertain, including its representations of the physical world. While conventional prompting discards the associated uncertainty, this paper, for the first time in the literature, systematically demonstrates that communicating such uncertainty in context provides useful additional cues for downstream decisions and yields improvements across a wide range of settings. We formulate the decision-making problem as a zero-shot foundation model-assisted image classification task supported by auxiliary text-based characteristics. We choose this due to the availability of well-defined auxiliary properties, associated uncertainty, and unambiguous task correctness. Using five downstream foundation models across three datasets, we extensively evaluate uncertainty arising from both humans and learned models when observing or predicting auxiliary evidence. Across all settings, we find that communicating uncertainty substantially improves downstream LLM classification performance: macro-F1 improves up to $26.18$ percentage points compared to when the uncertainty information is not communicated. We further observe how the foundation models process such uncertainty information and find that auxiliary uncertainty contains predictive information about the correctness of downstream decisions. Together, these results establish uncertainty as useful contextual information and motivate more systematic consideration of uncertainty communication in future foundation-model inference pipelines.

## Assessing Vision-Language Models for Early Prediction of Unsafe Situations in Robotics

DevFeed: [Assessing Vision-Language Models for Early Prediction of Unsafe Situations in Robotics](<https://devfeed.tech/articles/assessing-vision-language-models-for-early-prediction-of-unsafe-situations-in-robotics-142205.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24146>)

Author: Muhammad Farooq, Aitor Arrieta, Paolo Arcaini, Shaukat Ali, Thomas Peyrucain

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [Monitoring & Alerting](<https://devfeed.tech/topics/monitoring-alerting.md>)

Tags: [classification](<https://devfeed.tech/tags/classification.md>), [cs-ro](<https://devfeed.tech/tags/cs-ro.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [early-warning](<https://devfeed.tech/tags/early-warning.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [history](<https://devfeed.tech/tags/history.md>), [llava](<https://devfeed.tech/tags/llava.md>), [monitoring](<https://devfeed.tech/tags/monitoring.md>), [performance](<https://devfeed.tech/tags/performance.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [qwen](<https://devfeed.tech/tags/qwen.md>), [robot-safety](<https://devfeed.tech/tags/robot-safety.md>), [robotics](<https://devfeed.tech/tags/robotics.md>), [support](<https://devfeed.tech/tags/support.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [video](<https://devfeed.tech/tags/video.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>), [vision-language-models-vlms](<https://devfeed.tech/tags/vision-language-models-vlms.md>)

### AI overview

The abstract reports an empirical evaluation of Qwen, InternVL, and LLaVA-OneVision for early robot-safety monitoring across HuRoN, Pal Robotics, and CrowdBot, using video and odometry to classify situations as safe, potentially unsafe, or unsafe and assess uncertainty. Qwen had the strongest overall classification and prediction-reliability performance; InternVL showed stronger early-warning capability in several configurations. The effect of robot-state information and temporal history varied by model and evaluation objective, so the preferred configuration depends on whether accurate safety-state classification, reliable uncertainty estimates, or timely warning is the priority.

### Source excerpt

arXiv:2611.24146v1 Announce Type: cross Abstract: Vision-Language Models (VLMs) process visual information and textual instructions and, by using their ability to understand and reason about context, are well-suited for predicting robot failures before they occur. However, evidence supporting this remains limited. To address this, we present an empirical evaluation of VLMs for predicting robot failures. In particular, we experiment with Qwen, InternVL, and LLaVA-OneVision across three robotic case studies (HuRoN, Pal Robotics, and CrowdBot), analyzing both video and odometry data for early failure warning (i.e., before an unsafe situation occurs). Specifically, we evaluate their performance in classifying situations as safe, potentially unsafe, and unsafe, together with uncertainty quantified using various metrics, and study the relationship between uncertainty and performance. Our results showed that Qwen provides the strongest overall classification and prediction-reliability performance, while InternVL shows stronger early-warning capability in several configurations. The results further show that the effect of robot-state information and temporal history depends on the model and evaluation objective. Overall, our findings indicate that general-purpose VLMs can support early robot-safety monitoring, but the most suitable configuration depends on whether the priority is accurate assessment of safety-state, reliable confidence and uncertainty, or timely warning before unsafe events.

## The Limit Is in the Input, Not the Model: State-Dependent Command Safety in Natural-Language Interfaces to Industrial Controllers

DevFeed: [The Limit Is in the Input, Not the Model: State-Dependent Command Safety in Natural-Language Interfaces to Industrial Controllers](<https://devfeed.tech/articles/the-limit-is-in-the-input-not-the-model-state-dependent-command-safety-in-natural-language-interfaces-to-industrial-controllers-142217.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.24682>)

Author: Hemen Taukoor, Zhaoxia Wang, Venky Shankararaman

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Audit-then-score](<https://devfeed.tech/topics/audit-then-score.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [actions](<https://devfeed.tech/tags/actions.md>), [all](<https://devfeed.tech/tags/all.md>), [allowed](<https://devfeed.tech/tags/allowed.md>), [applying](<https://devfeed.tech/tags/applying.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [audit](<https://devfeed.tech/tags/audit.md>), [can](<https://devfeed.tech/tags/can.md>), [cases](<https://devfeed.tech/tags/cases.md>), [command](<https://devfeed.tech/tags/command.md>), [controllers](<https://devfeed.tech/tags/controllers.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [industrial](<https://devfeed.tech/tags/industrial.md>), [input](<https://devfeed.tech/tags/input.md>), [interfaces](<https://devfeed.tech/tags/interfaces.md>), [language](<https://devfeed.tech/tags/language.md>), [limit](<https://devfeed.tech/tags/limit.md>), [model](<https://devfeed.tech/tags/model.md>), [natural-language](<https://devfeed.tech/tags/natural-language.md>), [safety](<https://devfeed.tech/tags/safety.md>), [state](<https://devfeed.tech/tags/state.md>)

### AI overview

The proposed answerability audit tests whether a model receives the machine-state information needed to judge a command: it records model inputs and pairs the same command across two states requiring opposite decisions, counting a pair only if both answers are correct. In an author-built test of 15 pairs from three systems and four language models, none reliably distinguished the cases without state, even when given rules. Adding state improved three models; scores ranged from 0 to 13 correct pairs. A rule evaluator matched all pairs by using the same rules that defined expected answers, so this was not an independent safety test. The small-sample results do not establish general reliability; the audit is presented as a way to distinguish missing-input problems from model errors.

### Source excerpt

arXiv:2611.24682v1 Announce Type: cross Abstract: Language models judging commands to industrial controllers may not receive machine state, the equipment's current condition. The same command can require opposite safety decisions in different states, so an incorrect answer may reflect missing information rather than poor reasoning. We propose an answerability audit: does the model receive the information needed to decide? The audit records the model's inputs and pairs the same command with two states requiring opposite answers. The paired score counts a pair only when both answers are correct. We add rules, then state values, to the model's input. In an author-built test of 15 pairs from three systems with four language models, none reliably distinguished the paired cases without state, even with rules. With state, scores ranged from 0 to 13 correct pairs out of 15; three models improved. With text alone, one model always allowed commands and another always denied them. Their overall (raw) accuracies differed by 32 percentage points; both scored 50% balanced accuracy, which weights safe and unsafe cases equally, and no correct pairs. A rule evaluator matched all 15 pairs by applying the same rules that supplied the expected answers; this is not an independent safety test. The audit complements systems that check actions against safety rules: it checks whether the model receives the information needed for those decisions. This helps evaluators separate missing-input problems from model errors. These small-sample scores do not establish general model reliability.

