Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection

Yujing Wang; Yulan He; Wenya Wang; Hanqi Yan

cs.CLOct 16, 2026

Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection

Authors: Yujing Wang, Yulan He, Wenya Wang, Hanqi Yan

Organizations: King’s College London · Nanyang Technological University

Abstract

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.

Figures & tables

Figure 1: Internal (solid) and visible-CoT lexical-proxy (dashed) trajectories on TruthfulQA, projected onto the internal faithful–unfaithful J-space direction.

Figure 2: J-space trajectories across model depth at four normalized CoT phases on AQuA. Panels (a)–(d) follow the progression of the visible CoT. Arrows indicate increasing layer depth, and shading shows approximate 95% confidence intervals.

Figure 3: Overview. Top : A contrast CoT pair is mapped to a shared J-space for faithful anchor (

\tilde{J}

) discovery. The resulting anchor dictionary

\mathcal{A}_{\ell,p}

explains how reasoning stages shift across different model layers

\ell

as CoT progression (

p

) via representative tokens. Bottom : Given an evaluation CoT, the structured faithful detector is built to measure the CoT’s transitional features (Map, Profile, Dynamics) against the discovered faithful anchor.

Figure 4: Functional composition of faithfulness-related anchors on FaithCoT-Bench using Qwen3. (a)–(b) show changes across normalized CoT phases, while (c)–(d) show changes across model depth. Anchors are grouped into Context, Explanation, Checking, and Reporting stages, and each region reports its share of positive label-directed contrast mass. The complete words in each category are provided in Appendix B.7 .

		BonaFide		AQuA		LogiQA		TruthfulQA		HLE-Bio
	Method	Bal. acc.	AUROC	Bal. acc.	AUROC	Bal. acc.	AUROC	Bal. acc.	AUROC	Bal. acc.	AUROC
Qwen3	Perplexity	0.565	0.582	0.510	0.558	0.562	0.570	0.563	0.540	0.533	0.520
	Early Ans	0.537	0.528	0.578	0.598	0.695	0.715	0.722	0.766	0.505	0.548
	Info Gain	0.742	0.847	0.631	0.692	0.566	0.566	0.545	0.577	0.552	0.516
	Ans Tracing	0.631	0.683	0.668	0.736	0.709	0.735	0.722	0.742	0.545	0.577
	LLM judge	0.581	–	0.673	–	0.606	–	0.571	–	0.444	–

Table 1: Faithful detection evalaution results. VERITAS p denotes the best Dynamics-Map configuration, and VERITAS c additionally incorporates Profile. LLM judge produces hard labels, so its AUROC is unavailable.

Figure 5: Layer–phase anchor alignments (Map) on BonaFide using Qwen3. Each cell shows the group-balanced mean standardized alignment for faithful (A) and unfaithful (B) CoTs.

Panel of Figure 6 — Figure 6: Analysis of faithfulness-relevant anchor alignments using Qwen3 representations. (a) Mean Profile alignments for the 10 most discriminative anchors. (b) Group-balanced mean alignment across normalized CoT phases, averaged within the workspace and motor layers. Shaded regions denote 95% confidence intervals.

Panel of Figure 7 — Figure 7: Effect of key factors on Qwen3-4B. (a) Balanced accuracy of features within CoT-length terciles on BonaFide. (b) Balanced accuracy on FaithCoT when layers are merged into $K$ stages.

Panel of Figure 8 — Figure 8: Transfer on FaithCoT-Bench. (a) Cross-subset transfer with Qwen3. Anchor transfer reuses the anchors discovered on the source subset and refits the classifier on the target, whereas zero-shot transfer freezes both. (b) Cross-model transfer from Llama3.1 to Qwen3, applying the classifier trained on Llama3.1 unchanged. Percentages show the retained fraction of in-domain balanced accuracy.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	#CoTs	#Questions	#Faithful	#Unfaithful
BonaFide	1,016	541	127	889
AQuA	354	100	273	81
LogiQA	400	100	273	127
TruthfulQA	389	100	280	109
HLE-Bio	161	41	96	65

Appendix

Table 2: Dataset statistics.

Category	Tokens
Context	actual, context, external, fact, hidden, influence, information, input, internal, meaning, prompt, question, relevant, state, understand, visible
Explanation	claim, explain, explanation, justify, logic, reason, reasoning, support
Checking	check, correct, critical, error, evidence, faith, false, quality, true, valid, verified, verify, wrong
Reporting	answer, conclusion, decision, express, expression, final, output, report, reported, reporting, statement

Appendix

Table 3: Functional grouping of anchor tokens used in Figure 4 , before removing tokens shared by the faithful and unfaithful anchor pools.

