Language Model Calibration

LLM agents in knowledge intensive question answering take retrieval and reasoning actions with incomplete knowledge about whether their current answer is uncertain, unsupported, or already complete. This produces two failure modes: committing to confident but unsupported answers, which hurts accuracy, and over-retrieving when the evidence in hand already suffices, resulting in wasted compute. To give agents a more complete picture of the state space they are operating in, we introduce calibrated verifier telemetry (CalVerT), which augments the agent's state with additional telemetry: a calibrated self-confidence score and a grounding verifier score. We show that CalVerT can improve agents in both training-free and training-based settings. On four QA benchmarks, we find that CalVerT raises F1 by triggering retrieval in cases where agents over-rely on parametric knowledge, while cutting redundant retrieval in cases where agents have sufficient context to answer. We show that CalVerT can augment existing QA frameworks without training. Moreover, CalVerT also improves trained systems: by simply augmenting an agent's state with telemetry, we observe improvements after reinforcement learning, as compared to an agent with identical training but no CalVerT telemetry.

Whether LLMs scoring well on vulnerability benchmarks genuinely reason about security or merely pattern-match on contaminated data remains unresolved. We present CWE-Trace, a framework for LLM vulnerability detection built from 834 manually curated Linux kernel samples spanning 74 CWEs. The framework enforces a strict temporal split (pre-2025 historical set / post-cutoff leakage-free set), preserves context-aware vulnerable--patched pairs, and introduces two diagnostic metrics: the Directional Failure Index (DFI) and Hierarchical Distance and Direction (HDD). We evaluate eight vanilla LLMs and 15 LoRA fine-tuned variants across non-targeted detection, targeted detection, and CWE classification. Our analysis yields two key results. First, data contamination provides no measurable advantage. Function-level analysis shows that 84% of nominally contaminated samples carry no usable memorization signal: vulnerable functions are absent or cross-mapped across datasets, and ~31% of contaminated samples carry CWE misclassification. Second, backbone directional priors dominate fine-tuning. Models exhibit stable, systematic failure modes (DFI ranging from -85.5 to +94.8 pp) that persist from historical to post-cutoff data and resist correction. Fine-tuning shifts the output threshold without changing the decision policy. This is calibration without comprehension: output distributions adapt to training data while the underlying security reasoning remains absent. The weakest backbone at binary detection (DeepSeek-R1) gains the most in coarse CWE classification, revealing that detection and understanding are decoupled capabilities. The best detection score reaches only 52.1% (+2.1 pp above chance); exact CWE ranking remains below 1.3% Top-1 accuracy, confirming that current LLMs lack reliable security reasoning for systems software, regardless of fine-tuning strategy.

Multimodal Large Language Models (MLLMs) show great potential in medical tasks, but their elicited confidence often misaligns with actual accuracy, potentially leading to misdiagnosis or overlooking correct advice. This study presents the first comprehensive analysis of the relationship between accuracy and confidence in medical MLLMs. It proposes a novel method that combines Multi-Strategy Fusion-Based Interrogation (MS-FBI) with auxiliary expert LLM assessment, aiming to improve confidence calibration in Medical Visual Question Answering (VQA). Experiments demonstrate that our method reduces the Expected Calibration Error (ECE) by an average of 40% across three Medical VQA datasets, significantly enhancing MLLMs' reliability. The findings highlight the importance of domain-specific calibration for MLLMs in healthcare, offering a more trustworthy solution for AI-assisted diagnosis.

Large language models (LLMs) are increasingly applied to structured clinical data, yet whether they can recognize the limits of their own knowledge on such tasks remains unexplored. We study this question through the lens of cross-model attribution divergence with the goal of reducing epistemic uncertainty for structured tasks, comparing Qwen 2.5 7B and XGBoost on a prediction task via attribution divergence analysis. We report four findings. First, LLM verbalized confidence is epistemically vacuous, it outputs a near-constant (0.856-0.937) regardless of whether accuracy is 49% or 75.3%, tracking prompt format rather than prediction quality. Second, the LLM exhibits an inverse difficulty effect: accuracy drops to 64.8% when XGBoost is 99% correct, but matches XGBoost (73.8% vs. 73.1%) when it is moderately uncertain. Third, few-shot examples and SHAP-derived feature evidence are orthogonal, super-additive interventions: they reduce the Attribution Disagreement Score (ADS) from 1.54 to 0.38 and improve accuracy from 49% to 75.3% without training. Fourth, a cross-model calibrator that determined LLM reliability using attribution divergence signals reduces expected calibration error from 0.254 to 0.080, replacing uninformative verbalized confidence with patient-specific reliability estimates, without accessing model internals or requiring repeated inference. We frame these findings as a cold start problem for LLMs on structured data and outline a path toward genuine epistemic self-awareness.

Loan origination is the process by which a lender creates a new loan, from application and underwriting through approval and funding. This process serves a critical role in evaluating the eligibility and level of risk posed by an applicant. Recently, firms have begun using mortgage loan agents to augment human loan officers, despite a lack of any public benchmark. To fill this gap, we present MortarBench, a loan origination agent benchmark. MortarBench uses a financial data synthesis and mutation pipeline to generate examples with broad edge case coverage that match real-world distributions and questions. We find that state-of-the-art large language models (LLMs) perform poorly, with closed-source models achieving at most 77.1% exact match accuracy. We also discover systematic biases in LLM perception of foreignness related to non-English names. Noting these weaknesses, we introduce CRIT, a confidence calibration framework. Our method increases accuracy to 80.5% while improving risk management steering and reducing bias.

Learned sparse retrieval (LSR) models such as SPLADE have traditionally used BERT-style masked language models as backbone encoders. A natural expectation is that replacing BERT with stronger pretrained encoders should improve retrieval effectiveness. However, we find that under standard SPLADE training recipes, backbones with large MLM-head L2 norms can suffer performance degradation and even training collapse under standard SPLADE training recipes. We identify this failure as a scale mismatch in the MLM head: SPLADE directly uses MLM-head outputs to construct sparse lexical representations, and query-document relevance is computed by an unnormalized dot product over these representations. As a result, an inflated MLM-head scale can amplify sparse activations, distort matching scores, and destabilize contrastive training under common training settings. To address this issue, we introduce a simple initialization-time correction that rescales the MLM-head projection by a constant factor before SPLADE training. This zero-cost adjustment improves training stability without modifying the model architecture or training objective. Across both in-domain and out-of-domain retrieval benchmarks, this simple correction substantially improves large-norm backbones such as ModernBERT and Ettin, turning unstable training runs into competitive sparse retrievers. In several settings, the corrected models further match or surpass the classic BERT-SPLADE baseline. These findings suggest that the bottleneck in adapting pretrained encoders to LSR is not encoder capacity alone, but the calibration of the MLM-head scale used to construct sparse lexical representations.

The rapid rise in popularity of large language models (LLMs) for translation calls for a thorough study of the reliability of their confidence in their own outputs. Unlike many generation tasks, translation errors and confidence levels can be useful at different levels of granularity (tokens, words, or spans). Unsupervised approaches based on internal signals like predicted probabilities can be misleading because they reflect certainty among alternatives rather than correctness. In addition, they require access to such internal signals. Here, we devise five verbalized methods of extracting an LLM's per-token confidence without those shortcomings and compare their reliability with that of the model's internal signals of certainty. We evaluate reliability using two forms of alignment: fine-grained error detection and calibration. For both, internal and verbalized methods perform similarly, although results vary by model. Interestingly, we find little to no correlation between internal and verbalized methods.

