acceptodds
Under review as a conference paper at ICLR 2027

Reasoning in Two Dimensions: Quantifying Chain-of-Thought Faithfulness and Efficiency

Abstract

Chain-of-thought (CoT) reasoning exposes intermediate steps, but a correct answer does not establish that an LLM relies on the dependence expressed by those steps or reasons efficiently. We operationalize CoT faithfulness and efficiency within a common framework. Specifically, we represent a CoT as a semantic reasoning graph and decompose the model's predictive dependence along two complementary axes: dependence explained by semantic support versus additional reasoning history, and dependence allocated to answer-reaching versus off-path reasoning. This decomposition yields global and step-level measures of unfaithfulness and inefficiency. Controlled interventions show that the proposed metrics respond to hidden dependence and induced off-path reasoning as intended. Across mathematical and multi-hop reasoning tasks, answer-level unfaithfulness is consistently associated with greater predictive uncertainty, whereas both unfaithfulness and inefficiency exhibit weak or inconsistent associations with answer correctness. We further derive theoretically grounded, graph-aware fine-tuning objectives that improve the targeted reasoning properties without directly optimizing the computationally expensive metrics or compromising answer accuracy. Results indicate that accuracy and brevity provide incomplete assessments of reasoning quality: reliable CoT evaluation should also consider whether the LLM's predictive dependence is faithfully represented and efficiently allocated.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Chain-of-Thought, Reasoning Faithfulness, Reasoning Efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KdybOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a unified information-theoretic framework for evaluating chain-of-thought reasoning along two dimensions: faithfulness and efficiency. It represents a CoT as a semantic reasoning graph, measures whether the model’s predictive dependence is explained by the graph’s semantic support, and distinguishes dependence allocated to answer-reaching versus off-path reasoning. The resulting global and step-level metrics are validated through controlled interventions, analyzed in relation to uncertainty and correctness, and further used to construct graph-aware RFT objectives intended to improve reasoning faithfulness and efficiency.

Critical strengths
  1. This paper proposes a unified framework.
  2. It discusses a important issue.
  3. A new method is proposed to solve the problem.
Critical weaknesses
  1. Limited scientific and mechanistic insight. The paper mainly defines quantitative measures of faithfulness and efficiency and studies their associations with uncertainty and correctness, but provides little explanation of why unfaithful or inefficient reasoning arises or what internal mechanisms produce these behaviors.

  2. The technical novelty appears limited and lies mainly in integration and reformulation. Semantic dependency graphs, CoT faithfulness analysis, behavioral dependence measures, redundant-step detection, reasoning pruning, and faithfulness-oriented fine-tuning have all been explored in prior work. The main contribution is to combine these ingredients into a unified two-dimensional decomposition rather than introduce a fundamentally new reasoning mechanism or learning method.

  3. Empirical generalization is limited. The evaluation uses only two Qwen3 variants from the same model family and three relatively structured reasoning benchmarks. It remains unclear whether the proposed metrics and observed relationships generalize to other reasoning model families, frontier systems, or more diverse settings such as coding, planning, and agentic tasks.

  4. The framework critically depends on an LLM-generated semantic reasoning graph whose reliability is insufficiently validated. The semantic graph determines which dependencies are considered faithful, answer-reaching, or off-path, yet it is extracted by a single external LLM. The paper does not provide strong human validation, inter-annotator agreement, or robustness analysis across alternative graph extractors, making it difficult to separate properties of the target model from judgments introduced by the graph annotator.

  5. The framework is evaluated on heavily structured CoTs rather than natural reasoning traces. Models are required to produce explicitly segmented reasoning outputs, which greatly simplifies graph construction. It is unclear whether the same semantic graph extraction and metrics remain reliable for naturally generated free-form CoTs containing implicit dependencies, self-corrections, or more complex reasoning structure.

Questions for authors
  1. What substantive insight does the proposed framework provide about why unfaithful or inefficient reasoning arises, beyond measuring these behaviors?

