Faithfulness of Language Model Reasoning

Language-model agents often reason before they act, and the value of monitoring that reasoning depends on how it relates to their actions. We ask how an agent's actions depend on its reasoning when the correct action depends on a changing state. In a small file-system task, background events announce which file's lock has been toggled but not its new value. The agent therefore has to reconstruct the current locks from its history before it can move files safely. With Qwen3-8B and Qwen3.5-9B, we separate two roles of reasoning: generating it before each action, and retaining earlier reasoning in the prompt. Most of the reduction in lock-violating commands comes from reasoning written at the moment of action. Retaining earlier reasoning adds little to rule-following, although it helps one model complete the task. The benefit depends on what the reasoning says. Filler text, reasoning written for another history and an accurate table of the state do not reproduce it. When we give the agent reasoning written for a history that differs in one lock value, the agent acts on that reasoning even against its own history. Without reasoning, linear probes recover the state from the model's activations about as well as a text classifier recovers it from the transcript. However, decoding accuracy does not predict which interventions change the agent's actions. In this setting, the benefit of reasoning for rule-following comes mainly from generating it at each step. The state that can be decoded without reasoning is a limited guide to how the agent acts.

While a LLM's chain of thought (CoT) is a promising approach to interpreting model outputs, prior work shows CoT's content does not always accurately reflect the reasoning process. Emmons et al. (2025) finds that models can follow a question hint without revealing so in its chain of thought, but when the hint is sufficiently complex the computation is forced into the CoT. This claim, however, has only been tested on Gemini models, and the datasets and code is closed source. We replicate their proxy task across eleven models from six families and find the necessity effect generalizes. Every model follows simple cues above a no-hint baseline, while concealed complex-cue following falls back near baseline for most models. We also extend the study from Emmons et al. (2025) to measure CoT-necessity and unfaithfulness across five cue families (four of them new) and three difficulty levels. The relationship between when CoT is necessary and unfaithfulness is task-specific and model-specific. For some hint types, models often verbalize their hint-following behavior even when the hint can be decoded in a single forward pass. These results support the case for CoT monitoring when unsafe behavior requires non-trivial computation, and that properties of CoT unfaithfulness are sensitive to model and setting.

Current Chain-of-Thought (CoT) verification methods detect reasoning faithfulness from model outputs or static activations, but are limited in explaining how unfaithfulness develops during generation. We introduce VERITAS to verbalise the model's computation: how different reasoning anchors emerge, transfer across layers, and surface in the output. VERITAS adopts the layer-wise Jacobian projection to produce a verbalised account of the model's internal reasoning evolution at each generated CoT token. By training a classifier on the extracted features along the traces, we show that these structured features provide insightful perspectives on how and where reasoning diverges toward unfaithfulness. Experiments on four datasets from FaithCoT-Bench and BonaFide show that VERITAS achieves a significant performance gain over LLM-as-Judge, with a 28.4% gain in balanced accuracy, and an average improvement of 16.0% over activation probing. Further analysis shows VERITAS can be generalised to different datasets and model architectures. Our work shows that verbalising model computation provides a richer and more interpretable lens on CoT faithfulness, moving beyond detection toward a mechanistic understanding of how reasoning fails.

Detecting faithfulness hallucinations in large language models (LLMs) is crucial for their safe deployment in high-stakes domains. Recent approaches address this by contrasting contextual against parametric knowledge utilization, yet measuring context reliance typically demands noticeable overhead, such as materializing attention maps, running additional forward passes, or training auxiliary probes. In this work, we show that context reliance can be estimated by projecting hidden states onto the dark space - directions that compose weakly with the unembedding matrix. Beyond its recognized function as an initial-token attention sink, we show that this subspace actively mediates context processing, and that its share of hidden-state norm - termed the Dark Matter Ratio (DMR) - provides a direct score of context utilization. Pairing DMR with a parametric knowledge score yields DarkLens, a lightweight faithfulness detector operating strictly within a single forward pass. Evaluated across three QA and summarization benchmarks on six open-weight LLMs (4B-13B parameters), DarkLens matches or outperforms a wide range of baselines without requiring extra forward passes or probe training.

Chain-of-Thought (CoT) faithfulness matters for monitoring, explaining and safely deploying large language models, yet it has been studied mostly in English. Languages differ in level, the amount of text available to learn from, so models know less in lower-resource languages, which may change how much they rely on their reasoning. To our knowledge, we are the first to relate CoT faithfulness to both resource level and task type (knowledge- or reasoning-intensive). On parallel multiple-choice questions from Multilingual Massive Multitask Language Understanding (MMMLU) in eight languages, faithfulness rises as resource level falls on knowledge-intensive tasks but not on reasoning-intensive ones, a gap that points the same way across four models and several faithfulness measures. We propose, and find support for, the Reliance Hypothesis: models rely more on their reasoning when they cannot recall the answer, as in lower-resource languages on knowledge-intensive tasks. Model confidence partly explains the effect of resource level, but not the knowledge-reasoning gap. Languages in which models are less confident rely more on reasoning, yet less confident answers within a language do not, a Simpson's paradox. In lower-resource languages, models rely on their reasoning for more answers, not more strongly for each. CoT monitoring may therefore be most informative in lower-resource languages, where it is least often validated.

Large Language Models (LLMs) are increasingly deployed for automated software vulnerability analysis. Binary classification alone is insufficient; practitioners need explanations to triage bugs and engineer patches. Standard practice relies on Chain-of-Thought (CoT) prompting, but free-form reasoning allows models to obscure logical leaps, hallucinated execution steps, and internal inconsistencies behind plausible prose. Our manual audit reveals that approximately 60% of correct vulnerability verdicts are accompanied by fabricated or unverifiable claims, and free-form explanations allow reasoning errors to evade LLM-as-a-judge evaluation. We present Vulnerability Explanation Reasoning Auditor (VERA), an automated framework for auditing LLM vulnerability reasoning. Rather than accepting free-form text, VERA asks models to output a Structured Reasoning Record (SRR) encoding tracked pointers, memory operations, and state transitions in machine-readable fields. A multi-stage judge audits each SRR against eight reasoning failure modes using deterministic checks, with LLM calls reserved for semantic interpretation. The standardized SRR schema also enables automated mutation testing to benchmark judges at scale without human annotation. Our evaluation shows reasoning flaws occur in correct verdicts just as frequently as incorrect ones, and VERA exposes 87% of reasoning errors that free-form LLM-as-judge systematically miss.

Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour. This reasoning comes at an increased inference cost, motivating efficient methods that train models to solve tasks using fewer tokens. However, a common concern is that such training may cause models to skip important reasoning steps, so the CoT no longer faithfully reflects the model's decision. It is unclear whether or when this occurs in practice, since different efficiency methods apply length pressure to models' CoT in distinct ways, and faithfully explaining a model's decision takes more tokens on some tasks than others. To understand these dynamics, we fine-tune a variety of models with three methods that apply length pressure differently, namely a fixed generation budget, a per-example length target, and a group-relative length reward. We evaluate how efficient reasoning affects CoT faithfulness (i.e., how well the CoT reflects model decisions on related inputs) and monitorability (i.e., whether the CoT reveals when input interventions alter the output). We find that it affects faithfulness and monitorability differently. Faithfulness falls in most settings, primarily because the trained models are less consistent. Monitorability is more robust, as models keep acknowledging the influence on their answer even when the CoT is much shorter.

Chain-of-thought (CoT) can sound plausible yet be unfaithful to the model's underlying reasoning. Most prior work probes CoT faithfulness through input--output behavior or input attributions, leaving internal computation largely underexplored. We instead cast faithfulness as internal concept grounding: Does a large language model's (LLM) CoT reasoning engage the same internal concepts that support the LLM's direct prediction, and do the shared concepts causally drive its answer? Encoding a prediction pass and a CoT pass with a single shared sparse autoencoder (SAE), a reliable approximator of the latent concepts LLMs use, makes their internal concepts directly comparable. We introduce three correlational metrics of concept-level alignment and a causal metric, $Δp$, which ablates the shared concepts and measures the drop in answer probability. Across five LLMs and four datasets, concept alignment is generally high, as indicated by the correlational metrics; yet these only identify which concepts are shared, not how much they causally contribute. $Δp$ fills this gap: causal faithfulness varies substantially with model depth, peaking at mid-to-late layers rather than the final ones, and model scale reshapes the layer-wise profile. Moreover, causally important shared concepts are not always verbalized in the CoT. These dissociations suggest that faithfulness cannot be reliably assessed from surface-level or representational correspondence alone; assessing it requires causal tests of whether the internal concepts underlying a CoT actually drive the model's prediction.

Large language models are increasingly used as social simulators, including as synthetic survey respondents. Most evaluations ask whether simulated outcomes resemble human outcomes. We argue that this is necessary but too weak: a simulator can match the final answer while using the wrong rationale-derived reason pattern. We study this problem through a 94-person sunscreen concept test in which each respondent evaluated three product concepts and wrote open-ended rationales. We map those rationales into signed reason states $Z$, where positive signs support adoption and negative signs block it. This gives a practical audit: holding respondent descriptors $D$, category context $K$, and concept treatment $X$ fixed, do human rationale-derived reasons help predict behavior $Y$, and can an LLM simulate the same reason state without seeing the human rationale or outcome? Human rationale-derived reasons substantially improve held-out prediction of purchase intent. LLM-simulated reasons are more brittle: they often sound plausible, but frequently echo the concept board rather than recover the respondent's acceptance or rejection path. The paper contributes an evaluation framework for social simulators. Reason states do not identify natural causal effects by themselves, but they provide an interpretable test of whether a simulator's stated reasons align with human evidence.

A key question for AI safety is whether a language model expresses all of its reasoning in its output tokens. We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks. We evaluate 13 frontier language models across three tasks and find that many models benefit significantly from filler tokens, with accuracy improvements of up to 13 percentage points. The benefit depends on which tokens are used and differs across models. We further show that filler tokens enable Claude Opus 4.5 to satisfy a hidden modular arithmetic constraint without sacrificing accuracy on its primary task, demonstrating that invisible reasoning can serve objectives entirely invisible to CoT monitoring. Reinforcement learning gives Qwen3-235B strong preferences over filler token content, but neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Our results indicate that frontier models already perform consequential computation with no interpretable trace in their output tokens.

Large Language Models (LLMs) achieve strong performance on reasoning tasks, but whether this reflects faithful logical inference or heuristic approximation remains unclear. We study this question in legal entailment by comparing three paradigms, including pure LLM classification, LLM-based Formal Reasoning, and solver-based Formal Reasoning using the Z3 SMT solver, on a re-annotated subset of ContractNLI across five LLMs. Our re-annotation reveals a systematic and measurable gap between pragmatic legal interpretation and strict formal entailment, where a substantial proportion of legally sound inferences are not formally grounded without additional unstated assumptions. While introducing formal structure improves accuracy, with LLM-based Formal Reasoning achieving the highest benchmark performance, we show that this gain does not imply faithful reasoning. We identify three recurring failure modes: scope laundering, where LLMs report solver-inconsistent classifications without executing the underlying formal reasoning, producing conclusions that appear logically grounded but are not; implicit constraint blindness, where LLMs overlook logical constraints present in formal representations; and program synthesis failures, where LLMs generate incorrect Z3 code despite structured prompting. Critically, scope laundering persists across all models, raising serious concerns about the faithfulness of LLM-based formal reasoning as a proxy for symbolic execution. These results reveal a fundamental gap between benchmark accuracy and logical faithfulness.

Reasoning-based LLM guardrails improve safety moderation by generating explicit rationales before issuing final decisions. However, their rationales do not always lead to faithful enforcement: a model may recognize a harmful intent in its reasoning but still predict a safe label, or issue an unsafe decision without policy-grounded justification. We identify this safety-critical failure mode as the deliberation-to-enforcement gap. Unlike general chain-of-thought faithfulness, guardrail reliability requires policy execution consistency: the generated reasoning should be grounded in the safety policy, and the final decision should be entailed by that reasoning. We propose ConsisGuard, a consistency-aware framework for reasoning-based LLM guardrails. ConsisGuard performs Policy-to-Decision Trajectory Distillation and Functional Coupling Alignment, aligning the internal coupling between safety deliberation and decision enforcement. Experiments on prompt and response harmfulness detection benchmarks show that ConsisGuard improves detection performance while reducing policy execution failures. These results suggest that reliable reasoning-based guardrails require accurate faithful execution of safety policies.

