LLM Reliability

When language models use test-time sampling, they generate multiple reasoning trajectories and select an answer by majority vote. We show that these trajectories are not independent: for a given question, they concentrate into a small number of clusters, or reasoning basins, each defined by a normalized final answer and the solutions that reach it. A majority vote therefore selects the most stable basin rather than the most accurate one, which creates wrong-majority failures where the correct answer is present but outvoted. We introduce ARBITER, a model-agnostic approach that models interactions between basins using only the base model's own sampled outputs, hidden states, and derived evidence. Most direct correction strategies fail; ARBITER instead uses conservative additive evidence on top of consensus. In its simplest parameter-free form, ARBITER-Δ adds same-model evidence to the majority prior, while ARBITER-Enc augments this with bounded residual signals from hidden states over complete solutions. On GSM8K with Qwen3-4B, consensus over K=24 samples achieves around the mid-94% range, while a same-pool top-2 oracle reaches around the mid-96% range. ARBITER recovers a subset of these cases using zero external information. Across three model families and three math benchmarks, it yields consistent gains with no net-negative cases; for example, on Llama-3.1-8B MMLU-HS-Math, it improves accuracy from the mid-78% range to the mid-82% range, recovering about 22% of the available oracle headroom, indicating that this headroom can be partially recovered from the sample pool itself.

Reliably knowing when a language model is correct is almost as important as being correct. We introduce prover-verifier deliberation (PVD), an inference-time protocol grounded in interactive proof theory, as a mechanism for selective prediction: the protocol produces both an answer and a structured confidence verdict, allowing a system to report high-confidence answers while abstaining on uncertain cases. In each dialogue, a prover defends a candidate answer through checkable sub-claims while a verifier issues targeted challenges and returns \textsc{Accept}, \textsc{Challenge}, or \textsc{Reject}. Because frozen language models are imperfect provers and verifiers operating over a noisy channel, formal soundness and completeness guarantees do not transfer; instead, we characterize the protocol empirically through its coverage-precision behavior. Our main experiment uses Claude Sonnet 4.6 as prover and Claude Haiku 4.5 as verifier on GPQA Diamond. Questions accepted with no answer revision, which we call Accept + No Change (ANC), are reported as the high-confidence subset; we evaluate this subset by its precision and coverage. ANC separates reliable from unreliable answers, yielding a $\sim$30pp HC-Prec gap over the non-ANC complement. Robustness experiments with GPT and Gemini pairings show that high HC-Prec can transfer across model families, while verifier strictness and domain competence largely determine the size of the selection gap. On Humanity's Last Exam, weaker prover-verifier pairings can collapse or invert the ANC signal, illustrating a practical failure mode when the verifier operates outside its effective region. Comparisons with self-consistency, universal self-consistency, multi-agent debate, and Reflexion suggest that prover-verifier deliberation supplies a distinct argument-defensibility signal for selective prediction.

Chains of thought (CoTs) have become central in interpreting and auditing behaviors of large language models. Yet growing evidence suggests that these traces often fail to faithfully represent the computations behind a model's predictions. Several faithfulness metrics have been proposed, but whether they indeed measure faithfulness remains unknown. Answering this requires ground-truth labels, which are hard to obtain since internal computations are not directly observable. Consequently, most works proposing metrics report only absolute scores or comparisons to prior metrics, and the few existing benchmarks rely on proxies like plausibility or importance, properties orthogonal to faithfulness that can mislead about whether a CoT can be trusted. We address this challenge by constructing tasks whose outputs reveal which intermediate computations must have produced them, and developing an automated labeling pipeline that yields ground-truth faithfulness labels at both the step and CoT level. Building on this methodology, we present BonaFide, a benchmark of 3,066 labeled CoTs across 13 tasks and 10 models, and use it to conduct the first systematic evaluation of prominent faithfulness metrics. Our experiments show that most metrics perform near chance, exhibit strong prediction biases and degrade on longer CoTs. The best metric reaches only 0.70 AUROC at the CoT level while another reaches 0.59 at the step level, with neither transferring across settings, while entailing prohibitively high computational cost. Our results expose fundamental gaps in current faithfulness evaluation and call for the development of more reliable and efficient metrics.

