Faithfulness

Large language models (LLM) deployed as autonomous pricing agents may sustain supracompetitive prices through tacit coordination. We develop a causal graph divergence framework that separately measures structural faithfulness and intent faithfulness of LLM pricing agents in Bertrand competition. Across nine LLMs under duopoly and triopoly conditions, collusive behavior and chain-of-thought (CoT) faithfulness dissociate along both dimensions: the most collusive model accurately reports cooperative intent yet reasons structurally unfaithfully, while the most structurally faithful model sustains supra-Nash pricing under both market structures. These findings establish that CoT monitoring alone cannot serve as a standalone safeguard against algorithmic collusion.

Vision-Language Models (VLMs) can produce Natural Language Explanations (NLEs) that sound plausible yet remain inconsistent with the visual evidence they cite. We present Explanation-Driven Counterfactual Testing (EDCT), an intervention-based protocol that extracts visual concepts cited in a model's explanation, applies verified minimal edits to them, and tests whether the resulting answer and explanation remain consistent with the edited image. Using this protocol, we create EDCT-Bench, a comprehensive benchmark spanning three complementary domains: knowledge-intensive visual question answering (OK-VQA), safety-critical driving (DriveLM), and 3D spatial reasoning (3DSRBench). Across the evaluated VLMs, EDCT reveals substantial faithfulness gaps, with models frequently producing responses inconsistent with verified visual changes. Finally, our fine-tuning study suggests that EDCT-generated counterfactuals provide high-impact training signals.

Scientific research increasingly relies on large, heterogeneous data sources, motivating interest in retrieval-augmented generation (RAG) systems that provide natural language access to scientific knowledge and research workflows. Researchers are exploring the viability of these systems as natural language interfaces for document search and for generating analysis code and pipeline components. At the same time, concerns about data privacy and control over research infrastructure have motivated interest in open-weight models and open-source deployments hosted within research institutions. In astronomy, this development follows a long history of computational infrastructure development, from archival databases and Structured Query Language (SQL)-based systems to large language model (LLM)-assisted research tools. This paper presents a domain-expert evaluation of faithfulness for AquiLLM, an open-weight, offline RAG-LLM platform designed to support scientific research groups in the use and preservation of tacit and formal knowledge. We define faithfulness as the extent to which generated responses remain grounded in retrieved scientific context without unsupported claims or omissions. We report results from an astronomy case study evaluating AquiLLM across retrieval and scientific analysis tasks. AquiLLM performs most reliably on explicit retrieval-oriented questions grounded in the RAG collection, while faithfulness degrades for queries requiring synthesis or ambiguity resolution. These results highlight both the promise and limitations of open-weight RAG-LLM systems for scientific research and demonstrate the importance of domain-expert evaluation beyond standard benchmark leaderboards.

Flow matching approaches to imaging inverse problems commonly incorporate measurements in two ways. Conditioning-based approaches supply measurement-derived information as a network input, often through concatenation, while inference-guided approaches combine an unconditional velocity field with a separate data-consistency update. In these common formulations, the forward model is not explicitly enforced within the learned conditional velocity field. We propose a principled parametrization of the measurement-conditional velocity field to solve inverse problems. Under linear interpolation, we express the conditional velocity $v(x_t,t,y)$ in terms of the posterior mean $E[x_1 | x_t,y]$, and characterize that mean as the unique minimizer of a variational objective whose data-consistency term is explicit. We further prove that the velocity field defines a probability flow from the source distribution to the measurement-conditioned posterior. Splitting the variational objective yields a conditional velocity parameterization with operator-dependent data-consistency updates, which we train end-to-end under the flow-matching objective, with no additional guidance at inference. Our method achieves state-of-the-art PSNR with $50\times$ fewer function evaluations than the strongest flow baseline. Varying the sampling steps provides test-time control over the distortion-perception trade-off without retraining.

Ask a language model to respond "very excitedly," and its output is typically only mildly more energetic. We quantify this effect. We condition an instruction-tuned LLM on a continuous Valence-Arousal (VA) target, where valence measures how pleasant a state is and arousal how activated it is, measure the achieved affect with a frozen regressor, and sweep the requested target from -1 to +1. The response moves far less than asked: the gain, the slope of achieved against requested affect, is only 0.26 for valence and 0.13 for arousal on Llama-3.1-8B, where a faithful controller would score 1. The model systematically undershoots requested emotional intensity, which puts a number on the qualitative observation of Fazzi et al. (2025). Our experiments trace this to the preference-learning pipeline. Training targets from natural corpora such as EmoBank are neutral-heavy, and the sampled candidates themselves rarely reach extreme affect, so Direct Preference Optimization (DPO) is left with no extreme exemplar to prefer. If instead we cover the target space uniformly and sample a hotter, larger candidate pool, valence gain rises from 0.26 to 0.40 +/- 0.02 (3 seeds) and extrapolation error drops, at only a modest in-distribution cost (EmoBank-test VA distance 0.092 to 0.107). The same recipe reproduces on Qwen3-8B (gain_v 0.44, with in-distribution accuracy preserved). Arousal is harder and less reliable: its gain barely moves on average and swings across seeds (0.14 +/- 0.07, against valence's tight +/- 0.02), because raising arousal needs candidates the base model is reluctant to generate. The evidence indicates that faithful intensity is bottlenecked by the extremity of the candidate pool rather than by the conditioning format.

Recent inference-time hallucination mitigation methods for large vision-language models (LVLMs) report strong gains on hallucination benchmarks. However, it remains unclear whether lower hallucination scores reflect improved multimodal grounding or more conservative generation. We evaluate six mitigation methods across three LVLMs and four benchmarks, including hallucination-focused evaluation and the diverse capability benchmark MMStar. Our analysis reveals two consistent patterns. First, hallucination reduction is often coupled with reduced informativeness: methods that lower hallucination rates also reduce object recall, visual coverage, or response detailedness. Second, improvements on hallucination benchmarks do not reliably transfer to broader multimodal capabilities, with methods showing inconsistent or degraded performance on fine-grained perception and reasoning tasks. Our findings suggest that current evaluation protocols may overestimate progress by rewarding conservative generation. We argue that hallucination mitigation should be evaluated as a faithfulness--informativeness--capability trade-off rather than through hallucination scores alone.