  2. What fundamentally new capability does the two-dimensional formulation enable compared with existing work on CoT faithfulness, dependency graphs, behavioral importance, and reasoning pruning?

  3. Do the proposed metrics and empirical findings generalize beyond Qwen3 to other model families, frontier reasoning models, or more diverse reasoning tasks?

  4. How robust are the results to errors or variation in the semantic reasoning graph? In particular, would the conclusions remain stable with human annotations or alternative graph-extraction models?

  5. How well does the framework extend to naturally generated free-form CoTs without the imposed reasoning structure?

2 comments
AuthorsOpenReview ↗

Reply to Reviewer Kdyb

Dear Reviewer Kdyb,

Thank you very much for your time and detailed comments. We address your concerns by clarifying the contributions of our work and several potential misunderstandings.

[Weakness#1&2 & Question#1&2. The scope, novelty, and key contributions of our framework.] We would like to clarify the scope of our work. Explaining “why unfaithful or inefficient reasoning arises or what internal mechanisms produce these behaviors,” or introducing a “fundamentally new reasoning mechanism or learning method,” is beyond the stated objectives of this paper. Our focus is to introduce a novel framework for jointly quantifying these properties and demonstrate its reliability and usefulness, which constitutes a research contribution in its own right, as reflected in most of the literature reviewed in Section 2 and Appendix C.

More importantly, we disagree with your comment that our work lacks novelty. Our contribution is not merely to “combine these ingredients into a unified two-dimensional decomposition.” Our mutual information decomposition, grounded in a semantic reasoning graph, provides conceptually and methodologically distinct formulations of both CoT faithfulness and efficiency relative to prior approaches. Please refer to our General Response for details.

[Weakness#3 & Question#3. Experiments covering a wider range of models and tasks.] We thank the reviewer for suggesting experiments across more model–task combinations. Our expanded experiments span MATH, Omni-MATH-2, and 2WikiMultiHopQA, and include Qwen3-30B-A3B-Instruct-2507, Qwen3-4B-Instruct-2507, Ministral-3-8B-Instruct-2512, Ministral-3-3B-Instruct-2512, and Llama-3.1-8B-Instruct.

Given the space constraints, we refer the reviewer to the updated Tables 3–4 and Figure 2, and especially the expanded experimental results in Appendix D.6, Tables 30–52 of the revised manuscript. Overall, these additional results across these diverse settings support our main findings and conclusions from the original submission:

  • Section 4 (Tables 30–33): Controlled interventions produce the expected changes: A-YES increases answer unfaithfulness, S increases global unfaithfulness, and A-NO increases both among cases where the answer changes. On the efficiency side, E reduces global inefficiency and token usage, whereas R increases both.
  • Section 5 (Tables 34-47): Answer unfaithfulness and D1,YτD_{1,Y}^{\tau} consistently show clear positive associations with answer uncertainty. Associations with correctness and those involving efficiency-related measures are generally weaker or less consistent across settings.
  • Section 6 (Tables 48-52): Faithfulness-aware RFT improves its targeted metric relative to Standard RFT without sacrificing Pass@1 accuracy. Note that additional efficiency-aware RFT experiments are still ongoing due to the limited time available. We will add their results to Tables 50 and 51 in the revised manuscript once these experiments are complete.

Please refer to the above tables for detailed numerical results.

[Weakness#4 & Question#4. Robustness to graph extraction.] We thank the reviewer for raising this question. We have added Appendix B.2 to the revised manuscript to examine the sensitivity of our framework to the semantic graph extractor. For Qwen3-30B-A3B-Instruct-2507 on MATH, in addition to DeepSeek V4 Pro, we use GPT-6.1 Terra and Gemini 3.8 Flash to extract graphs from the same responses using the same extraction prompt. Equation (4) defines edge F1 and Jaccard similarity, and Table 12 reports values around 0.9 for both measures, which indicates high agreement across extractors. Table 13 further shows that unfaithfulness and inefficiency scores remain broadly comparable across the three graph annotations. More importantly, Tables 14–20 report the corresponding analysis from Sections 4 and 5 using GPT-6.1 Terra, and the results support all of our main conclusions. Together, these results support the robustness of our framework to the choice of semantic graph extraction model.