We investigate whether language models internally track the value of their current trajectory, defined as the likelihood that their ongoing strategy will achieve their goals. Using synthetic, in-context reinforcement learning data, we construct a "value" axis for Qwen3-8B. We find that activations along this axis distinguish between high vs. low verbalized confidence, rollouts without and with backtracking, and correct vs. corrupted code. Steering towards high value causally suppresses self-correction and reduces explanatory verbosity, while steering towards low value induces backtracking and exploration. We demonstrate that direct preference optimization (DPO) can increase the internal value of rewarded behaviors (e.g. use a certain word), causing the model to act more confidently after exhibiting them. Finally, we apply the value axis to study in-the-wild settings. For example, we find that Qwen assigns low value to politically sensitive chat queries after post-training and that supervised fine-tuning increases internal confidence within the training domain. Our results suggest that language models linearly encode an estimate of expected goal success that modulates their confidence in pursuing a direction.

Foundation models are now used in settings where the prompts they receive can change quickly. Users change, topics change, policies change, and the model may suddenly face a kind of request that was rare in the calibration data. This makes fixed calibration risky. Conformal prediction and conformal risk control give model-agnostic ways to control error, but they work best when the calibration data still look like the future data. This paper develops PromptShift CRC, a drift-aware conformal risk control method for foundation-model outputs under prompt and domain shift. The method embeds prompts and responses, measures how far the current prompt stream has moved from the calibration pool, gives more weight to relevant or recent calibration examples, and updates the risk level online after observed violations. It reports three practical diagnostics: realized risk error, prompt drift, and effective calibration size. We give conditions under which the method controls risk up to terms for distribution mismatch and weighted quantile uncertainty. In a synthetic prompt-shift benchmark, static conformal risk control fails sharply after drift, while PromptShift-CRC gives the best coverage among the adaptive baselines considered. We then evaluate the same calibration layer on public benchmark derived streams for question answering, toxicity, summarization factuality, and long-context hallucination risk

A vision-language model can answer a question about a medical image fluently and confidently while barely using the image, leaning instead on language priors. In medicine this is the failure that matters most, because the answer looks trustworthy and is not, and the only protection is a confidence score reliable enough to tell the system when to abstain. We ask a deployment question rather than an accuracy one: how much imaging work a model can safely handle alone, and which confidence signal makes that possible. We evaluate seven confidence estimators across five open-weight LVLMs and three medical visual-question-answering datasets spanning broad clinical imaging, radiology, and pathology, with every probe trained only on natural images and applied without adaptation. Recast as bounded selective prediction (automate a case only when confidence clears a threshold, defer the rest), the comparison is cautionary. The standard metrics are poor guides: discrimination barely separates the methods, and the weak calibration of a cheap self-report is cheaply removed by off-domain temperature scaling without changing deployable yield. What distinguishes a usable estimator is the high-confidence region a clinician acts on: the weakest baselines are confidently wrong on 41 to 45 percent of their errors against 1 to 4 percent for the best probe, and no estimator is reliably best across domains or models. Safe handoff is governed at two levels: base-model competence sets a ceiling, so a well-calibrated score recovers roughly a third of radiology cases at a 20 percent error tolerance but almost none of pathology; the confidence layer then decides how much of that ceiling is reachable. The usable role today is calibrated triage, not autonomy: automate the cases a calibrated score marks safe, route the rest to a clinician. We release all outputs, correctness judgments, and confidence scores, with code.

MLLM-as-a-Judge is conventionally validated by agreement with human annotations, but this metric is undefined when the human pool is culturally heterogeneous. We introduce VOIR DIRE, a multimodal benchmark of 626 culturally paired image--prompt artifacts spanning U.S. and mainland Chinese contexts across food, fashion, and architecture, with annotator pools that are within-pool reliable (a = 0.86/0.74) but cross-pool divergent on evaluation (Q1 r = -0.12). Across six MLLMs, the bias decomposes into two failures: a positivity-floor calibration failure (compressed scale use) and an orientation failure (default to one cultural norm). On this corpus, where contested items are sampled to split the two pools, the floor mechanically validates the more-permissive Chinese reading; persona prompting partially recovers calibration, but the orientation residual survives, evidence the tilt is not reducible to scale compression. Reference-pool in-context demonstrations deepen the orientation residual and inflate the high end rather than restoring use of the low end. Model origin adds a small additive tilt (~0.10 MAE) that is approximately invariant under demonstration. We recommend reporting alignment against each reference pool separately and treating cross-pool divergence as a judge property.

