Large language models produce chains of thought (CoTs) whose stated reasoning may not reflect the computation that actually produced the answer. Gur-Arieh, Marasović and Geva (2026) constructed BonaFide, the first benchmark with ground-truth faithfulness labels (derived from task designs whose outputs reveal which intermediate computations must have occurred), and showed that all eight prominent faithfulness metrics they test perform near chance, exhibit strong prediction biases, degrade on longer CoTs, and are too expensive for runtime monitoring (the best CoT-level metric, CC-SHAP, reaches only 0.70 AUROC at up to 10310^3 seconds per instance, while the best step-level metric, Filler Tokens, reaches 0.59 and neither transfers across settings). The open problem is to develop and validate a faithfulness measure that reliably distinguishes faithful from unfaithful reasoning at both the step and whole-CoT level against ground-truth labels, transfers across models and task types, and is efficient enough for practical use. No subsequent detector validates against ground-truth labels at both granularities.

Background

A chain of thought (CoT) is the step-by-step verbalized reasoning a large language model (LLM) emits before its final answer. A CoT is faithful if it contains the complete reasoning path the model actually followed and contains no steps misrepresenting that computation; it is unfaithful if it is a post-hoc rationalization or omits decisive influences (for example, following a misleading hint inserted in the prompt without mentioning it). Whether CoTs can be trusted matters for interpretability and for AI safety, where CoTs are proposed as a monitoring surface for catching misbehavior.

Existing operationalizations measure properties that are logically distinct from faithfulness. Plausibility asks whether the reasoning sounds like a reasonable route to the answer — but a convincing rationalization is by definition plausible. Importance asks whether a step causally influenced the answer, usually tested by perturbation: Lanham et al. (2023) delete or truncate steps and measure answer changes; Tutek et al. (2025) 'unlearn' a step and check whether the answer changes (FUR); Zaman and Srivastava (2025) evaluate metrics by editing a model's knowledge. A step can faithfully describe a real dead-end exploration without being important, and a fabricated justification the model conditions on can be important while being unfaithful. FaithCoT-Bench (Shen et al., 2025) provides expert annotations for instance-level unfaithfulness detection, but its labels come from human plausibility judgments; the monitorability literature (Baker et al., 2025) largely bypasses faithfulness measurement on the grounds that it is intractable.

Gur-Arieh, Marasović and Geva (2026) broke this impasse with a ground-truth construction: in 'outright' and 'diversionary' task designs, the model's output itself reveals which intermediate computations must have occurred (e.g., an answer that exactly matches an injected hint proves the model consulted the hint). An automated labeling pipeline (validated at 98.9% precision against human annotation) yields BonaFide: 3,066 CoTs across 13 tasks and 10 models from the Qwen3, OLMo 3, DeepSeek-R1-Distill and Llama 3.3 families, with 1,120 CoT-level labels (85% unfaithful — faithful CoTs are harder to certify because every step must be faithful or inert) and 1,946 step-level labels (roughly balanced). AUROC (area under the receiver operating characteristic curve, 0.5 = chance, 1.0 = perfect) is the headline discrimination measure. Evaluating eight prominent metrics spanning importance-based, semantic-utility, and attribution-based families, they find: most metrics sit near chance at both granularities; the best CoT-level metric (CC-SHAP, an attribution-based method that compares SHAP token-contribution distributions for generating the answer and for generating the CoT) reaches 0.70 AUROC while the best step-level metric (Filler Tokens, a perturbation test that replaces the reasoning text with uninformative repeated tokens and measures answer preservation) reaches 0.59, with neither transferring to the other setting; importance-based metrics label 90–96% of CoTs unfaithful and semantic-utility metrics label 94–96% faithful on a balanced sample, i.e., systematic prediction bias rather than measurement; accuracy degrades on longer CoTs (with the cause — importance conflation versus per-step context dilution — left open); and CC-SHAP costs up to 10310^3 seconds per instance, precluding runtime use.

Subsequent work sharpens rather than fills the gap. The FaithCoT-Bench authors' CIE-Scorer (Shen et al., 2026) traces compact sentence-level internal 'circuits' (sparse subnetworks of the model implicated in the computation) and scores their discrepancy from the verbalized reasoning graph, achieving state-of-the-art instance-level detection on FaithCoT-Bench — but it is validated only against plausibility-style annotations, not ground-truth labels, and reports no step-level results. Angdembay, Aryal and Rahimi (2026) show on FaithCoT-Bench that behavioral detection is structured by answer correctness: on incorrect answers, where most annotated unfaithfulness lives, no tested signal beats chance, and the standard step-removal metric anti-correlates with human labels. Sarfati et al. (2026) find that linear probes on internal activations (linear readouts trained to detect information in the model's hidden states) track hidden influences better than the reasoning trace, but only for forecasting models in one domain. Jia, Benton and Easley (2026) propose information-flow diagnostics (sufficiency, completeness, necessity of prompt→CoT→answer routing) validated against human-judged hinted reasoning, again without ground-truth labels at both granularities. A separate 2026 line (e.g., steering-vector, RL, and training interventions) improves faithfulness rather than measuring it.