## Operational Proto-Introspection in Looped Language Models: Reading Correctness Before the Answer

DevFeed: [Operational Proto-Introspection in Looped Language Models: Reading Correctness Before the Answer](<https://devfeed.tech/articles/operational-proto-introspection-in-looped-language-models-reading-correctness-before-the-answer-142543.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2607.18553>)

Author: Jan Kirin

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [model architecture](<https://devfeed.tech/topics/model-architecture.md>)

Tags: [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [generate](<https://devfeed.tech/tags/generate.md>), [hidden-states](<https://devfeed.tech/tags/hidden-states.md>), [introspection](<https://devfeed.tech/tags/introspection.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [layers](<https://devfeed.tech/tags/layers.md>), [logic](<https://devfeed.tech/tags/logic.md>), [loops](<https://devfeed.tech/tags/loops.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [signal](<https://devfeed.tech/tags/signal.md>)

### AI overview

In Ouro-RLTT, probes of hidden states from the first 48 reasoning tokens predicted correctness on a synthetic logic task (AUROC 0.758 versus 0.515 for token confidence), including ranking the correct answer higher in 74% of same-question pairs. However, question features plus token confidence performed about as well on a larger set, and the within-question result was unresolved on GSM8K and a second checkpoint. For finished candidate answers, the signal strengthened across loops and was largely shared between them. It improved abstention and answer selection over chance, but surface features performed about as well; once sample agreement was known, the probe added no measurable value. Attempts to use the signal to improve generation did not work. The findings concern what probes can read; they do not show the model uses this information itself.

### Source excerpt

arXiv:2607.18553v4 Announce Type: replace-cross Abstract: Can one tell, while a language model reasons, whether its answer will be right? Looped models are a good place to ask: they run the same layers several times per token, leaving one hidden state per loop for a probe to read. I probe Ouro-RLTT, a 2.6B-parameter looped model, for what these states reveal about how its reasoning will end. I call such information operational proto-introspection: operational because it is defined by what a probe can read, and proto because nothing here shows that the model itself uses it. In part, yes. Reading the first 48 tokens of reasoning, from answers that state no answer yet, a probe on the hidden states reaches an AUROC of 0.758 on a synthetic logic task, against 0.515 for token confidence. Question difficulty is a rival explanation: on 1,164 further questions, token confidence plus simple features of the question predict about as well. It is not only difficulty: between two answers to the same question, the probe scores the correct one higher 74% of the time. A fine-tuned language model reading the same tokens does about as well. On GSM8K and on a second checkpoint this within-question test is unresolved. A probe that ranks finished candidate answers is weak at layer 8 on the first loop, much better on the second, and as accurate as the last layers by the third, on four Ouro checkpoints. The direction it reads is largely shared across loops: a probe trained on the last loop reads the first nearly as well as a probe trained there. The signal helps decide what to do with finished outputs. It improves abstention over surface features and picks correct answers among candidates better than chance, though surface features pick about as well. Once several samples are finished, a baseline that knows their agreement leaves the answer-free probe nothing measurable to add. None of my attempts to use it to make the model generate better answers worked.

## Reliability is a user-specific property of large language models

DevFeed: [Reliability is a user-specific property of large language models](<https://devfeed.tech/articles/reliability-is-a-user-specific-property-of-large-language-models-142194.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23732>)

Author: Xingyu Zhao

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [LLM security](<https://devfeed.tech/topics/llm-security.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [capabilities](<https://devfeed.tech/tags/capabilities.md>), [capability](<https://devfeed.tech/tags/capability.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [customer-support](<https://devfeed.tech/tags/customer-support.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [failure](<https://devfeed.tech/tags/failure.md>), [inference](<https://devfeed.tech/tags/inference.md>), [interaction](<https://devfeed.tech/tags/interaction.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [llm](<https://devfeed.tech/tags/llm.md>), [modelling](<https://devfeed.tech/tags/modelling.md>), [operational](<https://devfeed.tech/tags/operational.md>), [probabilistic](<https://devfeed.tech/tags/probabilistic.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [requirements](<https://devfeed.tech/tags/requirements.md>), [research](<https://devfeed.tech/tags/research.md>), [roadmap](<https://devfeed.tech/tags/roadmap.md>)

### AI overview

The abstract argues that LLM reliability should be claimed for specified users and operating conditions, because capability rankings alone do not establish how often a service will fail in future use. It distinguishes fault-finding tests from reliability assessment and proposes six research workstreams: specifying claims, modelling changing operational profiles, handling graded outcomes and noisy judgements, measuring local probabilistic robustness, accounting for interaction dependence, and applying statistical inference with reassessment. Published findings and a customer-support example motivate the roadmap, but integrated validation remains open. Reliability claims are presented as complements to separate safety and security arguments.

### Source excerpt

arXiv:2611.23732v1 Announce Type: new Abstract: Public benchmarks support comparisons of large language model (LLM) capabilities, yet users can experience the same highly ranked model differently. Drawing on traditional software reliability, we argue that the reliability of an LLM service should be assessed for specified users and conditions of use. This requires more than a capability score: evaluation evidence must support a claim about failure during future operation. We distinguish testing to discover faults from testing to assess reliability and identify ten modelling considerations behind such claims. We propose a six-workstream research roadmap linking claim specification, dynamic operational profiles, graded outcomes and noisy judgements, local probabilistic robustness, interaction dependence, and statistical inference with reassessment. Local probabilistic robustness addresses the behavioural variation hidden when different concrete inputs are treated as the same task. Published findings and an illustrative customer-support example motivate the roadmap; its integrated validation remains an open research task. Reliability claims complement, but do not replace, separate safety and security arguments. Capability scores should therefore be complemented by versioned operational claims that state whose requirements are covered, under which conditions, and with what uncertainty.

## Diagnostic Accuracy and Navigation Behavior of Vision-Language Model Agents Reading Emergency CT and MRI Examinations

DevFeed: [Diagnostic Accuracy and Navigation Behavior of Vision-Language Model Agents Reading Emergency CT and MRI Examinations](<https://devfeed.tech/articles/diagnostic-accuracy-and-navigation-behavior-of-vision-language-model-agents-reading-emergency-ct-and-mri-examinations-142662.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23480>)

Author: Jonathan Lichtle, Aymen Meddeb, Antonin Birot, Ibrahima Niang, Lisa Gastaud, Riyad Hanafi, Sofiane Dujardin, Corentin Selomme, Guillaume Cachard, Aghiles Hamroun, Gr\\'egory Kuchcinski, Su Hwan Kim, Ba

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [behavior](<https://devfeed.tech/tags/behavior.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [diagnostic](<https://devfeed.tech/tags/diagnostic.md>), [emergency](<https://devfeed.tech/tags/emergency.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [navigation](<https://devfeed.tech/tags/navigation.md>), [reading](<https://devfeed.tech/tags/reading.md>), [vision-language-model](<https://devfeed.tech/tags/vision-language-model.md>)

### AI overview

In this retrospective evaluation of 100 emergency CT and MRI examinations, six vision-language model agents were more accurate when given all slices or selected key slices in one request (42.2% and 42.5%) than when navigating freely (32.6%) or starting from a key slice (37.9%). On 30 cases also read by radiologists, freely navigating models were correct in 30.0% of runs, compared with 77.0% for radiologists. Models formed an initial hypothesis early and usually retained it: final answers were correct in 12.4% of runs after a wrong first hypothesis, versus 90.4% after a right one. Median stated confidence was 90%, while 38.9% of final answers were correct. The abstract concludes that accuracy remained substantially below radiologists across the tested strategies.

### Source excerpt

arXiv:2611.23480v1 Announce Type: new Abstract: Purpose. To evaluate the diagnostic accuracy and image navigation behavior of vision-language model (VLM) agents under varying exploration strategies. Materials and Methods. In this retrospective study, six VLMs read 100 emergency CT and MRI examinations published in a public archive (March 2025-August 2026) under four conditions (free navigation, navigation from a key slice, and all slices or key slices in a single request), twice each (4800 runs). Nine radiologists read 30 of those cases. A large language model judge validated against two radiologists graded the answers. Conditions were compared with generalized estimating equations clustered by case. Results. Pooled accuracy was lowest under free navigation (32.6%; 391/1200), followed by navigation from a key slice (37.9%; 455/1200; P = .003), and highest with all slices (42.2%; 506/1200) or key slices (42.5%; 510/1200) in one request (both P < .001). On the 30 cases read by radiologists, freely navigating models were correct in 30.0% of runs (108/360) versus 77.0% for radiologists (208/270; P < .001). Models formed a first hypothesis after a median of 10.8% of the slices and kept it in 63.1% of runs (755/1196); final answers were correct in 12.4% of runs after a wrong first hypothesis (110/885) versus 90.4% after a right one (281/311; P < .001). Median stated confidence in the final answers was 90%, whereas 38.9% of them were correct (1861/4778). Conclusion. Across all exploration strategies, VLM agents remained far less accurate than radiologists. They adopted an early hypothesis and usually kept it, seldom correcting a wrong one.