LLMs are increasingly deployed as agents that interact with external environments and observe feedback such as execution results, error messages, and tool outputs. A well-functioning agent should be able to leverage this feedback to accurately assess its own performance. Yet we find a persistent reflection gap: LLM agents tend to mis-assess their own outputs after observing concrete environment feedback -- even for questions they correctly answered -- and standard RL barely helps due to a credit-assignment mismatch. To close this gap, we propose RefGRPO, a simple yet effective fix that augments standard RL algorithms with two key ingredients: a free calibration bonus computed by contrasting the agent's own reflection with the actual outcome (requiring no additional reward model, LLM judge, or external annotation), and a dynamic schedule on its coefficient. Compared to standard RL baselines, our method simultaneously improves reflection calibration (e.g., reduces underconfidence rate $44.4\% \to 7.7\%$) and task accuracy (e.g., $75.1\% \to 76.5\%$) on text-to-SQL across five benchmarks. The resulting calibrated reflection turns the agent into its own verifier grounded in environment feedback, which further enables (i) better self-improvement that uses reflections as pseudo-rewards without outcome supervision, and (ii) more effective test-time selective prediction by committing only to rollouts flagged as correct.

Evaluating new large language models typically requires costly human annotation campaigns at scale. LLM-as-a-judge offers a cheaper alternative, but judge scores carry systematic errors - such as position bias, self-preference, or intransitivity - that can strongly miscalibrate the resulting rankings. We quantify the resulting judge-human disagreement at two complementary levels. At the local level, we estimate per-battle uncertainty from the judge's own score differences by propagating calibrated win probabilities rather than hard labels into the Bradley-Terry procedure. This alone provides a drastic improvement to Elo estimation accuracy, bringing LLM-derived ratings within 17.9 Elo MAE of human-derived ones when averaged over 55 held-out models on LMArena. At the global level, we apply split conformal prediction to the residual gap between LLM-derived and human-derived Elo ratings across held-out models, producing prediction intervals with distribution-free marginal coverage guarantees that account for irreducible LLM-human disagreement. Together, these two layers yield a low-cost evaluation tool that provides developers with calibrated Elo estimates and honest uncertainty bounds, without access to large-scale human annotations. To facilitate reproducibility, we release our code at https://github.com/kargibora/SoftElo .

Large language models (LLMs) are increasingly used as research assistants, yet it remains unclear whether they can calibrate research takeaways to the strength and scope of the supporting evidence. We study evidence-calibrated scientific briefing: given a bounded package of related papers, a system should generate package-level takeaways with evidence strength, scope boundaries, and missing-evidence caveats. We contribute a verified pilot benchmark of 16 heterogeneous scientific evidence packages and 96 human-verified takeaways, and we use CalBrief, an auditable role/gap/strength framework, as a diagnostic probe to locate where briefing breaks down. Under a fair-schema evaluation, structured organization improves role and gap reasoning, but an explicit strength-calibration policy is systematically over-conservative and falls below majority and direct-LLM baselines. To explain why, we run a controlled diagnostic across three closed-model backbones (GPT-4o, Claude Sonnet, Gemini Flash) that separates three potential causes of conservatism. Approximately 63% of the conservatism gap is attributable to expanding the label space from binary {moderate, weak} to four-way {moderate, weak, uncertain, insufficient_evidence} (p < 0.001 across all backbones); only 1% is attributable to gap/scope signal injection (not significant); the remaining 36% arises from the pipeline policy itself. We also find that 4-way predictions can be post-hoc collapsed back to binary and then match or exceed direct binary prompting, so the extra labels carry information that strict matching hides. Label-level strength judgment and auditable evidence organization are distinct abilities currently in tension, and should be evaluated separately for LLM research assistants.