Problem Statement

Construct a measure (metric or detection procedure) of chain-of-thought faithfulness — a procedure that, given a model, a task instance, and the CoT the model produced, outputs a faithfulness score for individual steps and for the whole CoT — and validate it against ground-truth faithfulness labels. The required objective, which no existing method meets, has three jointly necessary components:

  1. Ground-truth-validated reliability at both granularities. The measure must reliably distinguish faithful from unfaithful CoTs at both the step level and the CoT level, evaluated against ground-truth labels obtained from task constructions whose outputs reveal which intermediate computations must have occurred (the BonaFide benchmark and its public label release are the natural evaluation vehicle; an equivalent or stronger ground-truth construction is admissible). 'Reliably' is judged against the field's quantitative conventions — discrimination well above the near-chance AUROCs of 0.5–0.70 that current metrics achieve, without the systematic prediction biases documented on BonaFide (importance-based metrics labeling 90–96% of instances unfaithful, semantic-utility metrics labeling 94–96% faithful), and without the current property that a metric strong at one granularity is at or below chance at the other.
  2. Transfer across settings. The measure must demonstrably generalize across model families and task types rather than working in a single setting — the evaluation vehicle itself spans multiple settings (BonaFide covers 10 models from four model families and 13 tasks), so ground-truth-validated reliability must hold across these settings rather than in a single one.
  3. Practical cost. The measure must be efficient enough for practical use as an evaluator or runtime monitor of contemporary long-CoT reasoning models — orders of magnitude below the 10310^3-seconds-per-instance cost of the strongest current metric (CC-SHAP).

The primary paper additionally discusses, as accompanying open explanatory sub-questions rather than as deliverables of the measurement goal, why existing metric families fail — the conflation of faithfulness with importance or plausibility — and why accuracy degrades on longer CoTs (whether driven by per-step context dilution or other factors, which the paper leaves unresolved). Both black-box methods (using only inputs and outputs) and white-box methods (using internal activations, circuits, or attribution) are admissible; no method class, model family, or task domain is prescribed.

References

  1. Yoav Gur-Arieh, Ana Marasović, Mor Geva. 'Faithfulness Metrics Don't Measure Faithfulness: A Meta-Evaluation with Ground Truth.' 2026. arXiv:2605.25052, doi:10.48550/arxiv.2605.25052, https://arxiv.org/abs/2605.25052
  2. Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, Kamilė Lukošiūtė, Karina Nguyen, Newton Cheng, Nicholas Joseph, Nicholas Schiefer, Oliver Rausch, Robin Larson, Sam McCandlish, Sandipan Kundu, Saurav Kadavath, Shannon Yang, Thomas Henighan, Timothy Maxwell, Timothy Telleen-Lawton, Tristan Hume, Zac Hatfield-Dodds, Jared Kaplan, Jan Brauner, Samuel R. Bowman, Ethan Perez. 'Measuring Faithfulness in Chain-of-Thought Reasoning.' 2023. arXiv:2307.13702, doi:10.48550/arxiv.2307.13702, https://arxiv.org/abs/2307.13702
  3. Martin Tutek, Fateme Hashemi Chaleshtori, Ana Marasović, Yonatan Belinkov. 'Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps.' 2025. EMNLP 2025, arXiv:2502.14829, doi:10.18653/v1/2025.emnlp-main.504, https://aclanthology.org/2025.emnlp-main.504/
  4. Kerem Zaman, Shashank Srivastava. 'A Causal Lens for Evaluating Faithfulness Metrics.' 2025. EMNLP 2025, arXiv:2502.18848, doi:10.18653/v1/2025.emnlp-main.1496, https://aclanthology.org/2025.emnlp-main.1496/
  5. Xu Shen, Song Wang, Zhen Tan, Laura Yao, Xinyu Zhao, Kaidi Xu, Xin Wang, Tianlong Chen. 'FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning.' 2025. arXiv:2510.04040, doi:10.48550/arxiv.2510.04040, https://arxiv.org/abs/2510.04040
  6. Xu Shen, Zhen Tan, Song Wang, Pingjun Hong, Rui Miao, Xin Wang, Tianlong Chen. 'Detecting Unfaithful Chain-of-Thought via Circuit-Guided Internal-External Discrepancy.' 2026. arXiv:2605.25603, doi:10.48550/arxiv.2605.25603, https://arxiv.org/abs/2605.25603
  7. Suramya R. Angdembay, Dikshant Aryal, Nick Rahimi. 'Two Regimes of Chain-of-Thought Unfaithfulness: Metric-Based Detection Fails Where Models Are Wrong.' 2026. arXiv:2607.23458, doi:10.48550/arxiv.2607.23458, https://arxiv.org/abs/2607.23458
  8. Raphaël Sarfati, Pratyush Ranjan Tiwari, Siddharth Boppana, Christopher J. Earls, Srikar Varadaraj, Eric Ho. 'What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness.' 2026. arXiv:2607.08046, doi:10.48550/arxiv.2607.08046, https://arxiv.org/abs/2607.08046
  9. Jinghan Jia, Joe Benton, Eric Easley. 'Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning.' 2026. arXiv:2605.24286, doi:10.48550/arxiv.2605.24286, https://arxiv.org/abs/2605.24286
  10. Bowen Baker, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y. Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, David Farhi. 'Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation.' 2025. arXiv:2503.11926, doi:10.48550/arxiv.2503.11926, https://arxiv.org/abs/2503.11926
  11. Adam Karvonen, Euan Ong, Subhash Kantamneni, Samuel Marks. 'Would this change your answer? Evaluating Explanations of LLM Behavior In The Wild with Counterfactual Experiments.' 2026. arXiv:2608.16747, doi:10.48550/arxiv.2608.16747, https://arxiv.org/abs/2608.16747