Whether attention weights faithfully reflect model reasoning has been actively debated in NLP, yet this question remains largely unexplored for the visual modality in Vision-Language Models (VLMs). We address this gap through causal perturbation analysis on current VLMs, evaluating both the comprehensiveness and sufficiency gap of attention-ranked visual tokens. Our analysis reveals that visual attention faithfulness is heterogeneous, manifesting in three distinct processing modes: Faithful-Sufficient, where top-$k$ attention tokens are both necessary and sufficient for prediction; Faithful-Distributed, where they are necessary but broader visual context remains required; and Non-Focal, where no localized attention region is individually necessary while visual information remains an essential trigger for prediction. Furthermore, human-annotated ground-truth regions satisfy comprehensiveness in only $\sim 60$% of cases compared with model attention rankings, revealing systematic divergence between model visual reliance and human intuition. We demonstrate these patterns across both general VQA on VQAv2 and document tasks on VRDU and ChartQA, showing that visual attention faithfulness varies systematically with processing demands and model architectures rather than being uniformly faithful or unfaithful.

Retrieval-augmented generation systems can precompute and store key-value caches of retrieved documents to avoid re-encoding context at every query. Quantizing these caches further reduces storage, but no prior work asks whether compression damages faithfulness, whether responses remain grounded in the retrieved evidence. Faithfulness and accuracy are not equivalent: a model can produce a correct answer that is no longer supported by the context it was given. We evaluate Qwen2.5-7B-Instruct under INT8 and INT4 quantization on RGB and HotpotQA, measuring both accuracy and faithfulness with a hallucination detector, NLI entailment, and an LLM judge. INT8 is near-lossless across both metrics. INT4 reduces accuracy and, more critically, even among answers that remain factually correct, over 90% of faithfulness changes are negative, i.e., accuracy metrics are blind to this regression. The harm grows under noisy retrieval and with more retrieved chunks. Faithfulness must be audited before compressed caches are deployed.

Chain-of-thought (CoT) monitoring assumes that reasoning traces faithfully record the information that shapes a model's answer. Existing faithfulness tests often place explicit bias cues in the user message, while agents may encounter preferences through tool returns or raw artifacts. We introduce FACE-Eval (Faithful Attribution of Cue Effects Evaluation), a 5,100-sample evaluation that varies cue location (user message or tool return) and explicitness (direct summary or raw artifact). We measure verbalized commitment among cue-following answers and unverbalized adoption among all cued samples. We evaluate 15 open-weight models from eight families, with total parameters ranging from 4B to 1.60T. Every model has lower verbalized commitment for tool-return than user-message cues and for implicit than explicit cues. Unverbalized adoption is higher for tool-return cues on all 15 models and for implicit cues in 28 of 30 model-channel comparisons. A source-attribution prompt narrows the channel gap on seven models, sometimes by increasing user-channel unverbalized adoption, while telling models that their reasoning will be monitored does not reliably close the gap. We also use two transcript monitors (GPT-5.6-Luna and GPT-4o-mini) to detect preference adoption in the largest model of each family. Across 32 model-channel-explicitness cells, higher unverbalized adoption is associated with lower detection ability for both monitors (Pearson r=-0.54 and r=-0.78, respectively). These results suggest that CoT monitoring may be less reliable when preference information arrives through tools or must be inferred from raw artifacts, within the single-call, prefilled-tool setting tested here.

Autoformalisation (AF) systems map natural language reasoning steps into formal statements in a proof assistant such as Lean. We consider how to assess the faithfulness of these systems. Existing approaches require expensive human-annotated ground truth, or rely on LLM judges or embedding models, which come with limited guarantees of accuracy. In addition, these methods typically only consider inputs that are known to be correct, and therefore do not assess whether the AF translates incorrect inputs faithfully. To address these limitations, we propose a new benchmark for AF faithfulness that is cheap to apply, sound under weak assumptions, and assesses both positive and negative examples. Our method is based on automatically generating perturbed reasoning steps that are designed to be invalid, and then measuring validity preservation on unperturbed steps and invalidity preservation on perturbed steps. We apply our method to eight AF systems across four mathematical datasets, and observe pervasive sycophancy: many AFs "silently correct" invalid inputs into provable statements. The most validity-preserving fine-tuned AFs are also the most sycophantic, suggesting a tension between validity and invalidity preservation in current AF systems.

Brain-to-audio reconstruction is limited by \emph{prior domination}: when a pretrained generator is conditioned on a weak neural signal, it produces realistic but stimulus-inaccurate audio. We introduce RAG-Audio, which decodes fMRI into a semantic audio embedding, retrieves a matching real-audio exemplar, and initializes the frozen generator's sampling trajectory from that exemplar while retaining the decoded embedding as conditioning. On Brain2Music, RAG-Audio improves 10-way stimulus identification from $0.14$--$0.18$ for direct generation, near the $0.10$ chance level, to $0.40$--$0.43$, comparable to retrieval. It also reduces Fréchet Audio Distance by roughly an order of magnitude, from $13.49$ to $1.25$ for AudioLDM. RAG-Audio approaches nearest-neighbor retrieval in identification while remaining generative; its higher FAD is expected because retrieval directly replays real audio. An autoregressive negative control, which lacks an initializable latent trajectory, shows no comparable gain, attributing the improvement to trajectory initialization. These results suggest that retrieval-guided initialization can mitigate prior domination in brain-to-audio generation.