Feature set	$d_{Q}$	BonaFide	FaithCoT	$d_{L}$	BonaFide	FaithCoT
	Qwen3-4B			Llama3.1-8B
Map	146	0.814	0.684	126	0.841	0.640
Profile	161	0.819	0.705	108	0.838	0.629
Dynamics	247	0.840	0.668	211	0.833	0.643
Dynamics $+$ Profile	408	0.831	0.689	319	0.854	0.622
Dynamics $+$ Map	393	0.813	0.703	337	0.867	0.696

Appendix

Table 4: Feature-set ablation in balanced accuracy.

d_{Q}

and

d_{L}

denote the feature counts under the Qwen3 and Llama3.1 models, respectively.

Figure 9: Effect of cross-source on the three structured features using Qwen3-4B. Zero-shot transfer on FaithCoT, where anchors and classifier are frozen on one source subset and evaluated on another.

Figure 10: Cross-source generalization in AUROC on FaithCoT-Bench. Anchor transfer refits the classifier on the target while freezing the transferred anchors, whereas zero-shot transfer freezes both. Error bars show standard deviations over the twelve ordered source pairs; percentages indicate performance retained relative to the in-domain setting.

Figure 11: Sensitivity to the number of normalized CoT phases

P

on FaithCoT-Bench using Qwen3. Shading marks the default setting

P=5

Figure 12: Sensitivity to the number of anchors per cell

\kappa

on FaithCoT-Bench using Qwen3. Results are macro-averaged over the four sources; shading marks the default setting

\kappa=8

CoT Faithfulness

Chain-of-thought faithfulness measures whether a model’s stated reasoning causally reflects and supports the answer it gives.

See all 72 papers

Detecting Unfaithful Chain-of-Thought via Circuit-Guided Internal-External Discrepancy
May 25, 2026Xu Shen, Zhen Tan, Song Wang +4CoT Faithfulness LLM Auditing
1Jilin University · University of Central Florida · 3Arizona State University +2
Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models?
Mar 23, 2026Richard J. YoungCoT Faithfulness Faithfulness of Language Model Explanations
University of Nevada, Las Vegas, Department of Management, Entrepreneurship and Technology, Lee Business School, Las Vegas, NV, USA · DeepNeuro AI, Las Vegas, NV, USA
Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning
May 22, 2026Jinghan Jia, Joe Benton, Eric EasleyCoT Faithfulness LLM Interpretability
†Dept. CSE, Michigan State University · ⋆Anthropic Fellows Program · ‡Anthropic

May 25, 2026J/K move · Enter open · S save

Detecting Unfaithful Chain-of-Thought via Circuit-Guided Internal-External Discrepancy

Xu Shen, Zhen Tan, Song Wang, Pingjun Hong+3

1Jilin University · University of Central Florida · 3Arizona State University+2

CoT Faithfulness LLM Auditing Circuits in Language Models Reasoning Trace Analysis

Open paper AskPDF arXiv

Chain-of-thought (CoT) reasoning improves the problem-solving ability of large language models (LLMs), but generated reasoning traces may not faithfully reflect the model's actual decision process. Existing CoT unfaithfulness detectors mainly rely on external signals from generated rationales, such as textual plausibility or answer consistency, while overlooking evidence from the model's internal computation. Although recent circuit tracing methods provide a way to obtain model-internal evidence by tracing how information flows through model components during reasoning, constructing full reasoning circuits for long CoTs is costly and difficult to scale. To address these challenges, we propose Circuit-guided Internal-External Discrepancy Scorer (CIE-Scorer), a framework for instance-level CoT unfaithfulness detection. The key idea is that faithful reasoning traces should align with the model's computational process, whereas unfaithful traces may diverge from it. CIE-Scorer efficiently traces compact sentence-level circuits from informative reasoning tokens, constructs internal and external reasoning graphs, and measures their discrepancy using Fused Gromov--Wasserstein distance. Experiments on four datasets from FaithCoT-Bench show that CIE-Scorer achieves state-of-the-art performance while reducing the cost of circuit construction, demonstrating the effectiveness of combining mechanistic interpretability signals with external reasoning traces for CoT unfaithfulness detection.

Abstract

CoT Faithfulness

Explore similar work

Detecting Unfaithful Chain-of-Thought via Circuit-Guided Internal-External Discrepancy

Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models?

Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning