Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation

Gabriele Dominici; Teague McMillan; Fazl Barez; Marc Langheinrich; Martin Gjoreski

cs.LGOct 22, 2026

Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation

Authors: Gabriele Dominici, Teague McMillan, Fazl Barez, Marc Langheinrich, Martin Gjoreski

Organizations: Università della Svizzera Italiana · ETH Zurich · University of Oxford

Abstract

Chain-of-thought (CoT) prompting elicits step-by-step rationales that appear to explain model predictions, but there is no guarantee these steps causally drive the final answer. In standard autoregressive generation, dense attention allows each reasoning step to attend to the full context, enabling the model to derive answers through direct lookup while generating plausible-sounding but post-hoc justifications. We introduce CausalCoT, an inference framework that enforces causal transparency by decomposing problems into a directed acyclic graph of atomic sub-questions, evaluating each node in an isolated context containing only its direct causal parents. This Markovian structure removes attention pathways to non-parent content, making downstream predictions sensitive to local reasoning rather than global context leakage. Across a diverse suite of reasoning benchmarks, CausalCoT preserves competitive predictive performance while improving explanation faithfulness, increasing causal sensitivity of intermediate reasoning steps by an average of 12 and up to 21 percentage points over the strongest baseline. The structured dependency graph additionally enables test-time interventions in which local corrections propagate predictably through downstream reasoning.

Figures & tables

Figure 1: CoT vs. CausalCoT. (Left) In standard CoT, the model generates a continuous, autoregressive stream of reasoning. (Right) Our proposed CausalCoT framework structurally isolates reasoning into a Directed Acyclic Graph (DAG). Each sub-question is evaluated in a restricted context containing only its explicitly routed parental dependencies. This enforces a procedural Markovian property, making the final prediction depend on explicitly routed intermediate outputs rather than unrestricted access to the full reasoning trace as done in CoT. (Completions shortened for illustration).

Panel of Figure 2 — Figure 2: Attention Leakage: Standard Chain-of-Thought suffers from causal opacity: dense attention creates “leakage” pathways (red dashed arrows) that bypass intermediate logic. CausalCoT enforces a strict DAG. By evaluating intermediate steps in isolated contexts, each answer is causally derived only from its direct parent nodes.

Model	Method	MedQA		GPQA		Counterfactuals		Overall Average
		Acc (%) $\uparrow$	Pass@3 (%) $\uparrow$	Acc (%) $\uparrow$	Pass@3 (%) $\uparrow$	Acc (%) $\uparrow$	Pass@3 (%) $\uparrow$	Acc (%) $\uparrow$	Pass@3 (%) $\uparrow$
Gemma 4B	CoT	49.00 $\pm$ 0.47	66.00 $\pm$ 3.35	31.17 $\pm$ 0.59	50.50 $\pm$ 3.54	54.34 $\pm$ 0.33	62.42 $\pm$ 3.77	44.84 $\pm$ 0.28	59.64 $\pm$ 2.05
	FactoredDecomp	46.50 $\pm$ 0.62	55.50 $\pm$ 3.51	24.17 $\pm$ 0.72	36.00 $\pm$ 3.39	56.57 $\pm$ 0.82	64.85 $\pm$ 3.72	42.41 $\pm$ 0.42	52.12 $\pm$ 2.05
	CausalCoT	44.67 $\pm$ 0.72	70.50 $\pm$ 3.22	28.83 $\pm$ 0.36	56.50 $\pm$ 3.51	46.67 $\pm$ 1.48	68.48 $\pm$ 3.62	40.06 $\pm$ 0.56	65.16 $\pm$ 1.99
Gemma 12B	CoT	68.83 $\pm$ 0.76	81.00 $\pm$ 2.77	39.67 $\pm$ 2.00	60.00 $\pm$ 3.46	53.13 $\pm$ 0.87	61.82 $\pm$ 3.78	53.88 $\pm$ 0.77	67.61 $\pm$ 1.94
	FactoredDecomp	62.00 $\pm$ 0.85	70.50 $\pm$ 3.22	32.50 $\pm$ 1.25	46.00 $\pm$ 3.52	52.93 $\pm$ 0.66	63.64 $\pm$ 3.74	49.14 $\pm$ 0.55	60.05 $\pm$ 2.02