## Cost-Efficient LLM Evaluation: A Framework for Narrow Confidence Intervals Under Fixed Budgets

DevFeed: [Cost-Efficient LLM Evaluation: A Framework for Narrow Confidence Intervals Under Fixed Budgets](<https://devfeed.tech/articles/cost-efficient-llm-evaluation-a-framework-for-narrow-confidence-intervals-under-fixed-budgets-142567.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2609.29140>)

Author: Yezhou Cheng, Runjia Du, Zeming Liu, Qibai Chen, Hang Lyu, Yilan Wei, Yankai Zeng, Bojun Lin

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [open-llm-leaderboard](<https://devfeed.tech/topics/open-llm-leaderboard.md>)

Tags: [budget](<https://devfeed.tech/tags/budget.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [confidence-interval](<https://devfeed.tech/tags/confidence-interval.md>), [cost](<https://devfeed.tech/tags/cost.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-it](<https://devfeed.tech/tags/cs-it.md>), [error](<https://devfeed.tech/tags/error.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [framework](<https://devfeed.tech/tags/framework.md>), [livecodebench](<https://devfeed.tech/tags/livecodebench.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-evaluation](<https://devfeed.tech/tags/llm-evaluation.md>), [lower-bounds](<https://devfeed.tech/tags/lower-bounds.md>), [math-it](<https://devfeed.tech/tags/math-it.md>), [model-performance](<https://devfeed.tech/tags/model-performance.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [tasks](<https://devfeed.tech/tags/tasks.md>), [testing](<https://devfeed.tech/tags/testing.md>)

### AI overview

The framework allocates a fixed LLM evaluation budget across tasks and repeated runs, either requiring coverage of every task or allowing tasks to be skipped. It uses disagreements between repeats to bound estimation error, enabling narrower confidence intervals when sampled repeats show little variability. The authors prove fastest possible worst-case narrowing rates when repeats produce identical outcomes, with matching lower bounds for both coverage settings, even when evaluators adapt later runs based on earlier results. On LiveCodeBench, it produced narrower intervals than uniform random sampling in 15 of 16 model configurations, with a median width reduction of 30.6% at the same confidence level and 1,100 outputs per configuration.

### Source excerpt

arXiv:2609.29140v2 Announce Type: replace Abstract: Large language models (LLMs) play an increasingly important role in both academic research and industry. Reliable evaluation helps researchers understand model capabilities and assess progress during model development. However, evaluation can be costly, and researchers often need to assess performance within a fixed budget. Because model outputs can vary across runs, this budget must be allocated between testing different tasks and repeating selected tasks. A confidence interval quantifies the uncertainty in the resulting performance estimate. Obtaining narrower intervals at the same confidence level and evaluation cost provides a more precise assessment of model performance, helping researchers draw clearer conclusions from limited observations. We introduce a framework that allocates a fixed evaluation budget across tasks and repeated runs, with or without full task coverage. The framework then uses disagreements between runs on the same task to bound estimation error, allowing narrower confidence intervals at the same confidence level when the sampled repeats show little variability. When repeated runs on each task yield the same outcome, we prove that our framework achieves the fastest theoretically possible rates for narrowing confidence intervals in the worst case as the number of tasks and the budget for repeats grow. We establish matching lower bounds both when every task must be covered and when tasks may be skipped, even for evaluators who use earlier results to decide which run to carry out next. On LiveCodeBench, our framework yielded narrower confidence intervals on average than uniform random sampling in 15 of 16 model configurations, with a median width reduction of 30.6% at the same confidence level and the same budget of 1,100 outputs per configuration.

## Vanguard: Knowing When to Trust the Draft in Speculative Vision-Language Agents

DevFeed: [Vanguard: Knowing When to Trust the Draft in Speculative Vision-Language Agents](<https://devfeed.tech/articles/vanguard-knowing-when-to-trust-the-draft-in-speculative-vision-language-agents-142724.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.23917>)

Author: Hanqi Mao, Jiajun Fan, Ruihan Guo, Shuaike Shen, Ziwen Wang, Hui Ren, Xianrui Zhong, Ge Liu

Published: 2026-11-19T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [vlm](<https://devfeed.tech/topics/vlm.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [Physical AI](<https://devfeed.tech/topics/physical-ai.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [better](<https://devfeed.tech/tags/better.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [decision-model](<https://devfeed.tech/tags/decision-model.md>), [faster](<https://devfeed.tech/tags/faster.md>), [latency](<https://devfeed.tech/tags/latency.md>), [mathematics](<https://devfeed.tech/tags/mathematics.md>), [mobile](<https://devfeed.tech/tags/mobile.md>), [probability](<https://devfeed.tech/tags/probability.md>), [screenshot](<https://devfeed.tech/tags/screenshot.md>), [small](<https://devfeed.tech/tags/small.md>), [speculative-decoding](<https://devfeed.tech/tags/speculative-decoding.md>), [target](<https://devfeed.tech/tags/target.md>), [training](<https://devfeed.tech/tags/training.md>), [vision](<https://devfeed.tech/tags/vision.md>), [web](<https://devfeed.tech/tags/web.md>)

### AI overview

Vanguard uses a screenshot-aware decision model to judge candidate action steps from a small vision-language model. In one forward pass, it estimates each step's probability of being correct; it keeps the best draft above a threshold and otherwise asks the large target model to write the step. The abstract reports that a 2B draft with Vanguard is at least as accurate as its 8B target across web, mobile and embodied tasks, and more accurate on Mind2Web and EB-ALFRED, while the target produces only a minority of steps. Vanguard also completes more closed-loop episodes and, with a 27B target from another model family, is faster and more accurate. Its calibrated threshold transfers to unseen websites, and its advantage grows when the draft's pass@4 is strong relative to target accuracy. These are abstract-level claims; detailed methods and measurements are not provided here.

### Source excerpt

arXiv:2611.23917v1 Announce Type: new Abstract: Vision-language agents that browse websites, operate phones and act in homes reason before every action, so every step waits for a large model, a latency and cost that stand in the way of deployment. Step-level speculation lets a small model draft each step and calls the large model only on rejected drafts, but its outcome rests on the judge, and a judge that reads only a step's text cannot tell whether the step fits the screen it acts on. We find that a few samples of a small model can contain a correct step about as often as the large model gets the step right, and not always on the same steps: checked against the screen, small drafts can be not only cheaper but better. We present Vanguard, a training-free judge that leaves the draft and target models unchanged: a general decision model, with the screenshot in view, gives each candidate step a probability of being right in one forward pass; the best candidate is kept if it clears a threshold, and otherwise the large model writes the step. On web, mobile and embodied agents, a 2B draft judged by Vanguard is at least as accurate as its 8B target and significantly more accurate on Mind2Web and EB-ALFRED, while the target writes only a minority of the steps; it completes far more closed-loop episodes, and with a 27B target from another family it is both faster and more accurate. Vanguard is the most accurate judge we compare on the agent tasks; it relies on the screenshot on web and mobile tasks, its calibrated threshold transfers to unseen websites, and across model families its gain over the target grows with the draft's pass@4 relative to the target's accuracy. Vanguard composes with token-level speculative decoding and matches trained process reward models on mathematics, turning speculation from a way to save the large model's work into one that can improve on it.