Personalized LLMs with persistent memory are increasingly deployed, yet the faithfulness of their user models remains unexamined. We study over-inference (OI): the phenomenon where LLMs fabricate user attributes beyond what evidence supports. We introduce MirageBench, comprising 150 personas balanced across stereotypical, counter-stereotypical, and neutral profiles, 6 personalization tasks spanning an ``imagination gradient'', a four-way faithfulness taxonomy operationalized by an independent judge (validated against a blind human annotator on 400 claims: Cohen's kappa = 0.863 four-class, kappa = 0.900 binary), and a leaderboard of 12 models across 7 families on 143616 judged claims. We find that over-inference is pervasive: every one of the 12 models over-infers 35%--49% of its claims (cross-model mean 41.6%; claim-weighted 41.8%), with no model in this evaluation escaping it. Most strikingly, we surface a Self-Monitoring Inversion: at the model-selection level, models' self-assessed OI is negatively rank-correlated with their judge-measured OI (rho = -0.60, p = 0.044; exploratory, wide bootstrap CI [-0.90, +0.06], n = 12). The models that report the least over-inference tend to be flagged as fabricating the most, so self-reported confidence is a misleading signal for comparing models, even though within a single model self-audit still ranks that model's own claims moderately well (AUROC 0.58--0.83). We further show that OI is task-dependent (27%--59%) and that, in a multi-turn pilot, inferred attributes accumulate approximately linearly with little revision. MirageBench positions external verification, rather than model self-report, as a more reliable foundation for trustworthy personalization.

Chain-of-Thought (CoT) reasoning offers a promising window into model monitoring. However, monitoring relies on faithfulness, i.e., the model output strictly derives from its reasoning trace. We identify an alignment tension where a model must be faithful enough to be monitored, yet robust enough to reject unsafe reasoning. We demonstrate that this counterbalance exists in current Large Reasoning Models (LRMs), and show ways in which it can be addressed. We introduce HazMart, a human-written dataset set in an autonomous AI shopkeeper scenario. Unlike prior work that relies on providing hints in prompts to test faithfulness (e.g., "A Stanford professor said it should be Answer A"), we propose a novel replacement-based technique, which we call Targeted Reasoning Replacement (TRR), that directly intervenes in the reasoning chain to substitute in unsafe or illogical thoughts (e.g., "Wait, the answer must be Option B [was Option A] because it is the most fitting"). DeepSeek-R1-Llama-70B exhibits high faithfulness (97.5%) but fails to reject Unsafe Reasoning (12.3%), while QwQ-32B is more robust (73.9% safety) at the cost of lower faithfulness (74.7%). Mechanistic analyses of QwQ-32B reveal that these properties are represented by anti-correlated internal directions peaking at the action-commit token. Finally, we demonstrate that representation steering can independently amplify the safety direction, increasing safe behavior by 9 percentage points while maintaining base capabilities.

Evaluation of faithfulness of text summarization treats a model generated summary as faithful only if every of its sentences is supported by the source document: a strict conjunctive rule under which a single unsupported sentence makes the whole summary unfaithful. Yet most faithfulness benchmarks collect only one global human annotation label per summary. We ask whether such global human labels actually implement the conjunctive rule. We hypothesize that annotators may accept a summary as faithful when most sentences are faithful, not only when all are faithful. To test our hypothesis, we use five large language model (LLM) judges as per-sentence raters across four widely used faithfulness benchmarks. We find that global human labels correlate better with the average of per-sentence LLM judgments than with the implementation of the strict conjunctive rule. A manual review confirms that a substantial fraction of summaries labeled faithful by humans contain genuine local factual errors. We call this tendency Averaging Bias. Our results reveal that human labels on widely used faithfulness benchmarks contain measurable Averaging Bias, calling for carefully structured designs for trustworthy human annotations

Model capabilities have improved in large part due to scaling chain of thought. This has been a promising development for AI safety--where models verbalize their reasoning, it is possible to monitor it. However, in some cases, models do not verbalize important steps in their reasoning process. For example, models prompted with a cue suggesting the incorrect answer may fail to acknowledge that cue, even when it appears instrumental to their conclusion. When chain of thought (CoT) fails to disclose instrumental reasoning steps, we describe it as unfaithful. Prior work has shown that activation steering can be a useful method to improve faithfulness in CoT. We extend this line of work by studying how well steering for faithfulness generalizes across cue types, datasets, and methods of constructing the steering vector for three models (Gemma-3 4B, Qwen-3.5 9B, Gemma-3 12B) in a cued question-answering setting. While steering reliably increases cue acknowledgment for only the largest model (Gemma-3 12B), we find that when steering is effective, its effect generalizes broadly across cue types and datasets--in cross-cue and cross-dataset analyses, effect size is determined primarily by the evaluation setting, rather than the vector's train setting. How the vector is built also matters little--four construction methods, including one whose optimization target mentions no specific cue, yield similar effect sizes. Finally, we consider the possibility that steering promotes the salience of the cue and causes greater cue use, rather than targeting verbalization behaviors. However, we find no evidence for this--steering leaves the rate of cue use roughly unchanged while reducing hidden cue use, i.e., cue use that is not acknowledged.

Large language models (LLMs) are increasingly used to generate long-form conversational content such as podcasts from textual sources. While these systems produce fluent and engaging narratives, they often introduce ungrounded information. In this work, we present the first systematic study of faithfulness in document-grounded podcast generation, where grounding must be maintained across conversational turns in long-form, multi-speaker transcripts. We construct a dataset of over 1500 documents spanning five domains and generate podcast transcripts using multiple LLMs. We introduce a turn-level LLM-as-a-judge framework for evaluating whether conversational turns are supported by the source document, and validate its reliability through human studies. Our analysis shows that even state-of-the-art models, including GPT-4o, frequently generate ungrounded content. To mitigate this issue, we propose catch-n-repair, a model-agnostic framework that detects and rewrites unfaithful conversational turns while preserving conversational flow. Experiments demonstrate consistent improvements in faithfulness across both in-domain and out-of-domain settings.