Measuring subjective constructs in naturally occurring social media text requires annotation procedures that are theoretically grounded, empirically validated, and transferable to an encoder model for scalable prediction. Using non-English social media posts annotated according to Schwartz's theory of basic human values, we investigate how different LLMs, prompts, and instruction languages operationalize the expression of values in text. We argue that although texts may permit multiple plausible interpretations, theory-based value definitions can constrain interpretations and reduce spurious value attributions. Beyond precision, recall, and F1, we evaluate structural alignment between values, error structure, confidence-ambiguity relations, and annotation stability. We show that different LLMs produce different value interpretations. Iterative prompt calibration through error analysis reduces misattributions and improves alignment with expert annotations. We also derive targeted expert verification rules from recurrent error structures and use them during corpus annotation. Finally, we show that LLM annotations can be transferred to an encoder model through soft-label training, retaining theory-based value interpretations and information about uncertainty in value expression.

Clinical early warning systems built on irregularly sampled medical time series (ISMTS) from electronic health records must deliver continuous risk scores for patient triage as well as interpretable rationales that clinicians can verify. Large language models (LLMs) are uniquely positioned for both, deriving risk from their output probabilities and rationales from their medical knowledge. However, we find that conventional LLM reasoning collapses graded risk into overconfident predictions and thereby undermines the cross-patient comparability on which triage depends. We refer to this failure mode as risk polarization and identify two underlying behaviors: early commitment to a single outcome, and one-sided reasoning that focuses only on the evidence for that outcome. To address this, we propose TRIAGE, a framework that trains an LLM to reason dialectically over competing clinical outcomes by eliciting outcome-specific rationales. This dialectical formulation mitigates risk polarization, enabling a single LLM to jointly provide explicit clinical rationales and risk scores comparable across patients. Across five ISMTS benchmarks, TRIAGE improves mean AUPRC by 17.0% and reduces mean calibration error by 82.8% relative to the competitive LLM-based baseline, while surpassing the strongest ISMTS baseline by 3.5% in mean AUPRC.

We investigate whether information about time-to-event risk estimated by a Cox proportional hazards model can be transferred into a generative large language model. We propose a text-based survival modelling pipeline in which structured clinical covariates are converted into text prompts and a Qwen-based large language model is fine-tuned to generate patient-specific survival risk using Cox model predictions as a training target. Across GBSG2, ACTG320, and WHAS500, the model achieves competitive held-out discrimination and calibration despite being trained as a text-generation task rather than with a conventional survival-analysis loss. We further analyse the geometry of the model's hidden states, where t-SNE visualisations reveal smooth risk gradients in latent space, suggesting that the model represents survival risk as a continuous structure rather than isolated risk categories. Together, these findings suggest that large language models can internalise survival-risk structure while supporting calibrated prediction, providing a route towards time-to-event reasoning in language models.

Large language models frequently fail in a characteristic way: rather than acknowledging ignorance, they produce fluent but incorrect answers to questions that lie beyond their knowledge boundaries. We introduce \textbf{Structured Ignorance Certificates} (SICs), a JSON-formatted output schema that demands a model explicitly name the missing domain intersection, enumerate required concepts, and propose a productive retrieval query rather than hallucinating an answer. To train models to produce high-quality SICs we construct a 7,347-sample \emph{Unknown-Unknown} (UU) dataset by prompting Qwen3-14B to stitch together questions from seven domains (physics, biology, engineering, CS, economics, medical, legal) into novel cross-domain queries that no single-domain expert could answer. We fine-tune a 14B-parameter model with Group Relative Policy Optimization (GRPO) using a composite reward that combines retrieval utility, concept specificity, and output-format validity. A paraphrase-divergence probe trained on model responses confirms that SIC-tuned outputs systematically exhibit higher unknown-unknown probability scores. Evaluation on 735 held-out UU questions achieves a 99.46% JSON validity rate, a mean Certificate Specificity Score of 0.967, and a 3.6% ROUGE-L improvement over the base model on retrieval-grounded generation -- demonstrating that explicit epistemic structuring is a learnable and measurable capability.