## Qwen4-Max vs. Claude Opus 5.5: The cheaper model costs you somewhere else.

DevFeed: [Qwen4-Max vs. Claude Opus 5.5: The cheaper model costs you somewhere else.](<https://devfeed.tech/articles/qwen4-max-vs-claude-opus-5-5-the-cheaper-model-costs-you-somewhere-else-141057.md>)

Original publisher: [Read original article](<https://thenewstack.io/qwen4-max-vs-claude-opus-5-5/>)

Author: Jessica Wachtel

Published: 2026-11-18T19:10:39Z

Content type: comparison

Language: en

Sources: [The New Stack](<https://devfeed.tech/sources/the-new-stack.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Claude](<https://devfeed.tech/topics/claude.md>), [Language models](<https://devfeed.tech/topics/language-models.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [accounting](<https://devfeed.tech/tags/accounting.md>), [agent](<https://devfeed.tech/tags/agent.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [alibaba](<https://devfeed.tech/tags/alibaba.md>), [anthropic](<https://devfeed.tech/tags/anthropic.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [billing](<https://devfeed.tech/tags/billing.md>), [bug-fix](<https://devfeed.tech/tags/bug-fix.md>), [claude-opus-5-5](<https://devfeed.tech/tags/claude-opus-5-5.md>), [code](<https://devfeed.tech/tags/code.md>), [coding](<https://devfeed.tech/tags/coding.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [costs](<https://devfeed.tech/tags/costs.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [qwen4-max](<https://devfeed.tech/tags/qwen4-max.md>), [software-testing](<https://devfeed.tech/tags/software-testing.md>)

### AI overview

In three single-run tasks, Qwen4-Max cost about a third as much as Claude Opus 5.5 overall, but took about 3.14 times longer and scored 170/183 versus Opus's 179/183. Results varied by task: Qwen scored 71/76 on a multi-app workflow, missing points after misreading an ownership rule; it edged Opus 59/60 to 58/60 on incident investigation at about 31% of the reported cost, while taking longer; and it scored 40/47 on a coding task versus Opus's 45/47, missing a payment-rounding bug and leaving a scratch file. Opus fixed all planted bugs but missed a flawed test the task required reporting. The author sees Qwen as a lower-cost option for investigation when extra time is acceptable, and favors Opus for coding or consequential work. Each model ran once, so small score differences may shift; the tests do not establish general performance rankings. The excerpt also notes different APIs, cost-estimation methods, and benchmark setups.

### Source excerpt

Qwen4-Max, made available in preview to Token Plan subscribers on November 11 and generally released on November 16, is quite The post Qwen4-Max vs. Claude Opus 5.5: The cheaper model costs you somewhere else. appeared first on The New Stack.

## Is model self-assessment enough for diverse LVLM ensembles?

DevFeed: [Is model self-assessment enough for diverse LVLM ensembles?](<https://devfeed.tech/articles/is-model-self-assessment-enough-for-diverse-lvlm-ensembles-141825.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.22956>)

Author: Michael Smith, Frank P. Ferrie

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [ai evaluation](<https://devfeed.tech/topics/ai-evaluation.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [aggregation](<https://devfeed.tech/tags/aggregation.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [best-of](<https://devfeed.tech/tags/best-of.md>), [candidate](<https://devfeed.tech/tags/candidate.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-cv](<https://devfeed.tech/tags/cs-cv.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [dataset](<https://devfeed.tech/tags/dataset.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [free](<https://devfeed.tech/tags/free.md>), [key](<https://devfeed.tech/tags/key.md>), [language](<https://devfeed.tech/tags/language.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [metrics](<https://devfeed.tech/tags/metrics.md>), [model](<https://devfeed.tech/tags/model.md>), [models](<https://devfeed.tech/tags/models.md>), [open](<https://devfeed.tech/tags/open.md>), [parameter](<https://devfeed.tech/tags/parameter.md>), [performance](<https://devfeed.tech/tags/performance.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>)

### AI overview

A unified evaluation of candidate-response selection across 34 confidence-estimation methods, 24 ensemble configurations, 20 datasets, and two prompt specifications finds that simple training-free approaches using each model's own confidence can outperform constituent models and compete with more involved methods on accuracy, including open-ended tasks. Performance varies substantially with dataset, ensemble composition, and prompt wording; ensembles can underperform their strongest member. The authors recommend training-free ensembles only when deployment conditions can be characterized precisely, suggesting Best-of-N configurations of strong models as a starting point. These findings are based on the supplied abstract.

### Source excerpt

arXiv:2611.22956v1 Announce Type: cross Abstract: Ensembles of diverse Large Vision Language Models have demonstrated performance beyond that of single models across multiple datasets and tasks. However, the necessary aggregation process usually relies on training, parameter fitting, or complex task-specific pipelines. Furthermore, evaluation is limited, making a fair comparison difficult. We create a unified pipeline where we perform candidate response selection: we build an ensemble by selecting the best answer from a set provided by different models. We comprehensively evaluate across 34 different means of confidence estimation, 24 ensemble configurations, 20 datasets, and two prompt specifications. We show that simple training and parameter-free approaches can be sufficient, even for ensembles of different architectures. These simple self-assessment approaches rely solely on each constituent model's confidence, without explicit calibration, yet are able to perform better than said constituent models. They are also competitive with more involved alternatives on standard metrics such as accuracy, including on open-ended problems. However, we also discover that dataset, ensemble composition, and prompt wording are key confounding variables insufficiently explored by prior work that significantly affect ensemble performance, such that the ensemble can perform worse than its best constituent model. These must be optimized for to get the most out of ensembles, at the cost of limiting their generalizability. We recommend that practitioners consider training-free ensembles only if deployment conditions can be precisely characterized, with Best-of-N configurations of strong models being a good starting point.

## Toward AI epidemiology: a measurement standardization framework for prospective risk detection

DevFeed: [Toward AI epidemiology: a measurement standardization framework for prospective risk detection](<https://devfeed.tech/articles/toward-ai-epidemiology-a-measurement-standardization-framework-for-prospective-risk-detection-139753.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2512.15783>)

Author: Kit Tempest-Walters

Published: 2026-11-18T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Human-AI evaluation](<https://devfeed.tech/topics/human-ai-evaluation.md>), [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [aggregate](<https://devfeed.tech/tags/aggregate.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-systems](<https://devfeed.tech/tags/ai-systems.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [analysis](<https://devfeed.tech/tags/analysis.md>), [application](<https://devfeed.tech/tags/application.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [bootstrap](<https://devfeed.tech/tags/bootstrap.md>), [bounded](<https://devfeed.tech/tags/bounded.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [deployment](<https://devfeed.tech/tags/deployment.md>), [detection](<https://devfeed.tech/tags/detection.md>), [domains](<https://devfeed.tech/tags/domains.md>), [experts](<https://devfeed.tech/tags/experts.md>), [fields](<https://devfeed.tech/tags/fields.md>), [framework](<https://devfeed.tech/tags/framework.md>), [governance](<https://devfeed.tech/tags/governance.md>), [inference](<https://devfeed.tech/tags/inference.md>), [measurement](<https://devfeed.tech/tags/measurement.md>), [risk](<https://devfeed.tech/tags/risk.md>), [standardization](<https://devfeed.tech/tags/standardization.md>)

### AI overview

The paper proposes standardizing expert-AI interactions into eight structured fields so large language models can assess evidential and policy alignment without access to model internals. It frames alignment scores as possible deployment signals and, after measurement is established, as a basis for studying associations with downstream outcomes in regulated settings. A minimal application to a published corpus reproduced scores across two runs under specified conditions; reliability at scale remains unvalidated. The paper specifies a testing protocol using paired bootstrap inference, DeLong's test as an AUC sensitivity check, a one-sided non-inferiority margin of 0.05, and Holm-Bonferroni correction. Population-level claims are proposed aims of a staged research program, not findings reported here.

### Source excerpt

arXiv:2512.15783v5 Announce Type: replace Abstract: This paper proposes a measurement standardization framework that compresses expert-AI interactions into structured, comparable fields for prospective risk detection in deployed AI systems, without access to model internals. This concept paper defines the framework's scope, semantically and statistically, and specifies a protocol for its empirical testing. The population-level claims it is designed to support therefore belong to a staged research program rather than to results claimed here. Measurement standardization underpins three claims. The first is a reliability claim: under bounded conditions, large language models can produce reliable, standardized assessments of the evidential and policy alignment of expert-AI interactions. The second is a governance claim: alignment scores give experts an immediate signal during deployment and give institutions a basis for monitoring alignment patterns across mission types, models, and domains. The third is an outcome validation claim: once measurement standardization is established, aggregate alignment scores could be used to study associations with downstream outcomes in regulated professional settings. This introduces the possibility of an "AI epidemiology", a form of risk detection based on correlated variables instead of mechanistic analysis, inspired by epidemiological reasoning. A minimal application of the protocol to a published expert-AI corpus shows that the judge reproduces its policy and evidential alignment scores across two runs under the specified conditions. Judge reliability at scale remains to be validated in future work. The paper sets out a defined grammar of eight interaction fields, together with a statistical protocol based on paired bootstrap inference, DeLong's test for paired AUCs as a sensitivity check, a pre-specified one-sided non-inferiority margin of 0.05, and Holm-Bonferroni correction.