Large Language Models (LLMs) are optimized to produce distributionally plausible continuations rather than to explicitly verify whether generated propositions are entailed by source documents. This inductive bias enables generalization, but it does not encode whether responses are grounded with respect to a reference. These issues limit the use of LLMs in domains where strict factual correctness is crucial, such as clinical decision support. Existing hallucination detection approaches improve factuality through retrieval augmentation, self-consistency, or claim verification, but generally do not learn directly over alignment topology. To leverage alignment topology as an inductive bias, we construct aligned bipartite graphs between reference information and LLM outputs and train a graph neural network (GNN) to model alignment structure using message passing. The method achieves state-of-the-art results on four diverse hallucination and question-answering datasets, outperforming all compared methods, including foundational LLMs such as GPT-4o.

Hallucination is often viewed as a direct consequence of missing knowledge: a model answers incorrectly when the correct answer is absent from its generation-time distribution, and correctly when it is present. We test this assumption by introducing a semantic notion of answer availability that aggregates token-level variants expressing the same answer concept, and asks whether the correct concept is already available at the moment the model commits to an answer. Across Qwen and Llama models from 0.8B to 72B in both Instruct and Base variants, 16-47% of Instruct hallucinations occur with substantial probability mass already on the correct concept, and the rate rises monotonically with scale. Comparing such failures against correct generations with matched semantic support, the distinguishing factor is not whether the correct concept is represented, but how its probability is distributed: correct generations concentrate mass on a single surface form, hallucinations disperse it across alternatives. The same sharpening asymmetry extends across multi-token generation and is detectable in pre-generation hidden states. Together, these results identify a single mechanism: instruction tuning sharpens answer commitment with scale, making helpfulness and confident hallucination two consequences of the same underlying disposition.

Machine unlearning aims to remove the influence of specific training data from a model while preserving reliable behavior on the remaining data, making reliable prediction and uncertainty estimation essential for evaluation. Calibration is commonly used as a proxy for reliability in language models, but low calibration error does not necessarily imply reliable decision rules, as models may rely on spurious correlations while remaining well calibrated. We investigate this gap in generative language models using the multiple-choice question-answering evaluation protocol on the TOFU benchmark, measuring probabilistic reliability with calibration metrics (ECE, MCE, Brier) and decision-rule reliability via attribution-based shortcut detection with Integrated Gradients and Local Mutual Information. We find that fine-tuned models achieve low calibration error (ECE ~ 0.04) compared to pretrained models (ECE > 0.5), and models after unlearning retain similarly low calibration despite reduced accuracy on the forget split, while attribution analysis shows increased reliance on correlation-based tokens. These results demonstrate that good calibration can coexist with shortcut-based decision rules after unlearning, extending the reliability paradox to the machine unlearning setting.

Medical large language models (LLMs), including custom medical GPTs (MedGPTs) and open-source models, are increasingly deployed on web platforms to provide clinical guidance. However, they pose risks of hallucination, policy noncompliance, and unsafe design. We conduct a large-scale assessment of 6,233 MedGPTs, evaluating a stratified sample of 1,500, together with 10 open-source LLMs. We introduce two frameworks: MedGPT-HEval for hallucination detection and an LLM-based pipeline for assessing policy violations and developer intent. Our results show that 25-30% of MedGPTs exhibit low factual accuracy, with bottom- and middle-tier models at highest risk; 33.6-54.3% violate operational thresholds, and 57.06% of Action-enabled models lack adequate privacy disclosures. Compared with open-source models, MedGPTs achieve higher factual accuracy and semantic alignment, though open-source models are more stable. These results reveal systemic gaps in hallucination and compliance, highlighting the need for multi-metric evaluation and stronger safeguards. We release HAA-MedGPT, a structured dataset that supports future research on the safety of web-facing medical LLMs.

The maximum softmax probability (MSP) represents a default approach when evaluating uncertainty quantification for language model generation with structured output. Although cheap, it is often miscalibrated. Methods that probe the model's internal activations feed raw hidden states into opaque classifiers, reading activations as static snapshots and leaving implicit the layer-wise trajectory by which a representation is formed. Yet, similar endpoints can arise from very different paths, and how evidence accumulates, reinforces, or reverses across depth might reveal uncertainty that final probabilities obscure. We extract eleven scale-invariant geometric features, tracing the cumulative path of per-layer MLP updates, and feed them to a sparse linear probe. The probe outperforms MSP under selective abstention, with gains scaling with baseline miscalibration up to 21 AURC points. Because every feature has a closed-form geometric meaning, the probe's coefficients trace how and where along depth errors take shape -- which layers commit prematurely, which contradict the running state, where trajectories drift away from their endpoint.

Large language models (LLMs) can enhance factuality via retrieval-augmented generation (RAG), but applying RAG to every query is unnecessary when the model-only answer is reliable. This motivates cascaded RAG: each query is first handled by an LLM-only branch, escalated to a RAG fallback only if the primary branch is uncertain, and abstained from when neither branch is sufficiently trustworthy. However, calibrating such cascades stage by stage may be conservative, since the final utility depends on joint uncertainty thresholding of LLM-only and RAG. In this work, we develop BalanceRAG to certify threshold pairs at a target risk level. Given uncertainty scores from the two branches, BalanceRAG frames each threshold pair as an operating point on a two-dimensional lattice and identifies safe operating points using sequential graphical testing. This enables risk-adaptive threshold calibration, controlling the system-level error rate among accepted points, while retaining more examples. Furthermore, BalanceRAG extends to multi-risk calibration, allowing retrieval usage to be bounded together with the selection-conditioned risk. Experiments on three open-domain question answering (QA) benchmarks across multiple LLM backbones demonstrate that BalanceRAG meets prescribed risk levels, preserves higher coverage and more accepted correct examples, and reduces unnecessary retrieval calls compared with always-on RAG.

Benchmark datasets are critical for reproducible, reliable, and discriminative evaluation of LLMs. However, recent studies reveal that many benchmark datasets are included in pretraining corpora, i.e., $\textit{contaminated}$, which diminishes their value as reliable measures of model generalization. In this paper, we argue that benchmark datasets should be $\textit{contamination-resistant}$, i.e., $\textit{unlearnable}$, but support $\textit{inference}$. To accomplish this, we first highlight the wide prevalence of benchmark dataset contamination and outline the properties of contamination-resistant datasets. Second, we highlight how the asymmetry between the inference and training pipelines in the Transformer architecture can be leveraged to support contamination-resistance. Third, we outline mathematical advancements to make these datasets interoperable across various LLM architectures. Based on the above, we call on the community to ensure the reliability of LLM benchmarking by: (i) advancing novel contamination-resistant methodologies, (ii) developing supporting methods and platforms, and (iii) adopting contamination-resistant benchmarks into existing evaluation pipelines.

Large Language Models have achieved strong performance on reasoning tasks with objective answers by generating step-by-step solutions, but diagnosing where a multi-step reasoning trace might fail remains difficult. Confidence estimation offers a diagnostic signal, yet existing methods are restricted to final answers or require internal model access. In this paper, we introduce Stepwise Confidence Attribution (SCA), a framework for closed-source LLMs that assigns step-level confidence based only on generated reasoning traces. SCA applies the Information Bottleneck principle: steps aligning with consensus structures across correct solutions receive high confidence, while deviations are flagged as potentially erroneous. We propose two complementary methods: (1) NIBS, a non-parametric IB approach measuring consistency without graph structures, and (2) GIBS, a graph-based IB model that learns subgraphs through a differentiable mask to capture logical variability. Extensive experiments on mathematical reasoning and multi-hop question answering show that SCA reliably identifies low-confidence steps strongly correlated with reasoning errors. Moreover, using step-level confidence to guide self-correction improves the correction success rate by up to 13.5% over answer-level feedback.