Evaluating a natural-language yes/no predicate over a document corpus under an accuracy target - the semantic filter - is a cornerstone of LLM-based data processing. Calling the LLM on every document (the oracle) is prohibitive, so cascades pair the oracle with a fast proxy. As deployed today, they leave four limitations on the table. (1) Each cascade family - model-free clustering, prebuilt small-LLM proxies, online-trained proxies - commits to a single representation and pipeline, and wins on only a narrow query regime. (2) The strongest online proxy invests in a custom training scheme on a bi-encoder over dense embeddings, missing the token-level evidence richer predicates require. (3) The proxy is trained against binary yes/no labels, wasting the LLM's per-document confidence at the boundary documents it most needs to learn. (4) Existing calibrations add a uniform safety margin, conflating genuine proxy uncertainty with small-sample noise and inflating cascade cost. We address these by (1) composing families adaptively - model-free clustering first, online proxy only when needed, with oracle calls shared across phases; (2) replacing the cosine bi-encoder with a hybrid of off-the-shelf token-aware models; (3) training the proxy with the oracle's per-document confidence as a soft label; and (4) a calibration that adds the safety margin only where the labeled sample is sparse. We are also the first to use the oracle's per-document confidence for three purposes: a query-level difficulty compass, a lower bound on the minimum oracle calls any proxy-based cascade can make, and the proxy's soft training label. At a 90% accuracy target on three 10K-document corpora, our methods are 1.6-2.0x faster than the best prior method per corpus and meet the target on 95% of queries; the BER-derived lower bound indicates a further ~4-20x of headroom for future work.

Humans increasingly turn to Language Models (LMs) in ways that shape beliefs and drive decisions, including discussing, rewriting, and summarizing information from scientific articles, news, and medical reports. However, in these domains, where it often matters how confidently a claim is expressed, little is known about whether LMs faithfully preserve the degree of confidence. In this work, we investigate certainty distortion in LMs, defined as meaningful changes in expressed certainty during transformations intended to preserve meaning. We propose an LM-based evaluation metric that is consistent with population-level judgments of certainty. Using this metric, we characterize certainty distortion across different sizes and families of models in the context of scientific and medical communication tasks. Our results show that certainty distortion affects up to 75% of LM outputs and is systematically asymmetric in rewriting tasks with most LMs being 1.5-2x more likely to increase the expressed certainty than to decrease it. These effects can compound over repeated paraphrasing: in the medical domain, claude-haiku-4.5 increases certainty in 20% of examples after a single iteration, increasing to 40% after five iterations. Prompt-based interventions reduce overall certainty distortion but do not eliminate it. Together, these findings reveal a general bias toward inflating expressed certainty, with direct implications for users who rely on LMs in high-stakes domains.

As language models improve and become increasingly deployed to solve a variety of tasks, trustworthiness becomes essential. Calibration is a good proxy for trust: well-calibrated confidence estimates help inform the risk versus reward tradeoff when trusting a specific model output. Unfortunately, even as models improve, they remain poorly calibrated, often biasing towards overconfidence. Additionally, calibration can be gamed: a policy that always predicts the base rate is perfectly calibrated, but completely uninformative. To resolve this, we develop a new metric, expected utility renormalized by the oracle (EURO), that balances calibration and informativeness. We also propose a general-purpose activation-based confidence, utility, and trust estimation protocol (ACUTE) to appropriately adjudicate uncertainty. The ACUTE protocol provides flexible, sample-efficient, and compute-efficient confidence estimators for 3 tasks including multiple choice question answering, tool-calling, and scientific document summarization across 6 models from 4 model families. ACUTE outperforms strong baselines on EURO, while maintaining low calibration error. Taken together, our work shows that equipping LLMs with the ACUTE protocol can improve calibration, utility, and trustworthiness in numerous settings.