[Weakness#5 & Question#5. Extension of our framework to free-form CoTs.] We have further added Appendix B.3 to examine free-form CoTs. In this setting, the model is not required to generate numbered steps or separate Subgoal, Reasoning, and Result fields. Our framework accommodates this setting through a separate annotation stage, in which DeepSeek V4 Pro segments the generated reasoning into steps. Tables 21–27 report the corresponding analyses from Sections 4 and 5 using Qwen3-30B-A3B-Instruct-2507 on MATH, and the results support all of our main conclusions from the original submission. Together, these results show that our framework extends naturally to free-form CoTs and that our main conclusions remain robust in this setting.

We are happy to answer any further questions you may have. If our responses above have addressed your concerns, we would truly appreciate a re-evaluation accordingly.

Reviewer_KdybOpenReview ↗

Keep my score

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns. However, I still have concerns about the validity of the extracted semantic graphs and the limited improvements achieved by faithfulness-aware RFT. While the rebuttal addresses several of my concerns, I remain unconvinced about the overall contribution and maintain my score.

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

This paper proposes an information-theoretic framework for evaluating chain-of-thought (CoT) faithfulness and efficiency. It constructs semantic dependency graphs and compares model predictions under different conditioning contexts to quantify unfaithful and inefficient reasoning. The authors further propose graph-aware fine-tuning methods and evaluate their framework on mathematical and multi-hop reasoning tasks.

Critical strengths
  1. The paper explores an interesting direction by jointly studying CoT faithfulness and efficiency through semantic dependencies and model predictions.

  2. The proposed efficiency-aware fine-tuning shows some improvements in reducing reasoning tokens while maintaining answer accuracy.

Critical weaknesses

Insufficient validation of the proposed metrics. The framework heavily relies on LLM-extracted semantic dependency graphs, but their reliability and robustness are not adequately evaluated. Moreover, it is unclear whether graph-based off-path reasoning accurately reflects causal necessity. Without stronger validation and concrete case studies, it is difficult to assess whether the proposed metrics meaningfully capture faithfulness and efficiency.

Limited experimental comparisons and unclear novelty. The experiments mainly compare against standard RFT, without sufficient comparisons to existing reasoning pruning methods or controlled ablations. The paper could also better discuss relevant prior work, including "Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning" (Yu et al., NeurIPS 2025) and "Understanding Chain-of-Thought in LLMs through Information Theory" (Ton et al., ICML 2025). These limitations make it difficult to establish the incremental contribution of the proposed framework.

Additionally, the presentation needs improvement. The abstract and introduction lack sufficient intuition, making it difficult to understand the proposed decomposition and its practical implications. A concrete end-to-end example would significantly improve clarity.

Questions for authors
  1. Why should the proposed graph-based decomposition accurately reflect reasoning faithfulness and efficiency? In particular, how does graph-based off-path reasoning relate to causal necessity?

  2. How sensitive are the metrics to different semantic graph annotations, reasoning step segmentations, or alternative dependency structures? Is there any human evaluation of graph extraction quality?

  3. Could the authors provide concrete case studies showing the reasoning trace, extracted graph, step-level scores, and corresponding model behavior under interventions?

  4. Can the authors compare against existing reasoning pruning approaches and ensure that all fine-tuning baselines use identical training examples?

  5. Under what assumptions do the restricted-context predictive distributions correspond to the conditional distributions required by the proposed mutual-information decomposition?

4 comments
AuthorsOpenReview ↗

Reply to Reviewer SXpS

Dear Reviewer SXpS,

Thank you very much for your time and feedback. We are glad that you find our research direction interesting and our fine-tuning methods effective. We clarify the contributions of our work and address your concerns and some potential misunderstandings.