Uncertainty Quantification (UQ) is widely regarded as the primary safeguard for deploying Large Language Models (LLMs) in high-stakes domains. However, we argue that the field suffers from a category error: mainstream UQ methods for LLMs are just unsupervised clustering algorithms. We demonstrate that most current approaches inherently quantify the internal consistency of the model's generations rather than their external correctness. Consequently, current methods are fundamentally blind to factual reality and fail to detect ``confident hallucinations,'' where models exhibit high confidence in stable but incorrect answers. Therefore, the current UQ methods may create a deceptive sense of safety when deploying the models with uncertainty. In detail, we identify three critical pathologies resulting from this dependence on internal state: a hyperparameter sensitivity crisis that renders deployment unsafe, an internal evaluation cycle that conflates stability with truth, and a fundamental lack of ground truth that forces reliance on unstable proxy metrics to evaluate uncertainty. To resolve this impasse, we advocate for a paradigm shift to UQ and outline a roadmap for the research community to adopt better evaluation metrics and settings, implement mechanism changes for native uncertainty, and anchor verification in objective truth, ensuring that model confidence serves as a reliable proxy for reality.

Large language models (LLMs) hallucinate with confidence: their outputs can be fluent, authoritative, and simply wrong. In medical, legal, and scientific applications this failure causes direct harm, and detecting it from internal model states offers a path to safer deployment. A growing body of work reports that this problem is increasingly tractable, with recent methods achieving high detection performance on widely used benchmarks. We show, however, that much of this apparent progress does not survive scrutiny. Four of the six corpora embed the ground-truth answer directly in the input prompt. A naïve text-similarity baseline we call \textsc{TxTemb} exploits this to achieve near-perfect detection scores without any access to model internals. To measure what genuine detection capability remains once these artifacts are controlled, we conduct a large-scale evaluation spanning twenty-two detection methods, twelve open-source models spanning six architectural families, and six corpora. We further introduce \textbf{DRIFT}, a supervised probe over inter-layer hidden-state transitions, as a point of comparison for live-generation detection. Our findings suggest that the field's reported progress on hallucination detection is substantially explained by benchmark construction artifacts in widely used corpora, and that the majority of established baselines perform near chance under controlled conditions; the consistent exceptions are SAPLMA and DRIFT, both supervised probes on upper-layer hidden states.

Large language models (LLMs) generate not only reasoning text, but also token-level confidence trajectories that record how uncertainty evolves during inference. Whether these trajectories are relevant to reasoning correctness remains unclear. Here we show that confidence trajectories encode a content-agnostic confidence geometry associated with trace-level final-answer correctness. Using only token-level confidence values, without access to the input question, reasoning text, hidden states, or external verifiers, we find that low-dimensional representations of confidence trajectories separate correct from incorrect reasoning traces. Across GSM8K, MATH, and MMLU, this geometric separation is quantitatively linked to downstream predictability: stronger clustering of correct and incorrect traces, measured by the Davies--Bouldin index, consistently corresponds to higher correctness-discrimination AUC. We further show that correctness-related information is enriched in the tail of reasoning, suggesting that late-stage confidence dynamics carry key correctness signals. We propose NeuralConf, a lightweight estimator that learns from confidence trajectories for correctness evaluation. Under a fixed trace budget, NeuralConf-derived scores improve confidence-weighted answer aggregation over majority voting, tail confidence, and other static baselines. These results reveal that LLMs expose trace-intrinsic statistical signals of correctness through their own confidence dynamics, offering a route to improve inference using information already present within generation.

When Large Language Models produce structured outputs such as travel plans, code solutions, or multi-step proofs, individual reasoning steps may appear correct while the output as a whole violates budgets, fails test cases, or contradicts earlier deductions. We propose a decomposed energy function that combines a learned quality scorer with deterministic analytical constraint penalties for verifying structured LLM outputs. The quality scorer is a heterogeneous ensemble of low-rank adapters on a single frozen encoder (3% trainable parameters); the ensemble mean ranks candidates while the standard deviation quantifies epistemic uncertainty, driving a two-pass inference loop that triggers targeted regeneration or abstention. Across five benchmarks (GSM8K, MuSR, TravelPlanner, TACO, Knights & Knaves), our 149M-parameter verifier orchestrating a pool of 7-26B open generators outperforms single-shot Qwen-72B on every benchmark, matches Claude Sonnet 4.6 on MuSR (67.7% vs. 68.0%), and reduces constraint violations by 53% relative to Opus 4.6 on TravelPlanner (oracle 0.028, random 0.231). The two routes are complementary: structural verification wins when constraints are checkable (the verifier captures signal frontier models cannot self-detect), while pretraining-scale priors win where they are not (narrative inference, code semantics). A cross-dataset confounding analysis confirms genuine quality discrimination on four reasoning tasks and identifies a model-identity shortcut on code, mitigated via last-layer retraining. Scorers trained on difficult data transfer zero-shot: a MuSR-trained scorer achieves 93.9% on GSM8K without seeing a math problem.