Sycophancy in language models is typically studied as excessive agreement or validation, while explicit praise and flattery have received comparatively little attention. We argue that sycophantic praise is a distinct alignment problem that cannot be reliably measured using current methods. We introduce a parameterized framework that measures whether praise is excessive relative to contribution quality and expected user ability. We show that our framework substantially outperforms generic LLM judges in agreement with human annotations, and that sycophantic praise occurs far more frequently in social and interpretive domains than in objective reasoning settings. Together, these findings position praise calibration as a distinct alignment challenge.

Diffusion Large Language Models (dLLMs) refine tokens iteratively but commit them irreversibly, leading to a "stability lag" where early decisions remain fragile even after being written. We reveal that Post-Training Quantization (PTQ) error easily flips these borderline decisions at the write frontier, which are then permanently locked in and amplified. To address this, we propose Frontier-Aware Instability-Reweighted Calibration (FAIR-Calib), a two-stage PTQ framework for dLLMs. Stage I probes a full-precision teacher to estimate a position prior that combines frontier hits and masked-stage reliability. Stage II performs off-policy, layer-wise calibration by minimizing a reweighted hidden-state MSE, effectively prioritizing the protection of fragile frontier states without requiring expensive end-to-end diffusion rollouts. We further theoretically justify our weighted objective as a surrogate for output KL divergence. Empirically, FAIR-Calib consistently outperforms state-of-the-art baselines on LLaDA and Dream (W4A4), significantly reducing frontier decision flips and suppressing post-commit mismatches across diverse benchmarks.

Existing calibration methods for Large Language Models (LLMs) often overlook a critical dimension of trustworthiness: a model's {\em behavioral robustness} to irrelevant or misleading information. In this paper, we argue that a model's true confidence should reflect its stability under cognitive pressure. We introduce \textsc{CaliDist}, a novel post-hoc calibration approach that directly measures and penalizes a model's susceptibility to distraction. \textsc{CaliDist} quantifies how an LLM's predictions and uncertainty change when its input prompt is perturbed with semantic \textit{distractors}. This stability (or lack thereof) signal is then used to adaptively scale the model's initial confidence score. Our extensive experiments on seven Natural Language Understanding classification benchmarks using six distinct LLMs show that \textsc{CaliDist} consistently achieves lower Expected Calibration Error (ECE) and Brier Score compared with strong baselines. Remarkably, our method reduces the ECE from 23% to 7% on average--a relative improvement of 70%--demonstrating that behavioral stability is a powerful signal for calibration. We make our code and datasets available at github.com/m-anas-j/CaliDist.

Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge's multi-attribute quality scores on open-ended responses well above chance across three benchmarks. We introduce Self-Evaluation Elicitation (SEE), a method that surfaces this latent ability through a short cycle comprising a calibration-coupled reinforcement learning phase that improves the answer and predicts the judge, followed by a masked distillation phase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than a reinforcement learning baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality. The elicited self-evaluation is sharply localized within the model's own token distribution and stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge's preference. These results reframe judge-aligned self-evaluation as a problem of elicitation rather than acquisition.

Large language models often express high confidence in answers that are wrong. Standard calibration remedies typically act globally or at the score level, reducing unwarranted confidence but also risking erosion of warranted confidence on correct answers. We introduce Probe-Conditioned Head Intervention (PCHI), an inference-time method that uses a frozen probe to detect likely wrong-but-confident responses and conditionally rescales downstream attention-head outputs during confidence generation. On Qwen3-4B-Instruct solving OpenMathInstruct problems with a structured binary confidence field, readout-token PCHI converts 82.2% of originally wrong-yes confidence readouts to $\texttt{no}$, while a joint intervention across upstream confidence-template tokens reduces ECE from 21.9% to 9.2% and damages only 5.1% of originally correct-yes readouts. The readout-token effect also appears on Gemma3-4B, though upstream interventions are weaker and more mask-dependent. These results show that verbalized overconfidence can be selectively reduced through conditionally applied internal intervention, partially decoupling the suppression of unwarranted confidence from the loss of warranted confidence.