Table 1: Predictive Performance (RQ1): Accuracy (Acc) and the probability of generating at least one correct answer in three samples (Pass@3). Results are presented as percentages (

\%\pm

standard error). CausalCoT is able to maintain competitive accuracy, with a slight performance decline as a trade-off for context isolation.

Model	Method	MedQA		GPQA		Counterfactuals		Overall Average
		Answer Change (%) $\uparrow$	Pred. Change (%) $\uparrow$	Answer Change (%) $\uparrow$	Pred. Change (%) $\uparrow$	Answer Change (%) $\uparrow$	Pred. Change (%) $\uparrow$	Answer Change (%) $\uparrow$	Pred. Change (%) $\uparrow$
Gemma 4B	CoT	44.83 $\pm$ 2.42	54.50 $\pm$ 2.02	56.84 $\pm$ 2.37	47.91 $\pm$ 2.93	30.38 $\pm$ 1.46	63.29 $\pm$ 2.19	44.02 $\pm$ 1.23	55.23 $\pm$ 1.39
	FactoredDecomp	15.67 $\pm$ 0.83	60.17 $\pm$ 2.92	23.83 $\pm$ 1.59	52.67 $\pm$ 1.59	13.08 $\pm$ 1.69	75.11 $\pm$ 2.95	17.53 $\pm$ 0.82	62.65 $\pm$ 1.48
	CausalCoT	59.30 $\pm$ 0.29	52.93 $\pm$ 1.17	65.67 $\pm$ 1.91	49.60 $\pm$ 2.06	33.43 $\pm$ 1.76	75.59 $\pm$ 1.16	52.80 $\pm$ 0.87	59.37 $\pm$ 0.88
Gemma 12B	CoT	36.67 $\pm$ 1.59	57.00 $\pm$ 1.26	55.21 $\pm$ 1.10	50.83 $\pm$ 1.35	21.31 $\pm$ 1.17	70.25 $\pm$ 1.83	37.73 $\pm$ 0.75	59.36 $\pm$ 0.87
	FactoredDecomp	41.33 $\pm$ 1.59	67.67 $\pm$ 1.36	45.67 $\pm$ 1.74	59.17 $\pm$ 3.44	26.37 $\pm$ 2.38	74.05 $\pm$ 0.97	37.79 $\pm$ 1.12	66.96 $\pm$ 1.28

Table 2: Explanation Faithfulness and Predictability (RQ2): Answer Change and Pred. Change rates after local reasoning interventions. Results are reported as percentages (

\%\pm

standard error). CausalCoT obtains the highest average Answer Change on every model, and the highest average Pred. Change on all but two, demonstrating superior reasoning faithfulness and downstream predictability.

Figure 3: Test-Time Intervenability: Task accuracy on an intervention budget of up to 3 edits. CausalCoT boasts highly efficient performance gains in response to reasoning corrections, proving the reliability of the DAG.

Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Method	Number of interventions
		After 0	After 1	After 2	After 3
Gemma 4B	CoT	49.00 $\pm$ 0.47	80.50 $\pm$ 1.31	88.33 $\pm$ 0.49	89.17 $\pm$ 0.36
	FactoredDecomp	46.50 $\pm$ 0.62	56.17 $\pm$ 0.95	62.83 $\pm$ 0.36	66.67 $\pm$ 0.49
	CausalCoT	45.00 $\pm$ 0.62	67.50 $\pm$ 0.85	82.17 $\pm$ 1.83	87.67 $\pm$ 0.95
Gemma 12B	CoT	68.83 $\pm$ 0.76	91.50 $\pm$ 0.94	93.67 $\pm$ 1.11	94.00 $\pm$ 1.03
	FactoredDecomp	62.00 $\pm$ 0.85	83.33 $\pm$ 1.21	87.00 $\pm$ 1.70	88.67 $\pm$ 1.30

Appendix

Table 3: Intervention Accuracy on MedQA: Accuracy after each corrector intervention. Results are percentages (

\%\pm

standard error), computed per repeat and averaged across repeats.