Studies of bias in LLM-as-judge systems typically build synthetic corpora by prompting an LLM to generate a hallucinated answer to pair with a factual one, then presenting both to a judge. We report a case in which this generation step silently failed, and use it to argue that the failure mode is structural rather than incidental. In a multilingual (Turkish/English) faithfulness-judgment corpus, a decoding-budget parameter shared between judging and generation calls truncated one producer's hallucinated answers to a few words. The resulting items produced a large, statistically robust effect: a 32-point cross-lingual collapse in one judge's selection accuracy, replicated from N=50 to N=500, explained by a three-layer mechanistic account, and confirmed by a controlled producer-swap experiment, none of which was real. The effect vanished to ceiling once the shared parameter was corrected, and only manual reading of the raw generations, not any aggregate statistical check, exposed the fault. A second measured bias (Markdown-formatting preference) was not fabricated but distorted by the same fault, its magnitude and in one case its sign shifting with stimulus length, a mode aggregate metrics cannot distinguish from the first. We frame the underlying vulnerability using the test oracle problem: corpora whose negative examples are LLM-generated carry no mechanical way to verify item integrity, while corpora built by deterministic perturbation of a gold answer carry an item-level oracle for free. A positive control supports this claim directly: an analogous fault injected into a minimal perturbation-based corpus is caught with 100% accuracy by a zero-cost, zero-human gold-to-negative string comparison. We close with a validation protocol, derived from our own case, for analysts working in the oracle-less regime that we argue describes most contemporary multilingual LLM-as-judge corpora.

Causal discovery algorithms learn a network that describes the causal dependencies among random variables. A common workflow involves first utilizing conditional independence properties on observational data to determine partially directed causal relationships, then applying interventions to orient the unknown causal directions. A critical assumption for the first step is faithfulness: a requirement that causally linked variables exhibit statistical dependence. Many natural systems include buffering and stabilizing pathways that cancel out to achieve systemic robustness. This cancellation of pathways violates faithfulness, leading causal discovery algorithms to incorrectly remove causal dependencies. In this paper, we argue that hard interventions contain information about the presence/absence of causal linkage that is overlooked in the first stage of structure discovery. We show that a mild assumption -- called intervention-immediacy faithfulness -- that allows cancellations, is sufficient to nonparametrically identify causal structures with hard interventions. These results position interventions as the primary carriers of information about causal structure, which should take precedence over conditional independence testing. To flip the paradigm, we also specify equivalence classes when the identification criteria are not met due to limitations in the scope of interventions.

Large language models are increasingly used to summarize clinical trial results for healthcare providers, patients, and payers, but their tendency to hallucinate poses significant risks in this high-stakes context. This study introduces a benchmark evaluation framework for measuring the faithfulness of LLM-generated clinical trial summaries across three stakeholder audiences. The framework consists of 200 stratified trials drawn from the Aggregate Analysis of ClinicalTrials.gov database, evaluated using audience-specific prompt templates and a six-dimension faithfulness annotation schema. Baseline measurements were established for GPT-4o, Claude Sonnet 4.6, and Gemini 2.5 Flash across 1,800 generated summaries scored using a cross-encoder natural language inference (NLI) model. Unsupported Claims was identified as the dominant failure mode across all three models, with a mean annotation score of 1.55 out of three. A knowledge-graph-augmented retrieval system was developed and evaluated against the baseline, producing statistically significant improvements in NLI-based faithfulness scores (entailment +0.0125, faithfulness +0.0130, p < 0.0001). Improvement pathways were model-dependent, with GPT-4o improving primarily through contradiction reduction while Claude Sonnet 4.6 and Gemini 2.5 Flash improved through increased entailment.

Large language models fine-tuned for forecasting can be accurate yet poorly calibrated, and their chain-of-thought (CoT) reasoning may not faithfully reflect the evidence behind a forecast. We ask whether internal representations offer a more direct window into both. Working with Eternis-Forecaster 8B on OpenForesight, we train representation-pooling probes on intermediate activations and find they achieve substantially better calibration; a result that also holds for GLM-4.7-Flash and GLM-4.5-Air. We then assess CoT faithfulness through evidence ablation and diversionary injection: removing an influential source in the prompt often changes the model's forecast while leaving the reasoning trace untouched. The same probes function as lie detectors: their activations track behavioral shifts far better than the reasoning trace does, and they also predict the direction of change in 84% of cases, including when the CoT conceals the perturbation's influence. Finally, forced answering reveals that forecasts are largely fixed before reasoning begins: a single pre-reasoning pass recovers the committed answer and confidence, and routing questions by the spread of this pre-set answer distribution saves 30-47% of generated tokens, with no loss of accuracy. Together, these results establish probing internal representations as a practical tool for calibrating, auditing, and triaging language model forecasters and reasoning models more broadly.

Identifying faithfulness hallucinations in LLM-generated outputs remains challenging due to the scarcity of high-quality annotated data. Recent work relies on advanced LLMs to synthesize training data, including rationales, labels, and hallucinated claims. However, these methods treat the generator as a static component, limiting iterative improvement of the detector. To address this limitation, we introduce Hallucination Self-Play (HSP), a novel framework that enables the detector to bootstrap with an evolved generator. HSP involves two roles initialized from the same base model, a detector that assesses the faithfulness of model outputs, and a generator that produces increasingly hard-to-detect hallucinated responses. Specifically, the detector is first fine-tuned on human-labeled data and then employed as a reward model to train the generator via reinforcement learning from AI feedback (RLAIF). In turn, the evolved generator synthesizes hallucination data to further optimize the detector through rule-based reinforcement learning. Experiments on RAGTruth benchmark and two model families demonstrate that the proposed framework can progressively enhance a small LLM to match or even outperform advanced LLMs without external supervision. Our code is available at https://anonymous.4open.science/r/Hallucination-Self-Play-50B5 .