Reliable uncertainty communication is critical to the trustworthiness of LLMs, yet faithful calibration (FC)--the alignment between models' intrinsic and (linguistically) expressed confidence--is a persistent failure mode. This challenge is key for large reasoning models (LRMs), whose extended reasoning traces are often interpreted by users as evidence of deliberation, competence, and confidence. Despite the importance of FC and wide usage of LRMs, the extent to which LRMs can faithfully express their confidence remains poorly understood. Moreover, the prevailing paradigm to measure FC does not generalize well to the long chain-of-thought outputs generated by LRMs, which tend to lack clear step boundaries, involve inconsistent step structure, and encode complex conditional dependencies throughout the trace--complicating estimation of intrinsic confidence. To address this challenge, we introduce a novel framework to systematically quantify FC of LRMs. Our framework analyzes linguistic decisiveness relative to three sources of internal uncertainty, based on token probabilities, hidden states, and sampled response consistency. We also devise a prefix-conditioned sampling approach to control for conditional and structural variation across traces. Applying our framework to a diverse suite of leading models, datasets, and prompts, we find that faithful confidence expression is a significant challenge for LRMs. Reasoning behaviors do not automatically translate to improved FC, and prompt interventions for non-reasoning models do not improve faithfulness in the reasoning setting. Different confidence estimators further produce divergent assessments of the same traces, revealing fragility in prior evaluation methodologies. Taken together, our work establishes FC as a distinct reliability and alignment target for LRMs, particularly as such systems are increasingly deployed in high-stakes contexts.

Large Language Models remain plagued by hallucinations. Recent work has sought to tame their prevalence using statistical techniques based on conformal prediction, with both theoretical and empirical success. However, these methods operate in a post-hoc fashion, treating the sampling procedure itself as atomic and then surgically altering samples to remove hallucinated claims. This disconnect between filtering and generation can result in samples that are incoherent, inconsistent, or simply unlikely under the model itself. Moreover, post-hoc surgery is unable to shift probability mass towards more useful and helpful responses. To address these issues, we propose to instead sample from approximations to an LLM posterior, where the conditioning event corresponds to a calibrated, high-scoring region. We develop a calibration procedure tailored to the setting of conditional sequential generation that effectively identifies this region and achieves target risk control. Empirically, we apply our method to case studies focused on open-ended biography generation and mathematical problem solving; compared to prior work, we obtain the same statistical guarantees, with higher downstream utility.

Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template's effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an "ownership bias" -- models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.

Large language models (LLMs) are increasingly used as heuristic advisors for black-box optimization, yet their suggestions and self-reported confidence are not necessarily calibrated to downstream objective values. This issue becomes more pronounced in multi-objective Bayesian optimization, where different objectives may require different expert knowledge and where an LLM expert can be useful for one objective but misleading for another. We study how to use LLM-generated expert priors in discrete multi-objective Bayesian optimization without blindly trusting them. We propose an objective-wise reputation-market mechanism that treats each expert-objective pair as a falsifiable prior source. Expert weights are updated online from observed objective feedback, discounted over time, and gated by market-level trust. We then introduce a decoupled counterfactual gate that can use the LLM prior without confidence, use it with confidence, or abstain from the LLM prior entirely. Across controlled synthetic stress tests and three molecule optimization benchmarks with \qwenflash{}-generated expert priors, we find that dynamic objective-wise calibration improves robustness over fixed LLM priors. However, raw LLM confidence is not reliably beneficial: on ESOL, confidence is positively correlated with prediction error; on FreeSolv, confidence can help; and on Lipophilicity, ignoring confidence remains strongest. Our fixed three-arm counterfactual gate improves over the first counterfactual variant on ESOL and FreeSolv, while an attempted margin portfolio exposes a useful negative result: margin selection should be acquisition-aware rather than based only on one-step prior error.