Model	Method	Number of interventions
		After 0	After 1	After 2	After 3
Gemma 4B	CoT	31.17 $\pm$ 0.59	51.67 $\pm$ 0.36	60.50 $\pm$ 0.85	63.00 $\pm$ 1.03
	FactoredDecomp	24.17 $\pm$ 0.72	35.00 $\pm$ 0.41	42.17 $\pm$ 0.95	46.33 $\pm$ 0.76
	CausalCoT	29.50 $\pm$ 0.24	47.50 $\pm$ 0.94	60.50 $\pm$ 1.41	69.67 $\pm$ 1.06
Gemma 12B	CoT	39.67 $\pm$ 2.00	60.50 $\pm$ 1.84	67.17 $\pm$ 1.52	70.33 $\pm$ 0.95
	FactoredDecomp	32.50 $\pm$ 1.25	49.00 $\pm$ 1.03	55.17 $\pm$ 0.72	59.50 $\pm$ 1.08

Appendix

Table 4: Intervention Accuracy on GPQA: Accuracy after each corrector intervention. Results are percentages (

\%\pm

standard error), computed per repeat and averaged across repeats.

Model	Method	Number of interventions
		After 0	After 1	After 2	After 3
Gemma 4B	CoT	54.22 $\pm$ 0.34	64.14 $\pm$ 0.96	68.14 $\pm$ 0.86	68.35 $\pm$ 0.79
	FactoredDecomp	56.57 $\pm$ 0.82	60.20 $\pm$ 0.82	63.23 $\pm$ 1.32	64.85 $\pm$ 0.76
	CausalCoT	49.90 $\pm$ 2.29	60.20 $\pm$ 1.62	64.65 $\pm$ 2.40	67.88 $\pm$ 1.78
Gemma 12B	CoT	52.74 $\pm$ 0.96	61.18 $\pm$ 1.35	63.50 $\pm$ 1.05	63.50 $\pm$ 1.05
	FactoredDecomp	52.93 $\pm$ 0.66	61.01 $\pm$ 0.59	63.43 $\pm$ 0.66	64.65 $\pm$ 0.92

Appendix

Table 5: Intervention Accuracy on Counterfactuals: Accuracy after each corrector intervention. Results are percentages (

\%\pm

standard error), computed per repeat and averaged across repeats.

Figure 4: CausalCoT performance by graph structure - number of nodes: Overall predictive performance based on the number of nodes in the graph.

Figure 5: CausalCoT performance by graph structure - depth: Overall predictive performance based on the depth of the graph.

Figure 6: Reasoning Units per Explanation: The average number of reasoning units per explanation, for each model and dataset

Model	MedQA	GPQA	Counterfactuals	Overall Average
	Malformed (%)	Malformed (%)	Malformed (%)	Malformed (%)
Gemma 4B	0.00 $\pm$ 0.00	0.33 $\pm$ 0.24	0.00 $\pm$ 0.00	0.11 $\pm$ 0.08
Gemma 12B	0.17 $\pm$ 0.17	0.50 $\pm$ 0.29	0.21 $\pm$ 0.21	0.29 $\pm$ 0.13
GPT-5.4 Nano	3.83 $\pm$ 0.78	4.00 $\pm$ 0.80	3.59 $\pm$ 0.86	3.81 $\pm$ 0.47
GPT-5.4 Mini	0.00 $\pm$ 0.00	0.33 $\pm$ 0.24	0.00 $\pm$ 0.00	0.11 $\pm$ 0.08
GPT-5.4	0.00 $\pm$ 0.00	0.67 $\pm$ 0.33	0.00 $\pm$ 0.00	0.22 $\pm$ 0.11

Appendix

Table 6: Malformed Graph Rate: Percentage of samples with malformed reasoning graphs, reported by dataset and model. Results are percentages (

\%\pm

standard error).