When a language model sees a document contradicting its training knowledge, it must choose: follow the document or trust itself. Prior work proved this choice depends on how well-known the fact is. We ask: does the model's chain-of-thought (CoT) reasoning faithfully report this mechanism? We introduce introspective faithfulness and test it across 200 questions, 8 models, and 4 prompt conditions. We find CoT reasoning is highly stable across opposite decisions: flip pairs retain 96% of same-answer similarity (d=0.34; confirmed by ROUGE-L, d=0.45). Yet self-rated confidence carries a faint genuine signal: for obscure facts where entity fame is uninformative, confidence still predicts decisions (p<0.001) and tracks item-level knowledge (r=0.134). GPT-4o is the only model with statistically reliable reasoning-decision coupling. Claude Sonnet 4.6 shows the widest confidence range (SD=1.39) but near-zero pooled correlation because the confidence-decision relationship reverses between conditions; a temperature ablation confirms this is model-specific. Internal thinking tokens show greater decision-sensitivity than user-facing CoT (p=0.033). CoT decomposes into a decision-invariant knowledge display (~96%) and a thin confidence layer with weak but real signal. For monitoring: read confidence, not the argument.

Chain-of-Thought (CoT) reasoning has advanced large language models (LLMs), but outcome-based supervision leads to pervasive post-hoc rationalization, producing plausible yet unfaithful reasoning chains. Most prior faithfulness assessment methods are either unscalable, expensive, or unreliable. We propose GeoFaith, a spatio-temporal framework that leverages latent geometric structure and entropy dynamics to diagnose and enforce faithful reasoning. We develop a scalable bootstrapping pipeline expanding step-level annotations from 1k to 20k samples across four domains, train an 8B faithfulness detector outperforming GPT-5 on standard benchmarks, and design a faithfulness-aware reinforcement learning framework jointly optimizing outcome correctness, process faithfulness, and trajectory consistency. Experiments show the proposed method achieves superior performance on both faithfulness detection and downstream reasoning, producing shorter, more interpretable chains without sacrificing accuracy. Our code will be made available publicly.

Linguistic cues such as "I believe" and "probably" offer an intuitive interface for communicating confidence, yet a generalisable, principled calibration framework for linguistic confidence expressions remains underexplored. In particular, co-occurring linguistic cues, contextual variation, and subjective audience interpretation pose unique challenges. We therefore model linguistic confidence as a distribution over plausible perceived probability values that a statement is correct, capturing interpretation variability that scalar representations discard. Within this distributional framework, we introduce faithfulness as a complementary evaluation dimension and present Faithfulness Divergence (FD), an information-theoretic metric quantifying the surprise induced in audience beliefs upon truth revelation. Building on these foundations, we present Retrieval-Augmented Linguistic Calibration (RALC), a lightweight post-hoc pipeline that propagates calibrated confidence signals back into natural language via retrieval-augmented rewriting. Across three QA benchmarks and five LLM families, RALC improves in-domain faithfulness and calibration up to 66% and 58%, respectively, outperforming black-box and grey-box calibration baselines.

LLMs deployed multilingually are often audited via English explanations for non-English inputs. We evaluate extractive explanations ''where the model identifies input token spans as evidence alongside a generated rationale'' and uncover a systematic trade-off: English-pivot explanations can achieve higher span agreement with human rationales while their evidence becomes less causally grounded in the model's prediction, as measured by both comprehensiveness and sufficiency. Across 3 tasks, 5languages, and 2multilingual LLM families, we find that English explanations frequently produce fluent but loosely anchored rationales, with comprehensiveness degrading by up to 5.7x relative to native-language conditions - even as task accuracy remains stable across settings. For socially nuanced classification, English pivots also fail to preserve pragmatic cues, reducing both faithfulness and span agreement. We recommend auditing explanations in the input language, reporting multi-faceted faithfulness metrics beyond lexical overlap, and treating English rationales as communication summaries rather than faithful decision traces.

Reasoning models post-hoc rationalize answers they have already committed to internally, producing chains of reasoning theater: deliberative-looking steps that contribute nothing to correctness. This wastes inference tokens, pollutes interpretability, and obscures what the model actually computed. We introduce ProFIL (Probe-Filtered Reinforcement Learning) to reduce theater, increase chain-of-thought faithfulness, and shrink chain length in a single, drop-in extension to Group Relative Policy Optimization (GRPO). A multi-head attention probe is trained once on the frozen base model to detect post-commitment steps from internal activations alone; during GRPO, rollouts whose probe score exceeds a threshold have their advantage zeroed. Our central finding is that a probe trained on a frozen base, with verifier-derived labels and no human annotation, provides a stable signal that suppresses theater while resisting the RL-obfuscation failure mode predicted by prior work. Across four reasoning domains (GSM8K, LiveCodeBench, ToolUse, MMLU-Redux) and two model architectures (Llama-8B, Qwen-7B), ProFIL reduces post-commitment theater by 11--100%, raises faithful-fraction (e.g., +24pp on LiveCodeBench under an independent Claude 3.7 Sonnet judge), and shortens chains by 4--19%, all while preserving or improving task accuracy. ProFIL also beats a matched length-penalty GRPO baseline, isolating the gain as semantic commitment-detection rather than chain compression. Probe weights, training configurations, and rollouts are released across all four domains.