The use of large language models to assess user states in conversational and adaptive systems is based on the assumption that the metrics used for such assessment are stable and interpretable at the level of individual scores. This paper empirically tests this assumption, focusing on the psychometric reliability of artificial intelligence (AI) measures of user states. This study employed replication evaluation procedures to assess the repeatability of a broad set of metrics across three different bimodal large language models (GPT-4o audio, Gemini 2.0 Flash, Gemini 2.5 Flash). Analyses include both individual score reliability and aggregated reliability, allowing us to distinguish metrics potentially useful for real-time adaptation from those that retain their value only in aggregated analyses. The results demonstrate that metric reliability cannot be considered a default property in interpretive domains. The lack of stability at the level of individual scores precludes the interpretation of such scores as indicators of user state in real-time adaptive systems, even if these metrics demonstrate stability after aggregation. At the same time, the study indicates that individually unstable metrics can retain analytical utility in post-hoc studies, identifying rules governing interactions and their relationships with user experience parameters such as satisfaction, trust, and engagement. The main contribution of this work, besides quantifying the severity of the problem (only 31 of 213 metrics met the criteria), is the proposal of a replicable evaluation framework, enabling measurable evaluations of metric applicability. This approach supports more responsible AI design of adaptive systems, in which the interpretation of results requires explicit validation of reliability and monitoring for violations over time.

As large language models (LLMs) are deployed in consequential settings such as medical question answering and legal reasoning, the ability to estimate when their outputs are likely to be correct is essential for safe and reliable use, requiring well-calibrated uncertainty. Standard reinforcement learning with verifiable rewards (RLVR) trains models with a binary correctness reward that is indifferent to confidence, providing no penalty for confident but wrong predictions and thereby degrading calibration. Recent work addresses this by training models to produce verbalized confidence scores alongside answers and rewarding agreement with correctness. However, verbalized confidence is calibrated at the token level and thus exhibits inconsistency across textual variations with same semantic meaning. We propose \textbf{Calibration with Semantic Reward (CSR)}, a framework that calibrates language models directly in semantic space without a verbalized confidence interface. CSR combines the correctness reward with a novel semantic calibration reward that encourages exploitation among correct rollouts by promoting semantic agreement, and exploration among incorrect ones by discouraging spurious consistency. Experiments across three model families on HotpotQA (in-distribution) and TriviaQA, MSMARCO, and NQ-Open (out-of-distribution) show that CSR consistently achieves lower ECE and higher AUROC than verbalized-confidence baselines across nearly all settings, reducing ECE by up to $40\%$ and improving AUROC by up to $31\%$ over verbalized-confidence baselines, with calibration behavior generalizing robustly across all four evaluation settings.

As large language models (LLMs) are increasingly deployed to perform tasks with minimal human oversight, it is crucial that these models operate robustly. In particular, a model that can solve a given problem should not fail simply because certain entities$\unicode{x2013}$such as names, numbers, or other contextual details$\unicode{x2013}$have changed while the underlying problem logic remains the same. Prior work suggests that current LLMs still struggle with this form of robustness: they often succeed on some variations of a problem but fail on others. However, existing evaluations often lack a systematic way to identify which logic-preserving variations are most likely to induce failure. Instead, they typically test a random subset of allowable variations, which can overstate robustness. To address this gap, we introduce logic-preserving difficulty scaling (LPDS), a framework that (i) quantifies the difficulty of a problem variation and (ii) systematically searches the space of allowable variations to find those that maximize difficulty and expose failures. We show that as difficulty increases, performance declines and errors in the models' reasoning chains become more pronounced. We further demonstrate that LPDS efficiently finds difficult problem variations for a model, resulting in performance drops up to 5 times larger compared to random sampling. Finally, we show that fine-tuning on more difficult variations leads to more consistent robustness gains than training on easier ones.