[Weakness#1. The novelty and key contributions.] We would like to clarify that our framework is conceptually distinct from prior approaches to CoT faithfulness and efficiency, and that our graph-based mutual information decomposition is novel. We explicitly discuss these distinctions in Section 2, Table 1, and Appendix C. Please also refer to our General Response for further clarification of the novelty and key contributions of our work. Specifically, [a] uses causal sufficiency and necessity to estimate individual steps’ causal influence, focusing on final-answer correctness. Our framework instead quantifies predictive dependence according to semantic support and answer reachability, and is able to directly characterize predictive dependencies among intermediate reasoning steps. Moreover, while [b] also uses mutual information, it measures each step’s information gain for predicting the ground-truth answer using an additionally trained model. Our mutual information decomposition instead characterizes predictive dependence along two dimensions: within versus beyond semantic support, and answer-reaching versus off-path reasoning. It does not require training an additional model. We thank the reviewer for suggesting these additional relevant works and have incorporated them into our revised related work discussion in Section 2 and Appendix C.

[Weakness#2. Clearer presentation with more intuition.] We thank the reviewer for their suggestion to improve the presentation and have revised Section 1 to provide more intuition for our framework.

[Weakness#3 & Question#1. Our faithfulness and efficiency metrics have a principled theoretical interpretation.] Our definitions are grounded in the proposed information-theoretic decomposition. In Section 3.2 (particularly Equation (1), lines 227–236, and Table 2), we explicitly explain why the resulting information quantities naturally characterize CoT faithfulness and efficiency. Please also refer to our General Response for further discussion. Moreover, regarding the relationship between off-path reasoning and causal necessity, off-path reasoning has no path to the final answer in the semantic support graph, whereas causal necessity concerns whether intervening on a step changes the model’s outcome. Our efficiency metric quantifies the proportion of predictive dependence allocated to off-path steps. It does not equate off-path status with the absence of causal influence on the model. Please also see our responses to Weakness#1 and Weakness#5 & Question#4 for the methodological and computational advantages of our framework relative to [a].

[Weakness#4 & Question#2. Robustness to graph extraction and reasoning step segmentation.] We thank the reviewer for raising this question. First, we have added Appendix B.2 to the revised manuscript to examine the sensitivity of our framework to the semantic graph extractor. For Qwen3-30B-A3B-Instruct-2507 on MATH, in addition to DeepSeek V4 Pro, we use GPT-6.1 Terra and Gemini 3.8 Flash to extract graphs from the same responses using the same extraction prompt. Equation (4) defines edge F1 and Jaccard similarity, and Table 12 reports values around 0.9 for both measures, which indicates high agreement across extractors. Table 13 further shows that unfaithfulness and inefficiency scores remain broadly comparable across the three graph annotations. More importantly, Tables 14–20 report the corresponding analysis from Sections 4 and 5 using GPT-6.1 Terra, and the results support all of our main conclusions. Together, these results support the robustness of our framework to the choice of semantic graph extraction model.

Second, we have further added Appendix B.3 to examine reasoning without a prescribed step format. In this setting, the model is not required to generate numbered steps or separate Subgoal, Reasoning, and Result fields. Our framework accommodates this setting through a separate annotation stage, in which DeepSeek V4 Pro segments the generated reasoning into steps. Tables 21–27 report the corresponding analysis from Sections 4 and 5 using Qwen3-30B-A3B-Instruct-2507 on MATH, and the results support all of our main conclusions from the original submission.

Finally, we conducted a small-scale human evaluation covering Qwen3-4B-Instruct-2507 and Qwen3-30B-A3B-Instruct-2507 on MATH, Omni-MATH-2, and 2WikiMultiHopQA. We randomly sampled 10 responses from each model–dataset combination. We then independently constructed semantic dependency graphs, which agreed in 54 out of 60 cases with those extracted by the LLM. This provides additional evidence for the reliability of graph extraction.

AuthorsOpenReview ↗

Reply to Reviewer SXpS – Continued

[Question#3. Concrete case studies under controlled interventions.] We have added Appendix D.7 to the revised manuscript to present two detailed case studies of faithfulness and efficiency interventions. Each includes the original and intervention prompts, model responses, extracted semantic graphs, and step-level and global scores, and shows how the metrics change across conditions.

The faithfulness case illustrates that, even when the visible mathematical derivation remains valid, our metrics can capture changes in unfaithfulness that are not apparent from the reasoning text alone. The efficiency case shows that our metric reflects increases and decreases in redundant reasoning. Besides, although the inefficiency metric does not directly measure token consumption, its changes align with the corresponding changes in token usage.

[Weakness#5 & Question#4. Controlled evaluations and comparisons with existing reasoning pruning methods.] We would first like to clarify that our existing RFT experiments also serve as a component ablation study. Specifically, standard RFT, faithfulness-aware RFT, efficiency-aware RFT, and joint-aware RFT correspond to four combinations of two components: whether to include the regularization in Equation (2) and whether to remove off-path reasoning steps. We also thank the reviewer for suggesting comparisons with [a] and [b]. Please refer to our response to Weakness#1 for the conceptual distinctions between our framework and these works. Regarding comparison with existing pruning methods, the method in [a] requires repeated counterfactual generation and verifier calls for individual reasoning steps. Our efficiency-aware method instead operates on a single generated trajectory and its extracted semantic reasoning graph, and can therefore avoid the substantial computational overhead. We have implemented the pruning procedure in [a] for evaluation on MATH and Omni-MATH-2. However, its generation and verification costs have prevented us from completing the comparison within the limited rebuttal period. We will include the results in the revised version once the experiments are complete.

[Question#5. The assumption underlying our mutual-information decomposition.] Our mutual-information interpretation assumes that predictions obtained by querying the model under restricted contexts equals the corresponding conditional distributions obtained by marginalizing the full-trajectory joint distribution. We would like to emphasize that using predictions under reduced conditioning contexts to estimate conditional or marginal distributions has substantial precedent [c–e]. We thank the reviewer for raising this point and have clarified this assumption and its role in the restricted-context computation in Appendix D.4.

We are happy to answer any further questions you may have. If our responses above have addressed your concerns, we would truly appreciate a re-evaluation accordingly.

References:

[a] Yu et al., "Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning", NeurIPS 2025.
[b] Ton et al., "Understanding Chain-of-Thought in LLMs through Information Theory", ICML 2025.
[c] Fernandes et al., "Measuring and Increasing Context Usage in Context-Aware Machine Translation", ACL 2021.
[d] Malkin et al., "Coherence boosting: When your pretrained language model is not paying enough attention", ACL 2022.
[e] Shi et al., "Trusting Your Evidence: Hallucinate Less with Context-aware Decoding", NAACL 2024.

Thank you for the detailed response and the new analyses. The extractor-sensitivity results (Appendix B.2), the unformatted-reasoning setting (B.3) and the case studies (D.7) address my questions on robustness and clarity. My main concern remains the validity of the metrics. The response states that off-path status does not imply the absence of causal influence, so it is still unclear what the scores capture beyond the graph structure. The assumption that restricted-context predictions equal the required conditional distributions is now stated (D.4) but not tested, the human check covers only 60 graphs, and the comparison with existing pruning methods is not yet available. I have also read the other reviews and the authors' responses to them. The additional models and datasets are welcome, but they do not address the concerns above. I will therefore keep my score.

Thank you for acknowledging that our additional analysis address your concerns about robustness and clarity. We would like to respectfully clarify three points.

First, our experiments do provide evidence for the validity of the proposed metrics. Beyond expanding models and datasets, Section 4 directly validates whether the metrics respond as expected to controlled interventions that induce unfaithful or inefficient reasoning. These findings are consistent across different models, datasets, and graph extractors. We would also like to note that such validation based on controlled interventions has established precedent in the LLM explanation faithfulness literature [f,g]. We believe this provides empirical evidence supporting the validity of our metrics.

Second, the distinction between off-path status and causal necessity is precisely part of our conceptual contribution. Our framework measures how the model’s predictive dependence aligns with the semantic reasoning graph, rather than whether individual steps are causally necessary for the final answer. Importantly, our scores are not determined by graph structure alone: they explicitly incorporate the model’s predictive distributions. Appendix D.7 (Table 54) shows that the same semantic graph can yield substantially different unfaithfulness scores. This distinction from prior work, together with our unified global and step-level formulation, has been explicitly explained in Section 2, Appendix C, and our General Response.

Third, regarding restricted-context predictions, this approach has established precedent in [c–e], which likewise use predictions under reduced contexts without directly verifying their equivalence to exact marginalization. Computing the exact marginal distributions would require integrating over an enormous space of omitted reasoning trajectories. This makes exhaustive evaluation computationally infeasible. We have acknowledged this approximation in Appendix D.4, while providing extensive experiments to validate the behavior of the resulting operational metrics.

We hope these clarifications help address the remaining concerns and would sincerely appreciate your reconsideration.

References:

[f] Zaman and Srivastava, “A Causal Lens for Evaluating Faithfulness Metrics”, EMNLP 2025.
[g] Mittal and Arike, “C2-Faith: Benchmarking LLM Judges for Causal and Coverage Faithfulness in Chain-of-Thought Reasoning”, GEM 2026.

Reviewer_8e5SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an information-theoretic framework to quantify two complementary properties of chain-of-thought (CoT) reasoning: faithfulness (the degree to which the model’s predictive dependence aligns with semantically necessary support) and efficiency (the extent to which predictive dependence is allocated to answer-reaching rather than off-path steps). The authors define a semantic reasoning hypergraph over generated steps, introduce five graph “views” that induce different conditioning structures, and decompose the KL divergence between full-autoregressive and no-history settings into orthogonal dimensions of faithfulness and efficiency with both global and step-level response scores. They validate the metrics via controlled interventions, relate them to uncertainty and correctness, and derive graph-aware rejection-sampling fine-tuning objectives that improve the targeted properties without sacrificing answer accuracy.

Critical strengths

1. The paper introduces a principled, unified decomposition of predictive information allocation across a CoT into two orthogonal dimensions: faithfulness (beyond vs within semantic support) and efficiency (answer-reaching vs off-path), grounded in a directed hypergraph representation of semantic dependency.

2. Fine-tuning experiments show that joint- and efficiency-aware RFT reduce Ineff and token usage, while maintaining or slightly improving Pass@1 relative to standard RFT.

Critical weaknesses

1. The improvements from faithfulness-aware RFT are modest (see Table 7 and 8). More extensive tuning/ablations would clarify headroom.

2. The authors should include additional training experiments across a wider range of models and tasks from different domains to demonstrate the general applicability of their conclusions and methods.

3. The directed-acyclic assumption under generation order may be strained by real-world CoTs that include backtracking or revisions, the impact on the graph construction and metrics is unclear.

Questions for authors

See Weakness

2 comments
AuthorsOpenReview ↗

Reply to Reviewer 8e5S

Dear Reviewer 8e5S

Thank you very much for your time and feedback. We appreciate your comments on the principled and unified nature of our framework and its grounding in semantic dependencies, as well as the effectiveness of our fine-tuning methods. Below, we address your concerns and provide additional clarifications.

[Weakness#1. More extensive tuning of faithfulness-aware RFT.] We would first like to clarify that our existing RFT experiments also serve as a component ablation study. Specifically, standard RFT, faithfulness-aware RFT, efficiency-aware RFT, and joint-aware RFT correspond to four combinations of two components: whether to include the regularization in Equation (2) and whether to remove off-path reasoning steps.

We have additionally conducted a more extensive investigation of the learning rate and regularization weight α\alpha for faithfulness-aware and joint-aware RFT on Qwen3-4B-Instruct-2507 with MATH. Table 52 in Appendix D.6 of the revised manuscript reports all 18 configurations across the two methods, covering learning rates in 10−5,5×10−5,10−4\\{10^{-5},5\times10^{-5},10^{-4}\\} and α∈5×10−4,10−3,5×10−3\alpha\in\\{5\times10^{-4},10^{-3},5\times10^{-3}\\}. For each method and learning rate, we summarize below the lowest observed unfaithfulness and the highest observed Pass@1 across α\alpha.

MethodLearning rateLowest Unfaithfulness (×100)Highest Pass@1 (%)
Standard RFT10−510^{-5}18.8478.30
Faithfulness-aware RFT10−510^{-5}18.75 (α=10−3\alpha=10^{-3})78.50 (α=10−3\alpha=10^{-3})
Joint-aware RFT10−510^{-5}16.92 (α=10−3\alpha=10^{-3})78.30 (α=5×10−4\alpha=5\times10^{-4})
Standard RFT5×10−55\times10^{-5}19.7778.20
Faithfulness-aware RFT5×10−55\times10^{-5}19.54 (α=5×10−4\alpha=5\times10^{-4})78.50 (α=10−3\alpha=10^{-3})
Joint-aware RFT5×10−55\times10^{-5}18.25 (α=5×10−4\alpha=5\times10^{-4})78.80 (α=5×10−4\alpha=5\times10^{-4})
Standard RFT10−410^{-4}19.6976.70
Faithfulness-aware RFT10−410^{-4}19.41 (α=5×10−4\alpha=5\times10^{-4})78.10 (α=5×10−4\alpha=5\times10^{-4})
Joint-aware RFT10−410^{-4}19.67 (α=10−3\alpha=10^{-3})77.70 (α=5×10−3\alpha=5\times10^{-3})

Overall, Table 7 and Tables 48, 50, 51, and 52 in Appendix D.6 together show that faithfulness-aware RFT generally improves faithfulness and often also improves Pass@1 over Standard RFT, while joint-aware RFT can achieve larger reductions in unfaithfulness while maintaining or improving Pass@1. We added this conclusion in the revised manuscript.

[Weakness#2. Experiments covering a wider range of models and tasks.] We thank the reviewer for suggesting experiments across more model–task combinations. In the revised manuscript, we added experiments spanning MATH, Omni-MATH-2, and 2WikiMultiHopQA, and include Qwen3-30B-A3B-Instruct-2507, Qwen3-4B-Instruct-2507, Ministral-3-8B-Instruct-2512, Ministral-3-3B-Instruct-2512, and Llama-3.1-8B-Instruct. Given the space constraints, we refer the reviewer to the updated Tables 3–4 and Figure 2, and in particular, the expanded experimental results in Appendix D.6, Tables 30–52 of the revised manuscript. Overall, these additional results across diverse settings further support our main findings and conclusions from the original submission:

  • Section 4 (Tables 30–33): Controlled interventions produce the expected changes: A-YES increases answer unfaithfulness, S increases global unfaithfulness, and A-NO increases both among cases where the answer changes. On the efficiency side, E reduces global inefficiency and token usage, whereas R increases both.
  • Section 5 (Tables 34-47): Answer unfaithfulness and D1,YτD_{1,Y}^{\tau} consistently show clear positive associations with answer uncertainty. Associations with correctness and those involving efficiency-related measures are generally weaker or less consistent across settings.
  • Section 6 (Tables 48-52): Faithfulness-aware RFT improves its targeted metric relative to Standard RFT without sacrificing Pass@1 accuracy. Note that additional efficiency-aware RFT experiments are still ongoing due to the limited time available. We will add their results to Tables 50 and 51 in the revised manuscript once these experiments are complete.

Please refer to the above tables for detailed numerical results.

AuthorsOpenReview ↗

Reply to Reviewer 8e5S – Continued

[Weakness#3. Our directed-acyclic assumption is compatible with backtracking and revisions.] First, CoT generation is autoregressive with each step using only preceding steps. If the semantic reasoning graph contained a cycle, at least one step would necessarily depend on a future step. Therefore, the acyclicity assumption holds. Second, each vertex represents an individual generated step. A later step that backtracks to or revises an earlier conclusion will be represented as a new vertex, rather than overwriting the earlier vertex. We illustrate this with two simple examples.