## One-Shot Calibration via Model Editing

DevFeed: [One-Shot Calibration via Model Editing](<https://devfeed.tech/articles/one-shot-calibration-via-model-editing-137586.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20562>)

Author: Jiaxing Qiu, Luke Hakso, Roxana Daneshjou, Ahmed Alaa, Thomas Hartvigsen

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [generalization in machine learning](<https://devfeed.tech/topics/generalization-in-machine-learning.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [vlm](<https://devfeed.tech/topics/vlm.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [datasets](<https://devfeed.tech/tags/datasets.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [model](<https://devfeed.tech/tags/model.md>), [performance](<https://devfeed.tech/tags/performance.md>), [target](<https://devfeed.tech/tags/target.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [vision-language-models](<https://devfeed.tech/tags/vision-language-models.md>)

### AI overview

SoftEdit uses multiple plausible user responses and their varying likelihoods to form a target distribution, allowing a model edit to calibrate generations rather than replace them with a single answer. To generalize to unseen samples, it uses co-plausibility relationships between answers while preserving real distribution shifts, addressing a tradeoff between generality and locality. The abstract reports state-of-the-art performance across three language models, three vision-language models, and two large datasets; it provides no quantitative results or further method details.

### Source excerpt

arXiv:2611.20562v1 Announce Type: cross Abstract: Model editing aims to correct errors in large, pretrained models by replacing model generations with user-provided responses. Existing editors assume a single target answer for each edit. However, many real tasks inherently entail uncertainty, resulting in multiple plausible answers with varying likelihoods in user feedback. Such feedback forms a target distribution to not only correct but also calibrate model generations. We therefore propose SoftEdit, the first model editor to leverage uncertainty in user responses and perform one-shot calibration during editing. To generalize calibration to unseen samples, SoftEdit uses co-plausibility relationships between answers while preserving real distributional shifts, balancing the new distributional generality-locality tradeoff. Across three language models, three vision-language models, and two large datasets, SoftEdit achieves state-of-the-art performance, ultimately showing that uncertainty in user responses enables generalizable calibration of pretrained models beyond single-answer corrections.

## Wrong with Confidence: Measuring Hallucination, Overconfidence and Identity-Conditioned Behaviour in Large Language Models

DevFeed: [Wrong with Confidence: Measuring Hallucination, Overconfidence and Identity-Conditioned Behaviour in Large Language Models](<https://devfeed.tech/articles/wrong-with-confidence-measuring-hallucination-overconfidence-and-identity-conditioned-behaviour-in-large-language-models-137334.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18728>)

Author: Md. Rejaul Korim Sadi, Toufiqur Rahman Tasin, Golam Mostofa Naeem, Dewan Fahad Chowdhury, Hatim Al Amin Chowdhury

Published: 2026-11-17T05:00:00Z

Content type: news

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Hallucination detection](<https://devfeed.tech/topics/hallucination-detection.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [AI Chat](<https://devfeed.tech/topics/ai-chat.md>)

Tags: [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [identity](<https://devfeed.tech/tags/identity.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [risk](<https://devfeed.tech/tags/risk.md>)

### AI overview

In a paired study of five open-weight language models, removing a question's referent increased errors among attempted answers from 39.1% to 52.0%; models asked for clarification on 33.7% of underspecified prompts. Wrong answers rarely signaled uncertainty: 75.3% had no hedge, and hedging weakly distinguished correct from incorrect responses (AUROC 0.56). Specialist-domain questions failed more often than general ones across all five models. The study also found occupational stereotypes in gendered pronouns, answers about 45% longer when a user identity was stated, and no reduction in safety caveats when users claimed expertise. The authors recommend evaluating clarification separately from answer accuracy and checking label validity as well as rater agreement.

### Source excerpt

arXiv:2611.18728v1 Announce Type: cross Abstract: Large language models answer underspecified questions readily, and accuracy-only evaluation rewards guessing over abstaining. Observational studies link query underspecification to hallucination, but they cannot separate the effect of missing information from item difficulty, and they rarely ask what the model tells the user. We report a pre-registered, paired behavioural study of five open-weight models (Gemma 3 27B, Gemma 4 31B, gpt-oss-120b, Qwen3-Next-80B-A3B and Nemotron 3 Super). Each of 142 factual questions was posed in a fully specified form and in a referentially underspecified form with the same answer key, and three probes measured identity-conditioned behaviour. Two raters scored 5,620 responses under a fixed rubric; effects were estimated with mixed-effects models. Without the referent, the models sought clarification on 33.7% of prompts and answered the rest. Among attempted answers, the error rate rose from 39.1% to 52.0% (mixed-effects odds ratio 4.12, 95% CI 2.76-6.14), and specialist-domain questions failed more often than general ones in every model. Unelicited verbal uncertainty did not track this risk: 75.3% of wrong answers carried no hedge, wrong answers to underspecified prompts were hedged no more often than wrong answers to clear prompts (27.3% vs 22.3%), and hedging separated correct from incorrect answers only weakly (AUROC 0.56). Gendered pronouns followed occupational stereotypes whenever a model committed to one; any stated user identity lengthened answers by about 45% relative to an anonymous user; claimed expertise did not reduce safety caveats. The models either ask or answer as if the question were well posed, and the wording of a guess does not reveal that it is one. Evaluations should score clarification as a distinct outcome and check label validity as well as rater agreement.

## Benchmarking Large Language Models for Linguistic Annotation

DevFeed: [Benchmarking Large Language Models for Linguistic Annotation](<https://devfeed.tech/articles/benchmarking-large-language-models-for-linguistic-annotation-137676.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21291>)

Author: Ond\\v{r}ej Tich\\'y (Faculty of Arts, Charles University), Anna Marklov\\'a (Faculty of Arts, Charles University), Barbora Bulantov\\'a (Faculty of Arts, Charles University), Magdalena Titlbachov\\'a (Fac

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [annotation](<https://devfeed.tech/tags/annotation.md>), [annotations](<https://devfeed.tech/tags/annotations.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [discourse](<https://devfeed.tech/tags/discourse.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [framework](<https://devfeed.tech/tags/framework.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models](<https://devfeed.tech/tags/large-language-models.md>), [methodology](<https://devfeed.tech/tags/methodology.md>), [predictions](<https://devfeed.tech/tags/predictions.md>), [prompts](<https://devfeed.tech/tags/prompts.md>), [python](<https://devfeed.tech/tags/python.md>), [python-benchmarking](<https://devfeed.tech/tags/python-benchmarking.md>), [quality](<https://devfeed.tech/tags/quality.md>), [rankings](<https://devfeed.tech/tags/rankings.md>), [semantics](<https://devfeed.tech/tags/semantics.md>)

### AI overview

dhAIBench is a Python benchmarking and annotation tool with a web interface for exploring and sharing results. It records prompts, configurations, and predictions so users can compare annotation quality, confidence calibration, repeatability, and estimated cost under documented conditions. The authors demonstrate it with 293 runs of 48 models across nine tasks covering morphology, syntax, semantics, discourse-pragmatics, spelling normalization, and learner-language correction. Models show high agreement with reference annotations on several tasks, including work on non-standardized historical varieties and tasks requiring contextual interpretation. The findings support evaluating models on specific linguistic tasks rather than relying on general rankings; current and future benchmarks are made publicly available.