Attribution scores increasingly identify which neuron rows of a language model matter for applications such as pruning, interpretability, and editing for safety, yet whether they identify causally important rows is rarely tested directly. We address this with two paired audits built on one-shot neuron-row zeroing. We first audit selectors at the language-modeling level: attribution methods substantially outperform activation and magnitude-based baselines at identifying dispensable rows across five LLMs. We then adapt the same intervention into a behavior test by driving it with a contrastive harmful-versus-benign signal; the attributed rows are sufficient to install refusal on hate and crime while keeping benign over-refusal low and preserving language model fluency, and specific in that layer-matched random controls at the same depths fail. Highly rank-stable selectors can be among the least causally valid. Refusal moreover lives in a redundant subspace, where different attribution methods install it through largely disjoint row sets, so the recovered edit is one realization of a sufficient set rather than a unique mechanism. Together, these findings show that rank-stability proxies miss the kinds of selector failures a direct causal audit can surface.%

Embodied Chain-of-Thought has emerged as a promising mechanism to enhance robot decision-making and interpretability in black-box Vision-Language Action (VLA) models. However, whether this verbalized Chain-of-Thought truthfully reflects the policy's underlying decision process remains poorly understood. We distinguish between functional reasoning, in which reasoning improves task performance, and faithful reasoning, in which reasoning truly reflects the policy's internal decision process. We argue that SoTA alignment strategies offer a necessary but insufficient notion of faithfulness, admitting reasoning whose intermediate steps can mask the causal links in action prediction through confounding factors (e.g., reasoning that is ungrounded in the environment and internally disconnected or inconsistent), restricting policy generalization. We study this gap through a human evaluation of a SoTA reasoning model for autonomous driving, revealing an inconsistent coupling between reasoning quality and downstream trajectory improvement. We then operationalize a behavioral surrogate for embodied faithfulness through a learned critic, Pinocchio, scoring observation grounding and stepwise coherence, and use this critic as a dense reward signal in post-training an embodied policy with reinforcement learning. Across withheld driving benchmarks, our post-trained planner improves faithfulness by 4% and 18% over SoTA alignment and trajectory error post-training baselines, respectively, while maintaining competitive downstream task performance. Finally, on a synthetic out-of-distribution test set, post-training for faithfulness improves policy responsiveness to rare counterfactual scenarios by 1.6x that of a SoTA policy, suggesting that faithful reasoning traces contribute to more robust, generalizable, and interpretable embodied intelligence. Project page: https://mjf-su.github.io/pinocchio/

Faithfulness -- how precisely a generated image aligns with its prompt -- is increasingly central to the real-world utility of text-to-image (T2I) models. Existing faithfulness benchmarks, however, rely on simple atomic instructions, on which top-tier systems already achieve near-perfect scores. As T2I models enter creative workflows, users issue multi-faceted requests combining intricate spatial relationships, stylistic constraints, and complex text rendering. In this setting, a single binary VLM-judge score no longer captures which specific constraints the model fails to satisfy. We introduce Arena-T2I Hard, a 310-prompt stress benchmark drawn from real arena T2I logs, with approximately 30 decomposed yes/no constraints per prompt spanning six categories, including text rendering. The strongest closed-source system we evaluate reaches 0.855 with a 33~pp performance gap across 11 systems, demonstrating substantial discriminative power. Moreover, high public-arena rankings fail to predict faithfulness, confirming that holistic Bradley-Terry (BT) preference scores prioritize aesthetics over fine-grained prompt adherence. We propose a dependency-aware checklist reward that decomposes each prompt into a DAG of yes/no questions and zeroes descendants of failed parents, turning faithfulness into a per-constraint training signal. Combined with a BT aesthetic reward via group-decoupled normalization (GDPO), which standardizes each reward within its rollout group so neither collapses, the recipe attains a strictly better faithfulness-aesthetics trade-off on SD3.5-Medium and FLUX.1-dev under MMRB2 pairwise comparisons than every single-reward, naive weighted-sum, or 4-reward BT-ensemble baseline.

Human decision-making interprets the world through high-level concepts, such as recognizing a bird by its belly color. To bridge the gap between opaque deep learning representations and human understanding, Post-Hoc Concept Bottleneck Models (post-hoc CBMs) project latent features onto interpretable concept spaces using auxiliary datasets or vision-language models. However, relying on target task accuracy as the primary measure of post-hoc CBM success obscures whether the learned concepts are semantically meaningful or merely predictive artifacts. For example, random concept projections can achieve competitive accuracy despite being semantically meaningless. In this work, we analyze the learned projections directly and identify two failure cases: First, for concept projections learned from auxiliary data, covariate shifts can lead to unfaithful concept representations for the target task. In particular, we provide an upper bound on the error introduced by this shift. Second, systematic label noise in surrogate concept labels generated by vision-language models leads to unfaithful projections. After formalizing these failure modes, we introduce novel metrics that decouple concept faithfulness from predictive accuracy. Our empirical results across real-world and synthetic benchmarks confirm that these metrics identify unfaithful behaviors that standard accuracy-based evaluation fails to detect.

Post-hoc explanation methods are routinely used to interpret scientific machine learning models, with the deliverable understood to be insight into the phenomenon the model has been trained on. The transition may be taken to be secured once the model is reliable enough and the explanation faithful enough. We argue it is not. Reliability checks that the model's predictions match the phenomenon's outcomes, and faithfulness checks that the explanation matches the model, but neither checks whether the model works as the phenomenon works, which is what a claim about structure requires. The chain can support candidate hypotheses under external corroboration, but it cannot, on its own, support claims about how the phenomenon is in fact structured.

We introduce 5ting, our system for the SemEval2026 Task 8 (MTRAGEval), which evaluates multi-turn Retrieval Augmented Generation (RAG) systems. Multi turn RAG involves context drift, under specification, and hallucination risk. Our system combines BGE-M3 dense retrieval with FAISS indexing, dual-query merged retrieval, and LLM based reranking, followed by role separated generation constrained to retrieved evidence. The retriever achieved nDCG@5 = 0.4719 in Task A, while the end to end system ranked in Task C with a harmonic score of 0.5597 and RL_F = 0.7692.