In a long conversation, an LLM may produce a fluent continuation that rests on premises the conversation has already abandoned. Context-manipulation attacks exploit precisely this weakness. We address this problem with a runtime verifier. An LLM Interpreter maps each utterance to one or more of eight epistemic operations, and then a symbolic engine applies these operations to a dependency map that records what every claim rests on and whether it still stands. Based on the dependency map, checking whether a continuation is grounded then reduces to a walk over the map, linear in its size and requiring no LLM call. Retraction propagates through the same map with a conflict-free guarantee and flags exactly the conclusions that lose support. Our experiments with five QA models demonstrate substantial improvements in QA accuracy at a low cost per query. On ReviseQA for belief revision and MemoryAgentBench's FactConsolidation split (MemAB-FC), the verifier outperforms a retrieval baseline and raises MemAB-FC single-hop accuracy from $0.46$--$0.95$ to $0.94$--$0.98$. With the verifier, even the small 7B model overtakes unaided GPT-4o. When a GPT-4o Interpreter extracts every update from raw text rather than taking the benchmarks' structured updates, the verifier still outperforms the retrieval baseline. Moreover, on MemAB-FC, QA prompts remain compact at $97$--$174$ tokens while the full-context baseline reaches $114.5$K. Retraction queries take less than a microsecond at $2000$ turns.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

Calibration is commonly evaluated by comparing model confidence with its empirical correctness, implicitly treating reliability as a function of the confidence score alone. However, this view can hide substantial structure: models may be systematically overconfident on some kinds of inputs and underconfident on others, causing global reliability diagnostics to obscure localised calibration failures. To address this, we formulate the problem of discovering hidden miscalibration regimes without assuming access to predefined data slices. We define the corresponding miscalibration field and propose a diagnostic framework for estimating it. Our approach learns a calibration-aware representation of the input space and estimates signed local miscalibration by kernel smoothing in the learned geometry. Across four real-world LLM benchmarks and twelve LLMs, we find that input-dependent calibration heterogeneity is prevalent. We further show that the discovered fields are actionable: they support local confidence correction and reduce calibration error in systematically miscalibrated regions where confidence-based methods such as isotonic regression and temperature scaling are less effective.

We present a framework for verifying the deterministic structured computations surrounding a large language model rather than the model itself, extending a Lean 4 trust-boundary architecture to the generic interfaces of modern LLM pipelines. Certificate validity is a Lean 4 kernel type-check plus a sorry-free transitive axiom audit against the trusted set {propext, Classical.choice, Quot.sound}; other assumptions are declared and partitioned by tier (mathematical placeholders, cryptographic assumptions, ML/human oracles). The technical contribution comprises three local certificate families and two operators. The families are conflict-aware bilattice grounding (with an emission-gate soundness lemma), embedding sensitivity and paraphrase stability, and Hoare-style agent action. The operators are a Maximal Certifiable Residue, which turns abstention into the maximum-weight certifiable residue with audit-logged dropped claims, and a Compositional Stability theorem, which yields a closed-form pipeline-wide perturbation budget from per-layer gains and margins. The three families plus a Universal Assurance Card consolidator form the per-call deliverable for high-stakes deployments: patent and legal retrieval, regulated finance, clinical decision support, and agentic systems with irreversible side effects. A compiled Lean 4 reference artifact (Lean v4.30.0-rc2, Mathlib) covers all 22 certificate types, with 17 of 46 kernel-audited declarations axiom-free, the rest depending only on the trusted set and declared assumptions, and zero uses of sorryAx or Lean.ofReduceBool. The three families are empirically tested through four registered pilots: bilattice grounding on adversarially perturbed HotpotQA, embedding sensitivity in short- and long-form settings, and Hoare-style agent action on a filesystem sandbox with adversarial prompt injection.

Large language models (LLMs) are increasingly deployed in settings where the available context is incomplete or degraded. We argue that an LLM generating answers under incomplete context can be viewed as an implicit imputer, and evaluated against a criterion from the multiple imputation (MI) literature: uncertainty should scale with the amount of missing information. We assess this criterion on SQuAD, using a controlled framework in which context availability is varied across five levels. We evaluate two answer-level uncertainty measures that can be estimated from repeated sampling: sampling-based confidence (empirical mode frequency) and response entropy. Confidence fails to reflect increasing missingness: it remains high even as accuracy collapses. Entropy, by contrast, increases with context removal, consistent with the MI analogy, and explains substantially more variance in accuracy than confidence across all evidence levels (quadratic $R^2$ gap up to 0.057). We further introduce a black-box diagnostic $ρ_R(α)$ that estimates the proportion of baseline uncertainty resolved by context level $α$, requiring only repeated sampling with and without context. These results suggest that entropy is a more responsive black-box uncertainty measure than confidence under incomplete context.

Scaling laws predict loss from compute but not how capabilities interact. We measure the coupling between reasoning and truthfulness across 63 base models from 16 families and find a regime change invisible to loss curves: below a family-dependent critical scale N_c, capabilities anticorrelate (r = -0.989, p = 4 x 10^{-5} nonparametric permutation test); above it, they cooperate. N_c ~ 3.5B parameters [2.9B, 13.4B] (bootstrap 95% CI), but model size is not the only variable that determines phase. Architecture, data curation, and training recipe each shift N_c independently: curated training eliminated the coupling dip between Qwen generations (0.025 to 0.830 at matched scale), Gemma-4 at 4B achieves coupling 0.871, characteristic of 13B+ standard-trained models, through distillation and architectural innovation, and Phi at 1B matches web-trained coupling at 10B through data curation alone. Width normalization eliminates the anticorrelation across all tested families, supporting an output-projection bottleneck. Internally, 38 of 40 models show zero competing attention heads. A sparse-regression ODE cross-predicts held-out Llama-2 at 5.6% error. The diagnostic requires no model internals -- only public benchmark scores across a model family. The cooperative regime extends to the frontier (r = +0.72, 34 models, 10 labs). A proof-of-concept intervention confirms the bottleneck is exploitable: adding a single truth-direction vector at the identified layer corrects 60% of misaligned outputs in the tax phase with zero retraining -- a surgical, per-inference correction that requires no weight modification. Code, data, an open-source steering CLI for any open-weight model, and an interactive dashboard for phase diagnosis are released: https://zehenlabs.com/cape/.

Large language models (LLMs) achieve strong performance across many tasks but remain vulnerable to hallucinations, making it important to systematically evaluate their reliability under realistic adversarial inputs. We formulate hallucination elicitation as a constrained optimization problem, where the goal is to find semantically coherent adversarial prompts that are equivalent to benign user prompts. Existing attack methods remain limited: discrete prompt-based attacks preserve semantic equivalence and coherence but search only over a limited set of prompt variations, while continuous latent-space attacks explore a richer space but often decode into prompts that are no longer valid rephrasings. To address these limitations, we propose REALISTA, a realistic latent-space attack framework. REALISTA constructs an input-dependent dictionary of valid editing directions, each corresponding to a semantically equivalent and coherent rephrasing, and optimizes continuous combinations of these directions in latent space. This design combines the optimization flexibility of continuous attacks with the semantic realism of discrete rephrasing-based attacks. Experiments demonstrate that REALISTA achieves superior or comparable performance to state-of-the-art realistic attacks on open-source LLMs and, crucially, succeeds in attacking large reasoning models under free-form response settings, where prior realistic attacks fail. Code is available at https://github.com/Buyun-Liang/REALISTA.

Large Language Models (LLMs) achieve strong performance on logical reasoning benchmarks, yet their reliability remains uncertain. Existing evaluations rely on static benchmarks, which fail to assess robustness under logically equivalent transformations and often overestimate reasoning capability. We propose LGMT (Logic-Grounded Metamorphic Testing), an oracle-free framework that leverages first-order logic (FOL) to evaluate LLM reasoning. By deriving metamorphic relations from formal logical equivalences, LGMT constructs semantically invariant test cases and detects reasoning defects through cross-case consistency checking. Experiments on six state-of-the-art LLMs show that LGMT exposes substantial hidden defects missed by traditional reference-based evaluations. We further find that models are particularly sensitive to symbol-level and conclusion-level variations, and that advanced prompting such as Few-shot CoT only partially mitigates these issues. These results suggest that LLM evaluation should move beyond isolated correctness toward robustness under logical invariance. LGMT provides a principled and scalable approach for diagnosing reasoning failures.