When copies of the same language model are prompted to debate, they produce diverse phrasings of one perspective rather than diverse perspectives. Multi-agent debate (MAD), and more broadly closed-system reasoning where agents iteratively transform each other's outputs, tends to preserve answer accuracy while degrading the reasoning behind those answers. We name the multi-agent case the Debate Trap and the broader phenomenon the Reasoning Trap, offering a programmatic theory of evidence-grounded reasoning failure.The framework has three parts: (i) SFS (Supported Faithfulness Score), a claim-level metric verifying decomposed atomic claims against provided evidence (decomposer-invariant rankings: Spearman rho=1.0); (ii) EGSR (Evidence-Grounded Socratic Reasoning), replacing adversarial argumentation with evidence-grounded inquiry; (iii) Theorem 1 (DPI Bound): under standard MAD, the chain E -> O^0 -> O^1 -> ... is Markov, and the Data Processing Inequality implies E[I(E;O^{t+1})] <= E[I(E;O^t)]. Three companion results -- open-system recovery (Theorem 2), EGSR accumulation (Lemma 2), and vote-aggregation floor (Proposition 1) -- partition multi-step LLM reasoning by its information-theoretic relationship to E. Across 16 conditions on SciFact (300 claims) and FEVER (1,000 claims), DebateCV (C13) preserves 88% of baseline accuracy while SFS drops 43%; majority-vote MAD (C15) reduces SFS to 1.7% of baseline (p < 10^{-6}, d = -0.96); EGSR recovers 98%. An R6 cohort study (Korean n=10x30 FEVER; English n=3x200 SciFact) finds inter-rater Fleiss kappa <= +0.018 with 0.8-1.4 Likert intra-rater shifts across language and domain -- the human agreement that faithfulness metrics have been calibrated against is not itself stable. We offer one falsifiable conjecture: any closed-system reasoning protocol preserving Theorem 1's Markov structure is, in expectation, subject to the same DPI bound.

When a language model answers a table question, users have no way to verify which cells informed which reasoning steps. We introduce RSAT, a method that trains small language models (SLMs, 1-8B) to produce step-by-step reasoning with cell-level citations grounded in table evidence. Phase 1 (SFT) teaches a structured JSON output format from verified reasoning traces. Phase 2 (GRPO) optimizes a composite reward centered on NLI-based faithfulness, alongside citation validity and parsimony. Across six models from two families-Qwen 2.5 (1.5B/3B/7B) and Llama 3 (1B/3B/8B)-RSAT improves faithfulness 3.7$\times$ over SFT alone (0.224$\rightarrow$0.826), with near-perfect citation validity (0.992). Post-hoc attribution collapses below 13% format success, confirming that attribution must be integrated into reasoning, not retrofitted. Ablations show the faithfulness reward is essential: removing it drops faithfulness from 0.97 to 0.03.

In large language model (LLM) agents, reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating memory. However, coherent reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate reasoning trajectories, we propose Self-Audited Verified Reasoning (SAVeR), a novel framework that enforces verification over internal belief states within the agent before action commitment, achieving faithful reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve reasoning faithfulness, we perform adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six benchmark datasets demonstrate that our approach consistently improves reasoning faithfulness while preserving competitive end-task performance.

Language models increasingly show their work by writing step-by-step reasoning before answering. But are these steps genuinely used, or is the answer rigid - fixed before reasoning begins? We introduce the Step-Level Reasoning Capacity (SLRC) metric and prove it is a consistent causal estimator (Theorem 1). We propose LC-CoSR, a training method with Lyapunov stability guarantees that directly reduces rigidity. Evaluating 16 frontier models (o4-mini, GPT-5.4, Claude Opus, Grok-4, DeepSeek-R1, Gemini 2.5 Pro, and others) across six domains at N=133-500, we find reasoning falls into three modes. OpenAI's o4-mini shows 74-88% step necessity on five of six tasks (73.8-88.3%) - the highest SLRC in our study. The critical differentiator is RL-based reasoning training, not thinking tokens: Grok-4's reasoning mode shows lower faithfulness than its non-reasoning mode (1.4% vs 7.2% necessity). We discover a faithfulness paradox - high-SLRC models are more susceptible to sycophancy - and propose the Reasoning Integrity Score (RIS = SLRC x (1-Sycophancy)), which significantly predicts error detection (rho=0.66, p=0.026). LC-CoSR achieves 2.6x less negative reward than FARL and CSR baselines without external model dependencies.