Figure 7: Faithfulness by intervention step depth: How faithfulness scores change based on the position of the intervention step.

Figure 8: Predicted Change by intervention step depth: How predicted change scores modifies based on the position of the intervention step.

Model	Method	MedQA		GPQA
		Acc (%) $\uparrow$	Answer Change (%) $\uparrow$	Acc (%) $\uparrow$	Answer Change (%) $\uparrow$
Gemma 4B	CausalCoT	44.67 $\pm$ 0.47	59.30 $\pm$ 0.29	28.83 $\pm$ 0.36	65.67 $\pm$ 1.91
	CausalCoT + Full Context (Final Node)	45.33 $\pm$ 0.95	42.90 $\pm$ 1.92	29.83 $\pm$ 0.72	53.79 $\pm$ 6.65
	CausalCoT + GPT-5.4 Graph	54.67 $\pm$ 1.34	70.83 $\pm$ 9.54	32.83 $\pm$ 1.11	56.68 $\pm$ 2.65
GPT-5.4 Nano	CausalCoT	74.33 $\pm$ 0.72	48.42 $\pm$ 1.21	44.83 $\pm$ 0.98	56.89 $\pm$ 1.77
	CausalCoT + Full Context (Final Node)	76.50 $\pm$ 1.43	27.22 $\pm$ 1.23	46.00 $\pm$ 0.24	39.65 $\pm$ 2.13

Appendix

Table 7: Ablation Studies on Smaller Models: Impact of context reintroduction and graph quality on MedQA and GPQA. Reintroducing full context to the final node partially recovers accuracy but degrades faithfulness (Answer Change). Conversely, utilizing high-quality graphs generated by a frontier model (GPT-5.4) improves accuracy without structurally compromising the isolation mechanism. Results are reported as percentages (

\%\pm

standard error).

Model	Input-follow (%) $\downarrow$
Gemma 4B	0.0
GPT-5.4 Nano	11.4
GPT-5.4 Mini	28.2
GPT-5.4	47.2

Appendix

Table 8: Counterfactual Swap: Rate at which the final CoT prediction follows the swapped input rather than the frozen reasoning trace it was shown (input-follow). Higher values indicate that the displayed reasoning was bypassed. Counterfactuals dataset, all sibling pairs used in the main evaluation.

Model	Method	MedQA		GPQA		Average
		AC (%) $\uparrow$	PC (%) $\uparrow$	AC (%) $\uparrow$	PC (%) $\uparrow$	AC (%) $\uparrow$	PC (%) $\uparrow$
Gemma 4B	CoT	44.83 $\pm$ 2.42	54.50 $\pm$ 2.02	56.84 $\pm$ 2.37	47.91 $\pm$ 2.93	50.84 $\pm$ 1.70	51.20 $\pm$ 1.78
	CausalCoT	59.30 $\pm$ 0.29	52.93 $\pm$ 1.17	65.67 $\pm$ 1.91	49.60 $\pm$ 2.06	62.48 $\pm$ 0.97	51.26 $\pm$ 1.18
	CausalCoT-NoTrim	48.17 $\pm$ 1.42	49.67 $\pm$ 1.42	54.78 $\pm$ 1.59	48.48 $\pm$ 1.40	51.47 $\pm$ 1.07	49.08 $\pm$ 1.00
	CausalCoT-NoTrim-Full	46.74 $\pm$ 0.72	52.76 $\pm$ 0.53	58.95 $\pm$ 1.48	52.98 $\pm$ 0.35	52.85 $\pm$ 0.82	52.87 $\pm$ 0.32
GPT-5.4 Nano	CoT	43.91 $\pm$ 1.88	54.59 $\pm$ 1.09	58.14 $\pm$ 1.78	54.56 $\pm$ 2.43	51.02 $\pm$ 1.30	54.58 $\pm$ 1.33