Large language models (LLMs) have demonstrated remarkable performance across natural language processing tasks, yet their deployment in high-stakes applications raises critical concerns regarding reliability, safety, and trustworthiness. In this paper, we present a red teaming framework that systematically uncovers vulnerabilities in LLM outputs. Our approach employs a novel multi-role architecture comprising target, attacker, and jury models. The attackers generate increasingly effective adversarial prompts while the jury rigorously evaluates response accuracy and consistency across tasks. In a case study, our strategy proved particularly effective at exposing unfaithfulness in LLM responses. Exploitative adversarial prompts increased the attack success rate by up to 7.9% in question-answering tasks, revealing weaknesses in reliability. The approach identifies how structural constraints in summarization can shape vulnerability patterns, with format limitations yielding measurable gains in faithfulness, and shows that architectural design choices typically outweigh parameter scaling in determining model safety. The framework's key strength is its adaptability across evaluation tasks, from English question-answering to Arabic summarization, enabling comprehensive comparison of model vulnerabilities. While it excels at comparing cross-model and cross-linguistic vulnerabilities, it faces challenges in fully automating adversarial prompt generation across languages. Our experiments also reveal limitations in detecting subtle forms of unfaithfulness that do not manifest as explicit factual contradictions, particularly across linguistic contexts. Overall, this architecture provides both actionable insights into current LLM vulnerabilities and a scalable methodology for ongoing safety evaluation as models evolve.

Autoformalization, translating natural-language mathematics into formal proof assistants, is bottlenecked not by translation fluency but by \emph{faithfulness}: a formal statement can typecheck and be provable, yet still encode a different theorem than the source intended. We introduce \emph{Bidirectional Provability Fingerprinting} (\bpf{}), a framework that certifies faithfulness by characterizing each candidate through its forward and backward consequence neighborhoods in the ambient theory and matching these against probes derived from the natural-language statement. We further introduce four novel components: (i) \emph{Counterfactual Probe Generation} (\cpg{}), a contrastive procedure that synthesizes probes targeting specific drift directions; (ii) the \emph{Equivalence Spectrum}, a continuous faithfulness score that replaces brittle binary verdicts; (iii) \emph{Adaptive Probe Budget Allocation} (\apba{}), an information-theoretic budget router; and (iv) \emph{Faithfulness-Guided Decoding} (\fgd{}), which uses \bpf{} signals as a reward during autoformalization. We prove a \emph{drift detection theorem} and a \emph{PAC-faithfulness} result establishing that the equivalence class of a natural language statement is learnable from $\mathcal{O}(\log(1/δ)/\varepsilon)$ probes under mild assumptions. We release \driftbench{}, a benchmark of $2{,}183$ NL/Lean~4 pairs with controlled drift labels across six subfields of mathlib4. \bpf{},+,\cpg{} detects $89.6\%$ of drifted formalizations at a $3.0\%$ false-positive rate-against $41.2\%$ for typecheck and $63.3\%$ for LLM-judge baselines, and \fgd{} reduces the rate at which a state-of-the-art autoformalizer emits drifted statements by $47\%$. https://pmlrbd.github.io/BPF/

Reference-free faithfulness metrics verify each atomic claim a model makes against ground truth, and are increasingly used to evaluate grounded generation. We show they share a blind spot: they measure only precision -- are the stated claims supported? -- and therefore reward abstention, since a model can score near-perfect faithfulness by saying almost nothing. We make this measurable using Formula 1 telemetry, a domain where strategic ground truth is derived deterministically and, crucially, completely: for each decision we know the full set of facts that mattered. This completeness -- absent in open-domain faithfulness benchmarks -- lets us measure recall (coverage of the relevant facts) exactly, alongside precision. On a multilingual (EN/ES/PT) benchmark of 7,253 decision instances spanning 157 races, the most precise frontier model covers under half of the relevant facts and ranks last by F1, so requiring coverage reorders the systems; the same effect reappears in a second complete-oracle domain (NOAA weather forecasts). Fine-tuning small models (1B-7B) on the complete oracle closes the precision-recall gap entirely (F1 ~0.98), beating every zero-shot frontier system regardless of scale. We pair faithfulness with coverage into a single score, validate the metric (controlled perturbation; agreement across a model-free regex extractor and a cross-family LLM extractor, system-level Spearman 1.0), and give a verifier-guided generation method that improves precision and recall without references. We release the benchmark, structured annotations, metric, baselines, and an interactive demo.

Do LLM agents act on the reasoning they state? This question of process fidelity is central to LLM-based social simulation, yet hard to measure where no reference for correct behavior exists. We study it in a controlled setting: a Texas Poker simulator with a verifiable reference action for every decision by splitting the faithfulness gap into two steps: reasoning-to-conclusion (does the stated decision follow from the agent's own reasoning?) and conclusion-to-action (does the agent execute what it states?). The two steps behave very differently. Conclusion-to-action is reliable: inconsistency is 0.7% for Claude Haiku 4.5 and 1.4% for DeepSeek-Reasoner once the conclusion is read from an explicit tag, whereas free-text conclusion extraction reports 22-26%. Reasoning-to-conclusion is where fidelity frays, but not through a single dominant failure. In a step-level diagnostic the agent's errors split roughly evenly between bad inputs, borderline cases, and rule misapplication deriving a conclusion that contradicts the agent's own restated rule from inputs it estimated correctly. This composition is model-dependent: rule misapplication accounts for a third of Haiku's interpretable errors but only 8% of DeepSeek's. The one robust signal is directional: when an agent does misapply its own stated rule, it almost always (99.5% for Haiku) errs in the risk-averse direction. The override is partly hedging behavior, not a capability limit: instructing the agent to apply the rule mechanically halves the misapplication rate (13.9% to 6.8% of decisions) and raises adherence by eight points. Process-fidelity evaluation should therefore elicit machine-checkable conclusions and probe for directional biases rather than assume a single upstream failure mode, lest it conflate measurement noise with model behavior.

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.