Chain-of-thought (CoT) reasoning has been proposed as a transparency mechanism for large language models in safety-critical deployments, yet its effectiveness depends on faithfulness (whether models accurately verbalize the factors that actually influence their outputs), a property that prior evaluations have examined in only two proprietary models, finding acknowledgment rates as low as 25% for Claude 3.7 Sonnet and 39% for DeepSeek-R1. To extend this evaluation across the open-weight ecosystem, this study tests 12 open-weight reasoning models spanning 9 architectural families (7B-685B parameters) on 498 multiple-choice questions from MMLU and GPQA Diamond, injecting six categories of reasoning hints (sycophancy, consistency, visual pattern, metadata, grader hacking, and unethical information) and measuring the rate at which models acknowledge hint influence in their CoT when hints successfully alter answers. Across 41,832 inference runs, overall faithfulness rates range from 39.7% (Seed-1.6-Flash) to 89.9% (DeepSeek-V3.2-Speciale) across model families, with consistency hints (35.5%) and sycophancy hints (53.9%) exhibiting the lowest acknowledgment rates. Training methodology and model family predict faithfulness more strongly than parameter count, and keyword-based analysis reveals a striking gap between thinking-token acknowledgment (approximately 87.5%) and answer-text acknowledgment (approximately 28.6%), suggesting that models internally recognize hint influence but systematically suppress this acknowledgment in their outputs. These findings carry direct implications for the viability of CoT monitoring as a safety mechanism and suggest that faithfulness is not a fixed property of reasoning models but varies systematically with architecture, training method, and the nature of the influencing cue.

We provide evidence of performative chain-of-thought (CoT) in reasoning models, where a model becomes strongly confident in its final answer, but continues generating tokens without revealing its internal belief. Our analysis compares activation probing, early forced answering, and a CoT monitor across two large models (DeepSeek-R1 671B & GPT-OSS 120B) and find task difficulty-specific differences: The model's final answer is decodable from activations far earlier in CoT than a monitor is able to say, especially for easy recall-based MMLU questions. We contrast this with genuine reasoning in difficult multihop GPQA-Diamond questions. Despite this, inflection points (e.g., backtracking, 'aha' moments) occur almost exclusively in responses where probes show large belief shifts, suggesting these behaviors track genuine uncertainty rather than learned "reasoning theater." Finally, probe-guided early exit reduces tokens by up to 80% on MMLU and 30% on GPQA-Diamond with similar accuracy, positioning attention probing as an efficient tool for detecting performative reasoning and enabling adaptive computation.

Chain-of-thought (CoT) monitoring has emerged as a compelling method for detecting harmful behaviors such as reward hacking in reasoning models, under the assumption that models' reasoning processes are informative of such behaviors. However, when a model's trained behaviors conflict with instructions provided at inference time, it faces two competing objectives it cannot simultaneously satisfy. We find that models resolve this tension through motivated reasoning---generating plausible-sounding justifications for violating their instructions while downplaying potential harms or contradictions---rather than genuine engagement with the spirit of the provided instructions. We investigate this phenomenon by training reasoning models via RL on objectives that conflict with their later constitutional constraints, and find that motivated reasoning emerges systematically across all settings studied, including beyond safety-related tasks. As motivated reasoning becomes more prevalent over the course of training, CoT monitors are increasingly fooled. Concerningly, the same monitor that correctly flags a violation when shown only the final answer is persuaded to approve it when shown the reasoning trace. While large reasoning models closely track human ability in detecting motivated reasoning, this is not a full solution, since frontier model providers commonly rely on smaller models for monitoring due to latency and cost constraints. Our results underscore the necessity for further research into the emergence and detection of motivated reasoning in model evaluation and oversight. Code for this paper is available at https://github.com/nikihowe/motivated-reasoning.