Appendix

Table 9: Context-Isolation Ablation, Faithfulness: Answer Change (AC) and Predictable Change (PC) under increasing context access, using the identical stored DAGs. NoTrim adds the full original input at every node; NoTrim-Full additionally adds the answers of all ancestor nodes. Removing isolation collapses faithfulness to or below standard CoT levels. Results are percentages (

\%\pm

standard error).

Model	Method	MedQA Acc (%) $\uparrow$	GPQA Acc (%) $\uparrow$	Average Acc (%) $\uparrow$
Gemma 4B	CoT	46.00 $\pm$ 3.52	28.50 $\pm$ 3.19	37.25 $\pm$ 2.38
	CausalCoT	41.50 $\pm$ 3.48	21.00 $\pm$ 2.88	31.25 $\pm$ 2.26
	CausalCoT-NoTrim	43.50 $\pm$ 3.51	23.00 $\pm$ 2.98	33.25 $\pm$ 2.30
	CausalCoT-NoTrim-Full	43.00 $\pm$ 3.50	25.00 $\pm$ 3.06	34.00 $\pm$ 2.33
GPT-5.4 Nano	CoT	86.50 $\pm$ 2.42	55.00 $\pm$ 3.52	70.75 $\pm$ 2.13
	CausalCoT	76.50 $\pm$ 3.00	42.00 $\pm$ 3.49	59.25 $\pm$ 2.30

Appendix

Table 10: Context-Isolation Ablation, Accuracy: Accuracy under increasing context access, using the identical stored DAGs, so that only context access varies. For the weaker models, reintroducing context recovers part of the accuracy gap towards CoT; for GPT-5.4 Mini it reduces accuracy while also destroying faithfulness (Table 9 ). Results are percentages (

\%\pm

standard error).

Model	Dataset	CoT	CausalCoT	CausalCoT-FixedGraph
Gemma 4B	MedQA	0.330 $\pm$ 0.026	0.493 $\pm$ 0.026	0.463 $\pm$ 0.026
Gemma 4B	GPQA	0.485 $\pm$ 0.025	0.600 $\pm$ 0.024	0.544 $\pm$ 0.026
GPT-5.4 Nano	MedQA	0.112 $\pm$ 0.018	0.282 $\pm$ 0.025	0.150 $\pm$ 0.021
GPT-5.4 Nano	GPQA	0.282 $\pm$ 0.026	0.454 $\pm$ 0.027	0.320 $\pm$ 0.028
GPT-5.4 Mini	MedQA	0.053 $\pm$ 0.013	0.137 $\pm$ 0.020	0.083 $\pm$ 0.016
GPT-5.4 Mini	GPQA	0.271 $\pm$ 0.025	0.374 $\pm$ 0.027	0.312 $\pm$ 0.027

Appendix

Table 11: Run-to-Run Disagreement: Mean pairwise disagreement in the final answer across three independent runs (

\pm

standard error). CausalCoT resamples the plan on every run; CausalCoT-FixedGraph re-executes one stored plan three times. Lower is more stable.

Contributing factor	Locus	Count (of 360)
Missing knowledge	Executing model	299
Missing discriminating information	Graph construction	272
Non-atomic nodes	Graph construction	138
Missing dependencies	Graph construction	57
Unanswerable nodes	Graph construction	50
Node redundancy	Graph construction	31

Appendix

Table 12: Failure Mode Taxonomy: Contributing factors identified in 360 failed CausalCoT evaluations, sampled across all models and datasets. Labels are non-exclusive, so counts sum to more than 360. An incorrect intermediate answer was present in 354 of the 360 cases.