### Source excerpt

arXiv:2611.21291v1 Announce Type: cross Abstract: Assessing large language models for linguistic annotation requires task-specific evaluation under documented conditions. We present dhAIBench, a Python benchmarking and annotation tool with a web interface for exploring and sharing results. It records prompts, configurations and predictions, enabling comparisons of annotation quality, confidence calibration, repeatability and estimated cost. We demonstrate the framework and the associated methodology through 293 runs of 48 models on nine linguistic annotation tasks spanning morphology, syntax, semantics, discourse-pragmatics, spelling normalization and learner-language correction. Models achieve high agreement with reference annotations on several tasks, including those on non-standardized, historical varieties as well as requiring contextual interpretation. The results support evaluating models against specific linguistic tasks, rather than relying on general rankings. To this end, we make the current and future benchmarks publicly available at dhaibench.silent3.ff.cuni.cz.

## Advancing the Calibration-Compute Frontier of LLM Classifiers via Asymmetric Duos

DevFeed: [Advancing the Calibration-Compute Frontier of LLM Classifiers via Asymmetric Duos](<https://devfeed.tech/articles/advancing-the-calibration-compute-frontier-of-llm-classifiers-via-asymmetric-duos-138820.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19020>)

Author: Berke Akturk, Ali Can Sahin, Rasit Cem Solatay, Emanuel Sommer, Yawei Li, David R\\"ugamer

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [classifiers](<https://devfeed.tech/tags/classifiers.md>), [compute](<https://devfeed.tech/tags/compute.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [fusion](<https://devfeed.tech/tags/fusion.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llms](<https://devfeed.tech/tags/llms.md>), [sidekick](<https://devfeed.tech/tags/sidekick.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [uncertainty-estimation](<https://devfeed.tech/tags/uncertainty-estimation.md>)

### AI overview

The Asymmetric LLM Duo calibrates classifier uncertainty by fusing a strong base model's logits with those of a smaller sidekick, using global scaling parameters tuned on validation data. Across diverse benchmarks, the authors report best or second-best expected calibration error and negative log-likelihood without reducing predictive accuracy, at a fraction of the computational cost of more complex baselines. They also report that the method can be combined with established calibration techniques and that its logit-space behavior mitigates base-model overconfidence. The supplied evidence is an abstract, so it does not provide benchmark details or quantitative results.

### Source excerpt

arXiv:2611.19020v1 Announce Type: new Abstract: The importance of reliable uncertainty estimation in large language models (LLMs) has become increasingly pronounced, yet modern models remain poorly calibrated and frequently overconfident. In this study, we address this by proposing an efficient method for LLM classifier calibration using the Asymmetric Duos framework. Our approach combines a strong base model with a smaller sidekick model via logit-space fusion, utilizing global scaling parameters optimized on a validation set. Across diverse benchmarks, this Asymmetric LLM Duo consistently achieves best or second-best performance on expected calibration error and negative log-likelihood, without degrading predictive accuracy. More importantly, these gains are realized at a fraction of the computational cost of more complex baselines, placing our method on the calibration-compute Pareto frontier. We further demonstrate that our method acts orthogonal to various established calibration-aware techniques, meaning it can be easily combined with them to consistently boost their performance. Finally, our analysis of the logit-space behavior illustrates how asymmetric fusion successfully mitigates base model overconfidence. Overall, our findings demonstrate that the benefits of an asymmetric fusion successfully transfer to the language domain, providing a simple, highly cost-efficient approach to improving uncertainty estimation in LLM classifiers.

## Scaling large language model-assisted line-of-therapy adjudication with human-in-the-loop evidence-calibrated selection

DevFeed: [Scaling large language model-assisted line-of-therapy adjudication with human-in-the-loop evidence-calibrated selection](<https://devfeed.tech/articles/scaling-large-language-model-assisted-line-of-therapy-adjudication-with-human-in-the-loop-evidence-calibrated-selection-139015.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20259>)

Author: Olatunji Oni, Sanhita Sengupta, Marc De Benedetti, Nathanial Eddy

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [error](<https://devfeed.tech/tags/error.md>), [human-in-the-loop](<https://devfeed.tech/tags/human-in-the-loop.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [manual-review](<https://devfeed.tech/tags/manual-review.md>), [q-bio-qm](<https://devfeed.tech/tags/q-bio-qm.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [scaling](<https://devfeed.tech/tags/scaling.md>), [selection](<https://devfeed.tech/tags/selection.md>), [simulations](<https://devfeed.tech/tags/simulations.md>), [stat-ap](<https://devfeed.tech/tags/stat-ap.md>), [study](<https://devfeed.tech/tags/study.md>)

### AI overview

Evidence-calibrated selection (ECS) uses clinician adjudications to calibrate confidence in LLM line-of-therapy assignments, send the least-supported cases for review, and feed completed reviews back as precedent. In simulations involving 136 clinician-adjudicated multiple myeloma patients, ECS reached 5% residual error after clinicians reviewed 44.6% of patients. Reaching that error level required review of 62.2% under LLM self-reported confidence and 81.9% under random sampling. The reported results suggest ECS can reduce review burden in study-specific adjudication; the supplied abstract describes simulations, not broader clinical validation.

### Source excerpt

arXiv:2611.20259v1 Announce Type: new Abstract: Large language models (LLMs) have enabled the ability to scale clinical adjudication tasks traditionally dependent on exhaustive manual review. Line-of-therapy (LoT) adjudication, a major bottleneck in oncology cohort construction, is a prime example. Line boundaries must be inferred from longitudinal treatment histories according to nuanced, study-specific definitions, at every regimen transition for every patient. LLMs can apply this adjudication at scale, but their reliability varies across patients. Treatment histories differ in complexity, and ambiguous or atypical patterns depend on expert clinical judgment. Reliable automation thus requires verified support for model assignments and the ability to identify cases that warrant manual review. Here, we present evidence-calibrated selection (ECS), a human-in-the-loop framework for scaling LLM-assisted LoT adjudication. ECS derives confidence from observed clinician adjudications, prioritizes patients with the least-supported assignments for follow-up review, and returns completed adjudications to the model as precedent. As the evidence base grows, ECS continuously calibrates confidence to estimate residual error, thus enabling more efficient use of clinician time. In simulations on 136 clinician adjudicated patients with multiple myeloma, ECS achieved a residual error of 5% after review of 44.6% of patients, compared to 62.2% when review was directed by LLM self-reported confidence and 81.9% under random sampling. These results support a scalable approach to study-specific LoT adjudication, using clinician evidence to inform automated predictions, minimize error, and reduce review burden.

## Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States

DevFeed: [Confidence Without Cause: Predicting and Localizing Physics-Shortcut Sensitivity in Language Model Hidden States](<https://devfeed.tech/articles/confidence-without-cause-predicting-and-localizing-physics-shortcut-sensitivity-in-language-model-hidden-states-137444.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19403>)

Author: Shubham Mishra, Atharv Sharma, Dhruv Kumar

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Language models](<https://devfeed.tech/topics/language-models.md>), [Ground truth / benchmark quality](<https://devfeed.tech/topics/ground-truth-benchmark-quality.md>), [qwen](<https://devfeed.tech/topics/qwen.md>), [machine learning overfitting](<https://devfeed.tech/topics/machine-learning-overfitting.md>)

Tags: [attention](<https://devfeed.tech/tags/attention.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [hidden-states](<https://devfeed.tech/tags/hidden-states.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [layer](<https://devfeed.tech/tags/layer.md>), [linear](<https://devfeed.tech/tags/linear.md>), [physics](<https://devfeed.tech/tags/physics.md>)

### AI overview

DECOY uses 155 families of four physics prompts, with a symbolic solver certifying that each changed detail leaves the answer unchanged, to measure sensitivity to irrelevant distractors. Among families with labels, Qwen2.5-7B-Instruct was sensitive on 50.0% and Llama-3.1-8B-Instruct on 22.9%; effects were largest when distractor values were numerically close to the answer. A Qwen hidden-state probe predicted sensitivity before generation (AUROC 0.760 with whole families held out), performing on par with a fitted token-entropy baseline (0.727) and achieving the best tested Brier score (0.204 vs. 0.214). On 15 families built after training, a frozen probe reached 0.820 pooled AUROC, but only 0.667-0.6875 within cue types. Knocking out four layer-16 attention heads canceled sensitivity on a 20-family panel. The results suggest detectability and a targeted intervention, with limited evidence for generalization across cue types.