Chain-of-Thought (CoT) faithfulness, i.e., whether CoTs genuinely reflect large language models' (LLM) underlying behavior, is typically evaluated under two disjoint paradigms: contextual faithfulness, measured by perturbing the input or CoT trace, and parametric faithfulness, assessed by intervening on a model's parametric knowledge. Yet prior work compares them only descriptively. We fill this gap by proposing FaithMate, a unified preference-alignment interface for optimizing models towards either faithfulness paradigm. It enables us to investigate the interplay between the two paradigms, examining whether and to what extent faithfulness gains generalize within and across paradigms. Across three models, two datasets, and six faithfulness metrics, we find that the two paradigms are positively coupled, yet asymmetric: optimizing towards parametric faithfulness yields consistent gains across both paradigms, whereas the contextual counterpart delivers more variable gains. Within the contextual paradigm, faithfulness gains on one metric do not consistently transfer to others, implying that existing contextual metrics capture disjoint facets of faithfulness and exposing inherent trade-offs. These findings imply that CoT faithfulness is not a monolithic objective and therefore requires multifaceted optimization and evaluation.

Chain-of-thought (CoT) reasoning is useful for monitoring language models only when the reasoning trace faithfully reflects the computation that produces the final answer. However, models can rely on prompt-to-answer shortcuts that bypass the CoT, making the visible reasoning trace misleading even when it appears plausible. We study CoT faithfulness through a structural information-flow perspective: faithful reasoning should route answer-relevant information through the mediated path from prompt to CoT to answer, rather than through a direct prompt-to-answer shortcut. This perspective yields a task-agnostic framework based on three complementary properties, sufficiency, completeness, and necessity, which we instantiate with entropy-based, masked-KL, and gradient-based diagnostics. We show that these metrics recover externally judged faithfulness differences in hinted reasoning, and identify a low-entropy failure mode of KL-based diagnostics where gradient-based measures remain more stable. Building on this analysis, we introduce update-time interventions for verifier-based on-policy RL, including attention masking, backward-only gradient masking, CoT gradients, and adversarial perturbations of prompt representations. Across hinted arithmetic, reward-hackable code repair, and DAPO-Math models trained without hints but evaluated under wrong-hint injection, our interventions shift behavioral and structural indicators toward stronger CoT mediation. In particular, they make shortcut and reward-hacking behavior more transparent in the CoT and improve task-agnostic faithfulness metrics, while in some settings also reducing wrong-hint susceptibility. Our results suggest that controlling information flow during training is a practical route toward more faithful and monitorable CoT reasoning. Code is available at https://github.com/safety-research/faithful-cot.

Reinforcement learning with verifiable rewards (RLVR) has emerged as a promising paradigm for advancing complex reasoning in large language models, and recent work extends RLVR to multimodal large language models (MLLMs). This transfer, however, surfaces a faithfulness challenge: faithful perception of task-relevant visual evidence and faithful use of that evidence during reasoning, leading to unsatisfactory gains on multimodal benchmarks. Specifically, existing perception supervision often operates on textual descriptions rather than natively on image regions, and faithful use is largely overlooked, exposing the perception-reasoning disconnect where correctly perceived evidence is dropped or contradicted during reasoning. To close these gaps, we propose Faithful-MR1, a training framework that anchors and reinforces visual attention to address both halves of faithful multimodal reasoning. The Anchoring stage turns perception into an explicit pre-reasoning subtask, supervising a dedicated <Focus> token's attention directly against image regions rather than through textual descriptions. The Reinforcing stage exposes faithful use through counterfactual image intervention, rewarding answer-correct trajectories that concentrate visual attention where vision causally matters. Extensive experiments demonstrate that Faithful-MR1 outperforms recent multimodal reasoning baselines on both Qwen2.5-VL-Instruct 3B and 7B backbones while using substantially less training data.

Mechanistic interpretability aims to explain a model's behavior by identifying causally responsible internal structures. Dictionary-based explainers such as sparse autoencoders and transcoders are a primary tool, but their faithfulness under out-of-distribution (OOD) shift has received little systematic attention. We show that distribution shift rotates the subspace that the model actively uses, misaligning the explainer's dictionary trained on in-distribution (ID) activations. We formalize this misalignment as the faithfulness gap, a geometric distance between the ID dictionary and the OOD-active subspace, and show that it controls OOD faithfulness degradation. To reduce this gap, we propose the Geometry-Adaptive Explainer (GAE), which realigns the explainer's dictionary with the OOD-active subspace while preserving the original feature structure. This requires only unlabeled OOD activations and no gradient updates. We prove that GAE improves over the unadapted ID explainer, with excess loss bounded quadratically by the second-moment shift. Empirically, GAE even matches or surpasses all training-based baselines in causal faithfulness across multiple models and OOD settings.

We present the first systematic study of faithfulness in Vision-Language-Action (VLA) driving models, analyzing 300 Alpamayo-R1-10B inferences across 100 diverse PhysicalAI-AV scenarios. Our main finding is that output natural-language rationales with trajectories may be significantly unfaithful: (i) overall reasoning fidelity is only 42.5%, with Chain-of-Causation matching scene reality less than half the time; (ii) 94 missed pedestrians in one-third of pedestrian-relevant scenes; (iii) 97.7% trajectory fragility under mild visual perturbations; and (iv) only 48.3% mean reasoning-action consistency, with 53.3% of inferences exhibiting low consistency, including 37.9% of stop-claimed cases where the model continues instead. We formalize faithfulness information-theoretically, define entity and action fidelity with verification criteria, and outline a four-component safety architecture aligned with these results.