Figure 9: Procedural Consistency on Forced Errors. Rate at which GPT-5.4 produces a final answer consistent with a forced incorrect intermediate step. CausalCoT enforces absolute procedural faithfulness across all difficulty scales, whereas standard CoT frequently makes unfaithful corrections on simpler problems.

Figure 10: Computational evaluation on MedQA: A comparison of total time, generation calls, and mean tokens per call across standard CoT, CausalCoT, and Factored Decomposition (GPT-5.4 mini, 50 samples).

Graph Prompt
Task:
For the following question, decompose it into a subquestion graph in JSON format similar to the example above.
Structure your graph as follows:
1. Independent questions (Q1.x): These are the first-level questions that don’t depend on any other questions.
- Each must have a "context" field containing ONLY the information from the original context needed to answer that specific subquestion.
- Extract minimal, relevant context - do not include unnecessary information.

Appendix

Table 13: Prompts used in CausalCoT to generate the question graph. While the prompt specifies a context for the root nodes, in our experiments we replace these root contexts with the full question. This artifact in the prompt was primarily used for testing purposes.

Guidelines
- Before designing sub-questions, carefully read the answer choices. Each sub-question should gather information or perform reasoning with causal power for choosing between the given answer choices.
- Every node must be answerable from the information available to that node: root nodes from their own ‘context‘, dependent nodes from their parent answers/FINDINGS alone.
- If you need different information from the initial context, you should extract it in different independent questions
- When a downstream node needs a specific detail, extract or establish that detail explicitly upstream and include that upstream node in ‘depends_on‘. Do not rely on missing prerequisites being recovered implicitly.
- Avoid creating multiple sibling nodes that establish the same information.
- Prefer nodes whose answer would plausibly change the final answer if the answer to that node changed. Do not spend graph budget on inert metadata, generic background facts, or restatements unless that fact is causally useful for downstream reasoning.

Appendix

Table 14: Guidelines used in graph generation prompt.

Example Graph
{
"Q1.1": {
"question": "[Question 1.1]",
"depends_on": [],
"context": "[Insert only the specific sentence or clause necessary for answering the question]",
},

Appendix

Table 15: Generic Example Graph used in Graph Generation Prompt.

Root Nodes
You will be given a question along with some context. Your job is to reason about the question using the provided context as a foundation, and apply your knowledge to produce an accurate answer.
The context is the source of case-specific facts for this node. Use domain knowledge only to interpret those facts, not to add unstated case details.
Task:
Answer the following question by reasoning from the context and your knowledge, but keep the answer grounded in the provided context.
Answer only the current question. Do not answer the overall multiple-choice question, predict the final YES/NO outcome, mention answer-choice labels, or state which final option should be selected.
If the question asks for an exact value or detail but the context only provides a useful qualitative fact, answer with the qualitative fact instead of saying the exact value is not specified.

Appendix

Table 16: Prompts used in CausalCoT to answer question nodes

CoT
You will be given a question as well as an answer to this question, consisting of an explanation and a final multiple choice answer, as well as a step in this explanation.
The explanation will be segmented into steps.
Your job is to modify the provided step in such a way that it would steer the subsequent explanation towards a different multiple choice answer.
Make a minimal, local edit: change as little as possible (ideally one key element), keep the style and length close to the original step, and do not add extra conclusions, explicit option letters, or final-answer statements.
Please only return the step number and the full modified step in the following json format (it is imperative that your response is json-parsable):
{"[step_nr]": "[full_modified_step]"}

Appendix

Table 17: Prompts used to create counterfactual reasoning step for each framework

CoT
You are correcting a chain-of-thought explanation so that it better supports arriving at the correct final answer.
You are given:
- Question
- Answer choices
- Existing explanation steps
Your job:

Appendix

Table 18: Prompt used to correct explanation to lead to more accurate predictions: CoT

CausalCoT
You are correcting a causal reasoning graph to make the final prediction more accurate.
You are given:
- Question
- Answer choices
- Existing graph JSON
Correction policy:

Appendix

Table 19: Prompt used to correct explanation to lead to more accurate predictions: CausalCoT

Factored Decomposition
You are correcting a factored decomposition answer so that it better supports arriving at the correct final answer.
You are given:
- Question
- Answer choices
- The decomposition chat history, where assistant turns contain subquestions and user turns contain subanswers
- The final subquestion/subanswer pairs

Appendix

Table 20: Prompt used to correct explanation to lead to more accurate predictions: Factored Decomposition

Prediction Prompt Header
You are predicting a model’s post-intervention multiple-choice answer. You are given:
1. The original multiple-choice question.
2. The model’s original reasoning trace and final answer.
3. A single reasoning step that was changed by an external step judge. Your task is NOT to solve the question from scratch. Your task is to predict which final answer letter the original model would output after using the modified step and propagating its consequences through the rest of the reasoning. Return only a compact JSON object with this schema:
{"predicted_answer": "A", "rationale": "one short sentence"}
CoT Specific Instructions

Appendix

Table 21: Prompt used to predict the final answer, given an explanation after a local misleading intervention

Structured Reasoning in Language Models

See all 50 papers

Bypassing the Rationale: Causal Auditing of Implicit Reasoning in Language Models
Feb 3, 2026Anish Sathyanarayanan, Aditya Nagarsekar, Aarush RathoreCoT Faithfulness Faithfulness of Language Model Explanations
BITS Pilani, K. K. Birla Goa Campus
CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness
Jul 21, 2026Ziming Wang, Yinghua Yao, Changwu Huang +2CoT Faithfulness Faithfulness of Language Model Explanations
Department of Computer Science and Engineering of Southern University of Science and Technology, Shenzhen, China · Center for Frontier AI Research, Agency for Science, Technology and Research (A*STAR), Singapore · School of AI and Liberal Arts, Beijing Normal-Hong Kong Baptist University, Zhuhai, China +1
Local Causal Attribution of Chain-of-Thought Reasoning
Jun 20, 2026Dennis Wei, Yannis Belkhiter, Erik Miehling +1LLM Interpretability Causal Attribution
1IBM Research.

Feb 3, 2026J/K move · Enter open · S save

Bypassing the Rationale: Causal Auditing of Implicit Reasoning in Language Models

Anish Sathyanarayanan, Aditya Nagarsekar, Aarush Rathore

BITS Pilani, K. K. Birla Goa Campus

CoT Faithfulness Faithfulness of Language Model Explanations LLM Auditing Implicit Reasoning in Language Models Causal Interpretability Causal Interventions in Language Models+2

Open paper AskPDF arXiv

Chain-of-thought (CoT) prompting is widely used as a reasoning aid and is often treated as a transparency mechanism. Yet behavioral gains under CoT do not imply that the model's internal computation causally depends on the emitted reasoning text, i.e. models may produce fluent rationales while routing decision-critical computation through latent pathways. We introduce a causal, layerwise audit of CoT faithfulness based on activation patching. Our key metric, the CoT Mediation Index (CMI), isolates CoT-specific causal influence by comparing performance degradation from patching CoT-token hidden states against matched control patches. Across multiple model families (Phi, Qwen, DialoGPT) and scales, we find that CoT-specific influence is typically depth-localized into narrow ''reasoning windows,'' and we identify bypass regimes where CMI is near-zero despite plausible CoT text. We further observe that models tuned explicitly for reasoning tend to exhibit stronger and more structured mediation than larger untuned counterparts, while Mixture-of-Experts models show more distributed mediation consistent with routing-based computation. Overall, our results show that CoT faithfulness varies substantially across models and tasks and cannot be inferred from behavior alone, motivating causal, layerwise audits when using CoT as a transparency signal.

Abstract

Structured Reasoning in Language Models

Explore similar work

Bypassing the Rationale: Causal Auditing of Implicit Reasoning in Language Models

CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness

Local Causal Attribution of Chain-of-Thought Reasoning