Large language models (LLMs) often produce answers with high certainty even when they are incorrect, making reliable confidence estimation essential for deployment in real-world scenarios. Verbalized confidence, where models explicitly state their confidence in natural language, provides a flexible and user-facing uncertainty signal that can be applied even when token logits are unavailable. However, existing verbalized-confidence methods often optimize answer generation and confidence generation jointly, which can cause confidence-alignment objectives to interfere with answer accuracy. In this work, we propose a decoupled and order-aware framework for verbalized confidence calibration. Our method first generates an answer and then estimates confidence conditioned on the fixed question--answer pair, allowing confidence optimization without directly perturbing the answer-generation process. To align confidence with correctness likelihood, we construct a sampling-based surrogate from multiple model completions and optimize rank-based reinforcement learning objectives that encourage responses with higher estimated correctness likelihood to receive higher verbalized confidence. Experiments on reasoning and knowledge-intensive benchmarks show that our method improves calibration and failure prediction performance while largely preserving answer accuracy. These results demonstrate that verbalized confidence can be more reliably aligned by decoupling confidence estimation from answer generation and optimizing the relative ordering of confidence across responses.

Monitoring coding agents for dangerous behavior using language models requires classifying transcripts that often exceed 500K tokens, but prior agent monitoring benchmarks rarely contain transcripts longer than 100K tokens. We show that when used as classifiers, current frontier models fail to notice dangerous actions more often in longer transcripts. In particular, on a dataset that requires identifying when a coding agent takes a subtly dangerous action, Opus 4.6, GPT 5.4, and Gemini 3.1 miss these actions $2\times$ to $30\times$ more often when they occur after 800K tokens of benign activity than when they occur on their own. We also show that these weaknesses can be partially mitigated with prompting techniques such as periodic reminders throughout the transcript and may be mitigated further with better post-training. Monitor evaluations that do not consider long-context degradation are likely overestimating monitor performance.

Large Language Models (LLMs) achieve strong performance on standard knowledge evaluation benchmarks, yet recent work shows that their knowledge capabilities remain brittle under question variants that test the same knowledge in different forms. Robustness augmentation of existing knowledge evaluation benchmarks is therefore necessary, but current LLM-assisted generate-then-verify pipelines are costly and difficult to scale due to low-yield variant generation and unreliable variant verification. We propose SAGE (Scalable Automated Generation of Robustness BEnchmarks), a framework for scalable robustness augmentation of knowledge evaluation benchmarks using fine-tuned smaller models. SAGE consists of VariantQual, a rubric-based verifier trained on human-labeled seed data, and VariantGen, a variant generator initialized with supervised fine-tuning and further optimized with reinforcement learning using VariantQual as the reward model. Experiments on HellaSwag show that SAGE constructs a large-scale robustness-augmented benchmark with quality comparable to the human-annotated HellaSwag-Pro at substantially lower cost, while the fine-tuned models further generalize to MMLU without benchmark-specific fine-tuning.

Large language models (LLMs) are increasingly used in social science as scalable measurement tools for converting unstructured text into variables that can enter standard empirical designs. Measurement validity demands more than high average accuracy, which requires well calibrated confidence that faithfully reflects the empirical probability of each measurement being correct. This paper studies the model miscalibration in LLM-based social science measurement. We begin with a case study on FOMC and show that confidence based filtering can change downstream regression estimates when LLM confidence is miscalibrated. We then audit calibration across 14 social science constructs covering both proprietary models, including GPT-5-mini, DeepSeek-V3.2, and open source models. Across tasks and model families, reported confidence is poorly aligned with tolerance-based correctness. As a simple mitigation, we propose a soft label distillation pipeline for calibrating Bert with LLM. The method converts an LLM score and its verbalized confidence into a soft target distribution, then trains a smaller discriminative classifier on encoder models for these targets. Averaged across datasets, this approach reduces ECE by 43.2% and Brier by 34.0%. These results suggest that LLM-based social science pipelines should treat calibration as part of measurement validity, rather than as an optional post-processing concern.