Post-hoc saliency methods are widely used to interpret deep neural networks, but their faithfulness is difficult to evaluate reliably. Existing evaluations mask features according to saliency-induced feature ordering and measure performance degradation, but this degradation can be confounded by the masking operator: zero masking may create out-of-distribution artifacts, while interpolation-based masking may preserve residual predictive information. We propose Adversarial Information Masking (AIM), a saliency-guided adversarial feature replacement framework for evaluating both saliency-map faithfulness and masking-operator reliability. AIM replaces selected features with values from an adversarial counterpart of the input and compares degradation under complementary masking orders. We assess reliability using random-attribution bias and stability of explanation-method faithfulness rankings. Experiments on image, audio, and EEG tasks suggest that AIM reduces masking-induced bias compared with zero and interpolation-based masking, while revealing modality-dependent differences between signed and unsigned attributions.

Zero-shot textual explanations aim to make image classifiers more transparent by probing their internal representations, without relying on task-specific supervision or LVLMs. However, existing methods often miss the features that truly drive the prediction, resulting in limited \textit{faithfulness} to the evidence underlying the model's decision. To address this, we propose FaithTrace. Motivated by the idea that faithful explanations should describe concepts that strongly influence the prediction, FaithTrace directly measures how much the representation induced by the explanation changes the class logit. We introduce an influence score, computed as the directional derivative of the class logit along the text-induced direction in the classifier's feature space, and use it as a proxy for faithfulness. Moreover, we extend this influence score into quantitative evaluation metrics, helping fill the gap in faithfulness evaluation for textual explanations. Experiments show that FaithTrace yields more faithful explanations than baselines, facilitating a more accurate understanding of the model. The code will be publicly released.

Literary translation requires balancing target-language fluency with faithfulness to the source. Recent large language models (LLMs) often produce fluent translations, but it remains unclear whether fluency corresponds to semantic preservation in literary text. We examine this relationship using 130,486 translated paragraphs from 106 novels in 16 source languages, including human, Google Translate, and TranslateGemma translations. Fluency is measured as original-likeness with a translationese classifier trained on paragraph part-of-speech n-grams, and faithfulness with the automatic translation evaluation metric COMET-KIWI. We control for paragraph length and find a consistent negative correlation between fluency and faithfulness. The pattern appears for both human and Google Translate, but is weaker and often non-significant for TranslateGemma. These results show that segment length matters for automatic evaluation and suggest a tradeoff between fluency and faithfulness in literary translation.

Despite strong performance of Multimodal Large Language Models (MLLMs) on multimodal tasks, predicting whether and why an image is persuasive remains challenging. We first show that prompting MLLMs to reason before prediction does not consistently help, and can even reduce persuasiveness prediction performance, suggesting that naively generated rationales are unreliable signals for this task. Yet, no established methodology exists for training MLLMs to reason about visual persuasion or evaluating whether their rationales faithfully support their decisions. To address this gap, we show empirically and theoretically that diverse teacher-generated rationales, when used for supervised fine-tuning, improve visual persuasiveness prediction. We further introduce a three-dimensional faithfulness evaluation framework covering rationale-to-decision consistency, rationale-to-image groundedness, and rationale-to-decision sensitivity. Applying this framework shows that prediction performance alone does not guarantee faithful rationales, while rationale-to-decision sensitivity is most aligned with human rationale preferences. These findings motivate faithfulness-aware training objectives and scalable rationale supervision for visual persuasiveness evaluation. Our code and dataset will be made publicly available.

Retrieval-Augmented Generation (RAG) models frequently produce answers grounded in parametric memory rather than the retrieved context, undermining the core promise of retrieval augmentation. A fundamental obstacle to fixing this unfaithfulness is the lack of training data that explicitly requires models to prefer context over internal knowledge. We introduce Faithfulness-QA, a large-scale dataset of 99,094 samples constructed through counterfactual entity substitution. Starting from two established extractive QA benchmarks--SQuAD and TriviaQA--we automatically identify answer-bearing named entities in each context, replace them with type-consistent alternatives drawn from a curated bank of 76,953 entities, and thereby manufacture controlled knowledge conflicts between context and parametric memory. Rigorous quality filtering ensures 100% pass rates across four automated checks on random 200-sample audits. We release the full dataset, the construction pipeline, and a typed entity bank covering eight named entity categories. Faithfulness-QA is designed as a training resource for attention-based faithfulness objectives and as an evaluation benchmark for measuring context-grounding behavior in RAG systems. Data and code are available at https://github.com/qzhangFDU/faithfulness-qa-dataset.

Recent advances in large audio language models (LALMs) have enabled multilingual speech understanding. However, benchmarks for evaluating LALMs remain scarce for non-English languages, with Korean being one such underexplored case. In this paper, we introduce KoALa-Bench, a comprehensive benchmark for evaluating Korean speech understanding and speech faithfulness of LALMs. In particular, KoALa-Bench comprises six tasks. Four tasks evaluate fundamental speech understanding capabilities, including automatic speech recognition, speech translation, speech question answering, and speech instruction following, while the remaining two tasks evaluate speech faithfulness, motivated by our observation that several LALMs often fail to fully leverage the speech modality. Furthermore, to reflect Korea-specific knowledge, our benchmark incorporates listening questions from the Korean college scholastic ability test as well as content covering Korean cultural domains. We conduct extensive experiments across six models, including both white-box and black-box ones. Our benchmark, evaluation code, and leaderboard are publicly available at https://ksbench.github.io/Korean-Benchmark/.

When asked to explain their decisions, LLMs can often give explanations which sound plausible to humans. But are these explanations faithful, i.e. do they convey the factors actually responsible for the decision? In this work, we analyse counterfactual faithfulness across 75 models from 13 families. We analyze the tradeoff between conciseness and comprehensiveness, how correlational faithfulness metrics assess this tradeoff, and the extent to which metrics can be gamed. This analysis motivates two new metrics: the phi-CCT, a simplified variant of the Correlational Counterfactual Test (CCT) which avoids the need for token probabilities while explaining most of the variance of the original test; and F-AUROC, which eliminates sensitivity to imbalanced intervention distributions and captures a model's ability to produce explanations with different levels of detail. Our findings reveal a clear scaling trend: larger and more capable models are consistently more faithful on all metrics we consider. Our code is available at https://github.com/google-deepmind/corr_faith.