acceptodds
Under review as a conference paper at ICLR 2027

Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation

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.

Primary area:
interpretability and explainable AI
Keywords:
chain-of-thought, faithfulness, interpretability, reasoning, test-time intervention, question decomposition

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_hCvyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CausalCoT to address the unreliability of reasoning traces by decomposing the input into a DAG of sub-questions, each solved locally using only its direct causal dependencies, thereby preventing superficial justifications in traditional chain-of-thought prompting.

Critical strengths

Unfaithful reasoning is a meaningful problem that the paper tries to solve. The method is tested on both open-source and closed-source models of different sizes. The paper presents 3 well-motivated research questions and reports results honestly.

Critical weaknesses

The scope is narrow, as the input needs to be decomposable into a DAG. Building a correct DAG also requires substantial iterative attempts with the LLM, and malformed DAGs can lead to degradation in performance

Performance is weak, as some questions might require the full context to answer and CausalCoT only focuses on local reasoning. The claim that multiple attempts might improve CausalCoT's performance is reasonable, but it still underperforms existing methods while introducing substantial computational overhead.

The faithfulness measurement based on answer changes and prediction changes is not fully reliable. Following the injected text does not necessarily mean the reasoning is faithful, it only shows that the model is responsive to the given text rather than its own reasoning. It also penalizes self-correction, where the model tries to correct the injected error.

Factored Decomp works fine and might achieve a more significant accuracy improvement than CausalCoT when increasing the number of interventions from 0 to 3 on MedQA and Counterfactuals.

There is no comparison of computational overhead or latency, which raises questions about the practical use of the method.

Questions for authors

Do you have a reliable way to construct a DAG from the given input?

Comparison of computation and latency could help a lot in understanding the trade-off between reliability and cost.

1 comment

Thank you for recognising the importance of the problem, the well-motivated research questions, the breadth of the evaluation and the honest reporting.

  1. Requiring a DAG does not narrow the scope. Standard CoT is itself a DAG in which every step can access the input and all preceding steps; CausalCoT restricts which dependencies are allowed, not which problems can be posed. A step that needs input information receives it from a root node through an explicit edge. We evaluate on discriminative reasoning, following standard practice (Lanham et al., 2023), since faithfulness is measurable there through answer changes. The method targets settings where a model must justify a decision (e.g., clinical diagnosis, legal eligibility), where post-hoc rationalisation matters.

  2. The graph is built in one zero-shot call, with no iteration. Our wording caused this confusion. The same prompt is used for all models and datasets, and Acc in Table 1 is single-attempt. Pass@3 asks whether any of three independent end-to-end runs succeeds; it is computed for every method and is not part of CausalCoT. On it, CausalCoT matches or exceeds CoT on all seven models. We will replace "repeated structural attempts" in Section 5.1 with "independently sampled runs". Malformed graphs average below 0.3% for six models (3.81% for Nano) and are already counted as errors (Table 6). Imperfect plans do reduce accuracy: in 360 analysed failures (H.9) the main factors are missing model knowledge and sub-questions omitting discriminating information; an incorrect node is visible in 354. On your question: structural validity is checked deterministically; plan quality comes from a prompt and has no formal guarantee, as in CoT, least-to-most and Factored Decomposition (FD). Iterative refinement or planner fine-tuning (Appendix E) could improve accuracy at extra cost; our aim is to show that zero-shot planning already suffices to add faithfulness (Table 2).

  3. There is an accuracy cost, a trade-off commonly accepted in the literature for gains in interpretability and safety. The average loss is 1.05-2.60 points for the four stronger models and 4.78-6.38 for the others. Precedents are concept bottleneck models (Koh et al., 2020) and our baseline FD, which also trades accuracy for faithfulness (Radhakrishnan et al., 2023) and is below CoT on every model in Table 1. The gap comes mainly from plan quality, not missing context. With graphs fixed, restoring context access recovers only 2-4.5 point gap for the weaker models (Table 10), but collapses faithfulness (Mini: Answer Change 58.7% to 37.1%, vs 41.0% for CoT; Table 9). A stronger planner's graphs recover more with isolation intact: Gemma 4B on MedQA rises from 44.7% to 54.7% (Table 7), above its CoT accuracy of 49.0%. The trade improves with capability: on GPT-5.4, a 2.2-point loss comes with Answer Change rising from 27% to 48%.

  4. Overall CausalCoT gains more from corrections than FD, and FD's gains support our thesis. You are right that FD gains more in three of six Counterfactuals settings and two of six on MedQA; we will correct H.1, which suggests otherwise. Across the 18 settings with FD results, CausalCoT gains more in 11 and FD in 6, averaging +22.2 points after three corrections against +18.2. FD's gains are expected: it shares decomposition and restricted context, the ingredients introduced to improve faithfulness, and our results show they also make a method respond to corrections. CoT largely does not for stronger models (GPT-5.4 on Counterfactuals: 50.0% to 51.5%).

  5. Answer Change is the standard metric in this literature (Lanham et al., 2023); its limits are why we proposed Predictable Change. We agree that sensitivity to an edited step is necessary, not sufficient, for faithfulness; Predictable Change adds a check that the answer moves in the direction the edit implies. On self-correction: in an audit of 270 unchanged CoT answers, 20 (7%) contained an explicit, supported correction of the edit; crediting them leaves CausalCoT ahead on average for all three audited models (new appendix). The correction experiments are free of this issue, since ignoring a correct edit is not self-correction, and there CausalCoT gains more than CoT in all 21 settings (Tables 3-5). H.5 further tests standard CoT without injecting an error: the model is given its own unedited reasoning for one input together with a minimally different input, and for GPT-5.4 the final answer follows the input, not the reasoning, in 47% of conflict cases, where by definition CausalCoT scores 0%.

  6. Overhead is already reported in Appendix H.11; we will move the numbers to the limitations. On 50 MedQA questions (GPT-5.4 Mini), CoT takes 1.31 minutes and 50 calls, CausalCoT 3.43 minutes and 237 calls, FD 3.77 minutes and 300 calls. The overhead relative to CoT is real (2.6x runtime) and lower than FD's. Independent nodes can run in parallel, so latency scales with graph depth.

Reviewer_qWmQOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CausalCoT, a framework that decomposes reasoning into a directed acyclic graph and evaluates each sub-question using only its direct parents' answers. The authors evaluate its performance, reasoning faithfulness, and test-time intervention capabilities on several benchmarks.

Critical strengths

The paper addresses the important problem of reasoning faithfulness and provides experiments across multiple models and benchmarks.

Critical weaknesses

The method is relatively simple, mainly combining question decomposition, DAG-based dependency routing, and context isolation. Similar ideas have already been explored in prior work, including Causal-CoT and Factored Decomposition [1, 2]. The paper does not sufficiently establish its novelty beyond these approaches. The substantial overlap with existing work also raises concerns about the originality and positioning of the contribution.

[1] Causal-CoT: Causal Chain-of-Thought for Validated Reasoning. [2] Radhakrishnan et al. Question Decomposition Improves the Faithfulness of Model-Generated Reasoning.

Questions for authors
  1. How does this work differ substantially from the existing Causal-CoT framework and Factored Decomposition?
  2. What is the key technical novelty beyond decomposing questions into a DAG and isolating the context of each reasoning step?
1 comment

Thank you for recognising the importance of reasoning faithfulness and the breadth of our experiments across models and benchmarks. We answer your two questions in turn and then address soundness.
1a. Causal-CoT [1] addresses a different problem with a different mechanism; the overlap is the name and the use of a DAG. [1] builds a causal graph over the events in a problem, adds mediators and confounders, elicits conditional probabilities from the LLM, and applies do-calculus to decide whether a causal claim holds. Its graph models the world, its LLM calls are not context-isolated, and it aims to improve answer accuracy. In our work, nodes are sub-questions and edges specify which earlier answers a step may read: the graph constrains the computation, not the world, and we measure whether the answer depends on the reasoning by intervening on its steps. We separate our approach from SCM-based methods of this kind in footnote 1 and Appendix D. We were unaware of [1]: it was desk-rejected from ICLR 2026 and, to our knowledge, never published, including as a preprint. We will nonetheless cite and discuss it, and if the similarity of names is a source of confusion, we are happy to rename our method.
1b. Factored Decomposition [2] is our main baseline; the difference is detailed in Appendix C and measured in every experiment. In [2], sub-questions are answered in separate contexts, but dependencies between steps are implicit: each answer can reshape the sub-questions that follow, and the final answer is produced with the original question and all sub-questions and sub-answers in context. CausalCoT fixes the dependencies in an explicit graph and gives every node, including the final one, only its parents' answers. CausalCoT's Answer Change is 7 to 35 points higher than FD's on all seven models (Table 2). An ablation isolates the role of the final step: giving CausalCoT's final node the original question lowers Answer Change by 12 to 21 points in all four model-dataset settings tested (Table 7).
Q2. The key technical novelty is enforcing isolation at every node, including the final answer, through an explicit dependency graph, and showing that this is what makes decomposition faithful. Prior decomposition isolates the intermediate steps but lets the final answer read everything, including the input. Closing this path is a small change with a large measured effect (above), and removing isolation on identical graphs returns faithfulness to roughly CoT level on the three models tested (Table 9). Beyond the method, we contribute Predictable Change. The answer-change metric this literature relies on has a confound: an answer may change because the method is unstable, not because it depends on the edited step. Predictable Change tests whether the answer moves in the direction the edit implies. We also contribute a counterfactual swap test showing that standard CoT bypasses its own reasoning more as models become stronger (Appendix H.5), and a study of test-time interventions on reasoning steps.
We agree the method is simple, and we consider this a strength. It needs no training and no access to model internals, and it works unchanged on seven models from three families, including closed APIs. The works that define this area, chain-of-thought and Factored Decomposition itself, are also prompt-level methods; their contribution was evidence about what a simple change does. Ours is that the faithfulness gains of decomposition depend on a design choice that prior work did not make.
Soundness: the review lists novelty as its only weakness, so we ask for the specific concerns behind this score. If there are concerns about our experiments or claims, we would be glad to address them.

Reviewer_cLJoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces CausalCoT, an inference framework that plans a DAG of subquestions, then evaluating them in isolated contexts to enforce causal transparency. The authors also propose Predictable Change, and evaluates accuracy, perturbation responses, and simulated corrections on MedQA, GPQA, and the Counterfactual Simulatability Suite. Results show higher answer-change rates across seven models, with mixed gains in predictability and lower overall single-sample accuracy than standard CoT.

Critical strengths
  1. The method provides a form of reasoning structural control. It makes dependencies explicit and supports local intervention.
Critical weaknesses
  1. The paper uses a planning step to create the DAG, which is claimed for improving faithfulness. This just moves the faithfulness issue to the planner itself and doesn't eliminate the faithfulness concern. There could be cases where the planner itself commits to an unfaithful plan.

  2. Unclear in comparison fairness and metrices. To show faithfulness and predictability, RQ2 perturbs different reasoning units across methods, how comparable the pertubations are remain unclear. It could be the case that one method receive less critical perturbation than the others. The assessment of "Answer Change" as a signal for faithfulness (RQ2) when receiving an injected pertubation in reasoning trace is arguable. If a model can self correct its own thinking during inference by looking at relevant information from the question and other thinking step can we say that it is unfaithful?

  3. CausalCoT uses substantially more compute than CoT, yet substantially underperforms CoT.

Questions for authors
  1. Do the authors have any principled way to improve faithfulness of the planning step? or can the authors add any evaluation to show the faithfulness of this step?

  2. How are intervention steps (in RQ2) selected, and how does the procedure ensure comparable interventions across methods?

  3. As CausalCoT uses substantially more compute than CoT, can the authors show accuracy at comparable inference budget when methods use majority voting of final results?

  4. Although appendix H.8 rates 73–82% of edits as plausibly capable of changing the answer, CausalCoT’s Answer Change for Sonnet 4.6 is 32% (Table 2). If the Markov property guarantees the final prediction is a faithful product of the intermediate steps (Section 3.2, line 215), why does a large share of plausible edits leave the answer unchanged?

1 comment

Thank you for recognising that the method provides structural control over reasoning, makes dependencies explicit and supports local intervention.
1. Planner. The answer depends on which notion of faithfulness is meant: we target post-hoc rationalisation, not counterfactual consistency, and for the former the displayed plan is the one executed. Section 2 distinguishes the two, and Section 7 states that we do not guarantee counterfactual consistency, i.e. that the stated reasoning predicts how the model behaves when the input changes. This limit applies to the planner. Isolation ensures that the answer is computed from the displayed steps: the planner can influence it only through the text of the graph, which is visible and can be inspected or edited before execution. Two results show the plan alone does not provide the answer. With the graph held fixed, editing one node's answer changes the final answer in 32-64% of cases (Table 2). And when GPT-5.4's graphs are executed by Gemma 4B, MedQA accuracy is 54.7%, against 93.5% for GPT-5.4 itself (Tables 1 and 7), so the plan does not carry the answer. On improving this step: non-final nodes are not allowed to commit to an answer or refer to the options, the plan can be audited before execution, and the planner can be fine-tuned with structural constraints (Appendix E). We will present the planner-transfer result as an evaluation of this step.
2. Comparability and the metric. We agree Answer Change can be confounded; it is the standard metric in this literature (Lanham et al., 2023), which is why we add Predictable Change and analyse results by edit position. The perturbed step is drawn at random for every method, and all edits are produced by the same judge. Raters judged edit validity to be similar across methods, within 2-3 items of 30 (Appendix H.8). Predictable Change accounts for how critical an edit is: an unchanged answer counts as correct when the edit should not alter the label. By position, CausalCoT's Answer Change is at or above the baselines' throughout (Figure 7).
Self-correction. Answer Change does count explicit self-correction against CoT; our audit shows it is a small share, and our conclusions hold with it credited and in an experiment free of it. We audited 270 unchanged-answer CoT continuations (three models, three datasets). In 20 (7%) the continuation explicitly identified the edited claim as mistaken and replaced it with a locally supported one, from 2 of 90 for GPT-5.4 Mini to 12 of 90 for Sonnet 4.6. Crediting these raises CoT's Answer Change to an estimated 36%, 32% and 22% for the three models (GPT-5.4 Mini, GPT-5.4 and Sonnet 4.6), against 64%, 48% and 32% for CausalCoT. The intervention experiment (RQ3) is not exposed to this issue: there, the edit is a correction, so ignoring it cannot be self-correction, and CausalCoT gains more than CoT with three corrections in all 21 model-dataset settings (Tables 3-5). We add this caveat to the metric description and report the audit in a new appendix. Appendix H.5 tests standard CoT without injecting any error. We take pairs of minimally different inputs with different answers. The model reasons about the first input; we then show it the second input together with its own unedited reasoning for the first, and ask for the final answer. If the answer came from the reasoning, it would be the first input's answer. GPT-5.4 instead gives the second input's answer in 47% of such cases: the answer is taken from the input, around the displayed reasoning. In those cases, CausalCoT scores 0% by definition. 3. Compute and accuracy. The extra compute buys faithfulness, not accuracy, and voting at a comparable budget does not change the picture. CausalCoT uses 4.7 calls per question and 2.6x CoT's runtime (Appendix H.11) and is less accurate than CoT (Table 1). As requested, we computed majority voting over three CoT samples, which take about as long as one CausalCoT run (3.9 against 3.4 minutes in the setting of H.11). Voting changes CoT's overall accuracy by -0.5 to +1.6 points across the seven models, so at a comparable budget the comparison in Table 1 is essentially unchanged: CoT remains ahead by 1.5 to 7.9 points. Voting does not make any single rationale more faithful.
4. Unchanged answers. A valid edit is one that could plausibly change the answer, not one that must; the main reason it often does not is redundancy between branches. The final node usually has several parents, so unedited branches can visibly outweigh one edited node, as discussed in Appendix H.3. Answer Change rises from roughly 35-50% for the earliest nodes to 70-85% for those closest to the final answer (Figure 7). The statement at line 215 concerns what each node can read, namely only its parents' findings. It does not imply that every edit changes the answer, and we will reword it. On Sonnet 4.6, the same edits change CoT's answer in 10% of cases, against 32% for CausalCoT.

Comments on the paper

1 comment

We thank the reviewers for their time and for recognising the importance of reasoning faithfulness, the breadth of the experiments, the honest reporting, and the structural control and local intervention the method provides. Below is a summary of the main issues, our short answers, and where each is addressed in full.

Main issues

  • Answer Change and self-correction (hCvy20, cLJo). Answer Change is the standard metric in this literature and does count explicit self-correction against CoT. In an audit of 270 unchanged CoT answers, 20 (7%) contain an explicit, supported correction; crediting them leaves CausalCoT ahead on average for all three audited models. The intervention experiment is free of this issue, and there CausalCoT gains more than CoT in all 21 settings. See response to hCvy20 (point 5) and to cLJo (Self-correction); Tables 3-5; new appendix.
  • Novelty with respect to Factored Decomposition and Causal-CoT (qWmQ). FD lets the final answer read the original question and all sub-answers (as well as potentially updating subquestions based on everything else that happened before); CausalCoT isolates every node, including the final one, which raises Answer Change by 7 to 35 points over FD on all seven models. Causal-CoT is a different method (do-calculus over a causal graph of the world, aims to improve accuracy); it was desk-rejected at ICLR 2026, and we found no other version or preprint. See response to qWmQ (Q1a, Q1b, Q2); Tables 2 and 7; Appendix C.
  • Accuracy cost (hCvy20, cLJo). The average loss against CoT is 1.05-2.60 points on the four stronger models and 4.78-6.38 on the others. This is a common trade-off for gains in interpretability and faithfulness, accepted in prior work such as concept bottleneck models (Koh et al., 2020) and our baseline Factored Decomposition (Radhakrishnan et al., 2023). It comes mainly from plan quality and capabilities, not from missing context. See response to hCvy20 (point 3) and to cLJo (point 3); Tables 1, 7 and 10.
  • Compute overhead and equal-budget voting (hCvy20, cLJo). CausalCoT uses 4.7 calls per question and 2.6x CoT's runtime, below FD. Majority voting over three CoT samples changes CoT's accuracy by -0.5 to +1.6 points, so the comparison is unchanged at a comparable budget. See response to hCvy20 (point 6) and to cLJo (point 3); Appendix H.11; new appendix.
  • Building the graph and the planner's faithfulness (hCvy20, cLJo). The graph is built in one zero-shot call with one prompt for all models and datasets; malformed graphs are below 0.3% for six of seven models. We target post-hoc rationalisation, not counterfactual consistency: the displayed plan is the one executed. The plan does not carry the answer: when GPT-5.4's graphs are executed by Gemma 4B, MedQA accuracy is 54.7%, far below the 93.5% GPT-5.4 reaches with the same graphs. See response to hCvy20 (point 2) and to cLJo (point 1); Tables 6 and 7.
  • Comparability of perturbations (cLJo). We acknowledge that Answer Change can be confounded, which is why we propose Predictable Change as a new metric and analyse results by the position of the edited step. The perturbed step is drawn at random for every method, edits come from the same judge, and raters judged edit validity to be similar across methods. See response to cLJo (point 2); Appendix H.8; Figures 7 and 8.
  • Unchanged answers under CausalCoT (cLJo). A valid edit can be outweighed by unedited branches; Answer Change rises from roughly 35-50% for the earliest nodes to 70-85% for those closest to the final answer. See response to cLJo (point 4); Appendix H.3; Figure 7.
  • Interventions compared with FD (hCvy20). Both decomposition methods respond to corrections and standard CoT largely does not. Across the 18 settings with FD results, CausalCoT gains more in 11 and FD in 6. See response to hCvy20 (point 4); Tables 3-5.
  • Scope (hCvy20). Requiring a DAG does not restrict the inputs, since standard CoT is itself a fully connected DAG. We evaluate on discriminative tasks, following standard practice. See response to hCvy20 (point 1); Appendix J.

Changes to the paper

  • A caveat on self-correction in the description of the faithfulness metrics, and a new appendix reporting the audit.
  • A new appendix reporting majority voting for CoT.
  • A summary of the overhead analysis moved to the limitations.
  • Rewording of the sentence at line 215, and of Appendix H.1 on the comparison with FD.
  • Citation and discussion of Causal-CoT in the related work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.