### Source excerpt

arXiv:2611.19403v1 Announce Type: new Abstract: A language model's confidence on a physics problem deserves trust only if it ignores details irrelevant to the answer. We introduce DECOY (Distractor-Embedded Certified cOunterfactuals for phYsics): 155 families of four prompts differing in one irrelevant detail. A symbolic solver certifies that each edit preserves the answer. Families are labelled by the largest drop in the model's log-probability of the certified answer, using a threshold fixed before probe training. Qwen2.5-7B-Instruct is sensitive on 50.0% of the 140 families yielding labels, and Llama-3.1-8B-Instruct on 22.9%. Effects are largest when distractor values lie numerically close to the answer. A linear probe on Qwen's layer-18 hidden state, averaged over four renderings, predicts sensitivity before generation at AUROC 0.760 with whole families held out. A separately tuned single-rendering probe reaches 0.747. A frozen layer-16 probe reaches 0.820 pooled AUROC on 15 families built after training, though within-cue-type AUROC is lower (0.667-0.6875). Under a paired family bootstrap, the probe performs on par with a fitted token-entropy baseline (AUROC 0.727) and achieves the lowest Brier score among tested predictors (0.204 versus 0.214). Knocking out four attention heads at layer 16 cancels sensitivity on a 20-family panel, localising it to a narrow mechanism. These results show that Qwen's confidence can track irrelevant details, that this failure is detectable before generation, and that targeted intervention can suppress it. DECOY provides certified ground truth and paired-tested baselines for evaluating hidden-state monitors.

## SG-LegalClarify: A Large-Scale Benchmark for Legal Reasoning and Ambiguity Recognition in Singapore Law

DevFeed: [SG-LegalClarify: A Large-Scale Benchmark for Legal Reasoning and Ambiguity Recognition in Singapore Law](<https://devfeed.tech/articles/sg-legalclarify-a-large-scale-benchmark-for-legal-reasoning-and-ambiguity-recognition-in-singapore-law-137478.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19643>)

Author: Ziwei Gu, Michael S. Lee, Drew Rein, Yash Maurya, Vikneswaran Kumaran, Darshini Ramiah, Chenguang Wang, Udari Madhushani Sehwag

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [llm-reasoning](<https://devfeed.tech/topics/llm-reasoning.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [domains](<https://devfeed.tech/tags/domains.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [language-models](<https://devfeed.tech/tags/language-models.md>), [large](<https://devfeed.tech/tags/large.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [law](<https://devfeed.tech/tags/law.md>), [legal](<https://devfeed.tech/tags/legal.md>), [llms](<https://devfeed.tech/tags/llms.md>), [queries](<https://devfeed.tech/tags/queries.md>), [questions](<https://devfeed.tech/tags/questions.md>), [reasoning](<https://devfeed.tech/tags/reasoning.md>), [recognition](<https://devfeed.tech/tags/recognition.md>), [research](<https://devfeed.tech/tags/research.md>), [scale](<https://devfeed.tech/tags/scale.md>), [singapore](<https://devfeed.tech/tags/singapore.md>), [web](<https://devfeed.tech/tags/web.md>), [web-search](<https://devfeed.tech/tags/web-search.md>)

### AI overview

SG-LegalClarify tests whether language models can reason under Singapore law and recognize when lay queries omit outcome-relevant facts. It includes 1,100 lawyer-authored questions across 10 domains, expanded to 1,860 evaluation instances with hidden facts and fully specified counterparts, and uses lawyer-authored IRAC rubrics. In a multi-turn evaluation of 19 models, a simulator reveals hidden facts only when a model asks for clarification. The benchmark remains unsaturated: the best model met 69.8% of mandatory rubric items closed-book and 77.8% with web search; averages were 53.8% and 63.0%, respectively. Models also struggled to identify when clarification was needed.

### Source excerpt

arXiv:2611.19643v1 Announce Type: cross Abstract: Singapore offers a distinctive testbed for legal reasoning by large language models (LLMs). Its legal system combines inherited English common law with jurisdiction-specific statutes, requiring models to distinguish local doctrine from generic common-law intuition. This setting also highlights the challenge of recognizing underspecification in lay legal queries, because the facts material to a legal outcome depend on the governing doctrine. Yet legal benchmarks rarely cover Singapore law or test this form of ambiguity recognition. To address these gaps, we introduce SG-LegalClarify, a large-scale benchmark for open-ended legal reasoning and ambiguity recognition in Singapore law. SG-LegalClarify contains 1,100 lawyer-authored source questions spanning 10 legal domains, including 340 fully specified and 760 underspecified questions. Each underspecified question has a fixed hidden-fact assignment and a fully specified counterpart, yielding 1,860 evaluation instances. Responses are graded with lawyer-authored Issue, Rule, Application, and Conclusion (IRAC) rubrics. We evaluate 19 LLMs, closed-book and with web search, using a multi-turn pipeline in which models decide whether to answer or clarify, and a user simulator reveals hidden facts only when asked. We find the benchmark far from saturated: the best models satisfy only 69.8% of mandatory rubric items closed-book (average 53.8%) and 77.8% with web search (average 63.0%). Models also struggle to recognize when clarification is necessary. We will release the public split of our benchmark, the model outputs on it, and the evaluation pipeline to support future research.

## Tight PAC-Bayes Generalisation Guarantees for Large Language Model Judges

DevFeed: [Tight PAC-Bayes Generalisation Guarantees for Large Language Model Judges](<https://devfeed.tech/articles/tight-pac-bayes-generalisation-guarantees-for-large-language-model-judges-139098.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20818>)

Author: Tom A. Lamb, Philip H. S. Torr, Tim G. J. Rudner

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [lora](<https://devfeed.tech/topics/lora.md>), [peft](<https://devfeed.tech/topics/peft.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-agents](<https://devfeed.tech/tags/ai-agents.md>), [bayes](<https://devfeed.tech/tags/bayes.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [certificates](<https://devfeed.tech/tags/certificates.md>), [classification](<https://devfeed.tech/tags/classification.md>), [coding](<https://devfeed.tech/tags/coding.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [compression](<https://devfeed.tech/tags/compression.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data](<https://devfeed.tech/tags/data.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [formal](<https://devfeed.tech/tags/formal.md>), [framework](<https://devfeed.tech/tags/framework.md>), [generator](<https://devfeed.tech/tags/generator.md>), [guarantees](<https://devfeed.tech/tags/guarantees.md>), [guardrail](<https://devfeed.tech/tags/guardrail.md>), [language-model](<https://devfeed.tech/tags/language-model.md>), [large-language-model](<https://devfeed.tech/tags/large-language-model.md>), [layers](<https://devfeed.tech/tags/layers.md>)

### AI overview

The paper applies PAC-Bayes bounds to certify safety judges on prompt-response pairs from a deployed generator. It reports that standard parameter-efficient fine-tuning can produce non-vacuous, sometimes tight certificates for classification and Brier risk without a held-out validation set. Its LoRA-GT method shares one quantisation scale across layers and gives the tightest certificates with minimal predictive-performance degradation. Across five guardrail models and four benchmarks, certificate tightness depends more on adaptation complexity than base-model scale. The authors also examine when quantisation raises risk enough to outweigh the tighter bounds from shorter descriptions.

### Source excerpt

arXiv:2611.20818v1 Announce Type: new Abstract: Large language model (LLM) judges are increasingly used to monitor the outputs of LLM-based AI agents, yet formal guarantees on the reliability of these judges remain largely absent, limiting confidence in their use for safety-critical oversight. We consider certifying such a judge and define its risk over safety classifications of prompt--response pairs drawn from a deployed generator. We apply a PAC-Bayes framework to this setting, refining existing compression-based methodology with the use of tighter bounds and more efficient coding. We show that standard parameter-efficient fine-tuning can yield non-vacuous and even tight certificates for safety judges, for both classification and Brier risk, without holding out data for validation. We introduce LoRA-GT, which shares a single quantisation scale across layers and achieves the tightest certificates with minimal degradation in predictive performance. Across five guardrail models and four benchmarks, we find that certificate tightness is governed more by adaptation complexity than by base model scale. Finally, we show how far quantisation can be pushed before the resulting increase in risk outweighs the bound tightening gained from shorter descriptions. Together, these results provide a practical route toward certifiable scalable oversight and offer insight into how the complexity of safety adaptations shapes generalisation.

## Black-Box Uncertainty Quantification for Large Language Models via Ensemble-of-Ensembles

DevFeed: [Black-Box Uncertainty Quantification for Large Language Models via Ensemble-of-Ensembles](<https://devfeed.tech/articles/black-box-uncertainty-quantification-for-large-language-models-via-ensemble-of-ensembles-135200.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.18453>)

Author: Wang Ma, Debarun Bhattacharjya, Junkyu Lee, Nhan H Pham, Harsha Kokel, Qiang Ji

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>)

Tags: [bayesian](<https://devfeed.tech/tags/bayesian.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [bias](<https://devfeed.tech/tags/bias.md>), [class](<https://devfeed.tech/tags/class.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [continuous](<https://devfeed.tech/tags/continuous.md>), [contribute](<https://devfeed.tech/tags/contribute.md>), [covariance](<https://devfeed.tech/tags/covariance.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [decoding](<https://devfeed.tech/tags/decoding.md>), [embedding](<https://devfeed.tech/tags/embedding.md>), [error](<https://devfeed.tech/tags/error.md>), [estimator](<https://devfeed.tech/tags/estimator.md>), [eu](<https://devfeed.tech/tags/eu.md>), [failure](<https://devfeed.tech/tags/failure.md>), [gap](<https://devfeed.tech/tags/gap.md>), [generation](<https://devfeed.tech/tags/generation.md>), [large-language-models-llms](<https://devfeed.tech/tags/large-language-models-llms.md>), [ranking](<https://devfeed.tech/tags/ranking.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [uncertainty-quantification](<https://devfeed.tech/tags/uncertainty-quantification.md>)

### AI overview

The method estimates prompt ambiguity (aleatoric uncertainty) separately from model knowledge gaps (epistemic uncertainty) using only black-box sampling. It samples stochastic answers to a fixed prompt, then applies meaning-preserving prompt perturbations; embedding-space variability and the law of total covariance decompose within- and between-perturbation uncertainty. The work also gives a finite-sample bias correction, bounds linking its estimates to Bayesian targets, and identifies failure modes consistency-only black-box methods miss. Across five short QA benchmarks and four instruction-tuned models, it was competitive with white-box baselines and matched leading black-box methods on error ranking at equal generation budgets. A long-form pilot reports claim-level factuality assessment without retraining.

### Source excerpt

arXiv:2611.18453v1 Announce Type: cross Abstract: Reliable use of large language models (LLMs) requires uncertainty estimates that separate ambiguity intrinsic to a prompt from genuine knowledge gaps. Bayesian and deep-ensemble methods deliver this aleatoric/epistemic decomposition in principle, but are computationally prohibitive at LLM scale and require white-box access. Existing black-box methods, in contrast, summarize to a single scalar that cannot make the distinction. We introduce a two-level ensemble that bridges this gap. Stochastic decoding of a fixed prompt forms the inner ensemble; meaning-preserving semantic perturbations of the prompt form the outer ensemble. Variability is measured in a continuous embedding space, and the law of total covariance gives an exact within-/between-perturbation decomposition to estimate aleatoric and epistemic uncertainty (AU/EU); it requires only black-box sampling access. On top of this estimator, we contribute three theoretical results that ground the framework: (i) a finite-sample bias identity for nested Monte Carlo with a closed-form bias-corrected epistemic estimator; (ii) a Bayesian bridge bounding the gap between perturbation-population AU/EU and Bayesian targets; and (iii) a separation result identifying a class of failure modes that any consistency-only black-box method provably misses while perturbation-based EU detects. Across five short QA benchmarks and four instruction-tuned LLMs spanning two families and two scales, our estimator is competitive with strong white-box baselines, and at a matched generation budget it matches leading black-box methods on error ranking while decomposing AU and EU; controlled paired interventions show that AU and EU respond selectively to distinct, actionable sources of failure: prompt ambiguity and model knowledge gaps; and a long-form pilot on the NQ long-answer split shows the framework extends to claim-level factuality without retraining.

## Model-Free Assessment of Simulator Fidelity via Quantile Curves

DevFeed: [Model-Free Assessment of Simulator Fidelity via Quantile Curves](<https://devfeed.tech/articles/model-free-assessment-of-simulator-fidelity-via-quantile-curves-134952.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2512.05024>)

Author: Garud Iyengar, Yu-Shiou Willy Lin, Kaizheng Wang

Published: 2026-11-16T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [ai evaluation](<https://devfeed.tech/topics/ai-evaluation.md>), [ai alignment](<https://devfeed.tech/topics/ai-alignment.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [abstract](<https://devfeed.tech/tags/abstract.md>), [accessible](<https://devfeed.tech/tags/accessible.md>), [ai](<https://devfeed.tech/tags/ai.md>), [ai-models](<https://devfeed.tech/tags/ai-models.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [assessment](<https://devfeed.tech/tags/assessment.md>), [call](<https://devfeed.tech/tags/call.md>), [challenge](<https://devfeed.tech/tags/challenge.md>), [comparisons](<https://devfeed.tech/tags/comparisons.md>), [condition](<https://devfeed.tech/tags/condition.md>), [conditional](<https://devfeed.tech/tags/conditional.md>), [confidence](<https://devfeed.tech/tags/confidence.md>), [continuous](<https://devfeed.tech/tags/continuous.md>), [critical](<https://devfeed.tech/tags/critical.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [data](<https://devfeed.tech/tags/data.md>), [free](<https://devfeed.tech/tags/free.md>), [model](<https://devfeed.tech/tags/model.md>), [quantile](<https://devfeed.tech/tags/quantile.md>), [simulator](<https://devfeed.tech/tags/simulator.md>), [stat-me](<https://devfeed.tech/tags/stat-me.md>)

### AI overview

The method assesses simulator fidelity by estimating a quantile curve for a robust proxy of the sim-to-real gap across scenarios. Because real and simulated population parameters are latent and observed through finite, potentially unequal samples, it constructs confidence sets for those parameters before estimating the proxy's quantile function. The resulting risk profile supports summaries such as Conditional Value-at-Risk, inference about real outputs in a new scenario, and comparisons across simulators. The approach is described as model-agnostic and applicable to categorical and continuous multidimensional outputs; the abstract reports an evaluation of four major LLMs against human populations on WorldValueBench.

### Source excerpt

arXiv:2512.05024v4 Announce Type: replace-cross Abstract: As generative AI models are increasingly used to simulate real-world systems, quantifying the ``sim-to-real'' gap is critical. For each input setting of interest -- which we call a \emph{scenario}, such as a survey question or operating condition -- the real and simulated systems are associated with unobserved latent population parameters, and their discrepancy varies across scenarios. A fundamental challenge is that, for any given scenario, this discrepancy cannot be observed directly, since both systems are accessible only through finite samples, often of heterogeneous sizes across scenarios. Standard predictive inference methods are therefore ill-suited, as they quantify uncertainty in observable outputs rather than latent population parameters. To address this, we construct confidence sets for these latent parameters and use them to derive a robust proxy for the sim-to-real discrepancy. We then estimate the quantile function of this proxy to obtain a distribution-level risk profile of the simulator, which supports a broad range of statistical summaries, including statistical inference for the real output distribution in a new scenario, the calculation of risk measures like Conditional Value-at-Risk (CVaR), and principled comparisons across simulators. Our method is model-agnostic and handles general output spaces, such as categorical survey responses and continuous multi-dimensional data. We demonstrate the practical utility of this method by evaluating the alignment of four major LLMs with human populations on the WorldValueBench dataset.

[Next page](<https://devfeed.tech/topics/uncertainty-quantification-llms.md?cursor=WyIyMDI2LTExLTE2VDA1OjAwOjAwKzAwOjAwIiwgImRhOTgxMzgyLWI2NWQtNDljZC04NGVmLTA5ODAyN2Q0NGUwYyJd>)