Backtracking. Question: xx is either 1 or 2, and x2=4x^2=4. Find xx.

  • A1A_1: “Try x=1x=1: this gives x2=1x^2=1.”
  • A2A_2: “This violates x2=4x^2=4, so exclude x=1x=1.” (backtracking)
  • A3A_3: “The remaining candidate is x=2x=2.”

The dependency structure is: A1→A2→A3.A_1 \rightarrow A_2 \rightarrow A_3.

Revision. Question: Find all real solutions to x2=4x^2=4.

  • A1A_1: “The only real solution is x=2x=2.”
  • A2A_2: “x=−2x=-2 is also a solution.”
  • A3A_3: “Revise the answer to include both solutions: x=2x=2 and x=−2x=-2.” (revision)

The dependency structure is: A1→A3←A2.A_1 \rightarrow A_3 \leftarrow A_2. Neither process introduces a directed cycle or conflicts with our acyclicity assumption.

[Question to the reviewer.] We noticed that the “Flag for Ethics Review” field is marked “Yes,” but we could not find an accompanying explanation. Could you please clarify the concern that prompted this flag? We would appreciate the opportunity to understand and address it.

Finally, we also provide further discussion of the novelty and key contributions of our framework in the General Response. We are happy to answer any further questions you may have. If our responses above have addressed your concerns, we would truly appreciate a re-evaluation accordingly. Thank you very much.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We really appreciate the reviewers’ time and evaluation. We noticed that some comments may suggest a misunderstanding of the novelty of our framework. The conceptual and theoretical contributions of our framework are explicitly presented in Sections 1 and 3 and Appendix E, with detailed comparisons to prior work in Section 2 and Appendix C. We therefore take this opportunity to further clarify the novelty and key contributions of our work. For more details, please refer to our updated pdf file that has highlighted the changes.

Key novelty and contributions of our framework. As shown in Section 3.1 and Figure 1, introducing a semantic reasoning graph allows us to insert three intermediate graph views, FYF_Y, HH, and HYH_Y, between the full autoregressive graph FF and the no-history graph BB (intuitive explanations of these views are provided in lines 186–195). These intermediate views allow us to decompose the total predictive information $D := D_{\mathrm{KL}}\left(p_{\theta}^{F}(T\mid q)\,\Vert\,p_{\theta}^{B}(T\mid q)\right) = \sum_{i=1}^{n} I_{\theta}(A_i;A_{\le i-1}\mid q) + I_{\theta}(Y;A_{\le n}\mid q)$ (see line 200) into four mutualinformation components D=D1,↛Y+D1,→Y+D2,↛Y+D2,→Y,D = D_{1,\not\to Y} + D_{1,\to Y} + D_{2,\not\to Y} + D_{2,\to Y}, as defined in Equation (1). These four components form a two-by-two characterization of reasoning faithfulness and efficiency (The correspondence between these information quantities and faithfulness/efficiency is explicitly explained in lines 227–236 and Table 2). Moreover, because the decomposition is additive across reasoning steps and the final answer, it supports both global scores and local step-level scores, which we formalize for each response in Definition 2. To our knowledge, this information decomposition of CoT faithfulness and efficiency grounded in a semantic reasoning graph is novel in both its formulation and the measurement of reasoning properties it enables.

To demonstrate the reliability and usefulness of this framework, we provide three complementary investigations in Sections 4–6: Section 4 examines whether our metrics capture the expected changes under interventions inducing unfaithful or inefficient reasoning; Section 5 investigates their relationships with predictive uncertainty and answer correctness; and Section 6 evaluates the faithfulness-aware and efficiency-aware fine-tuning methods derived from our framework. These methods are supported by the theoretical guarantees in Theorems 1 and 2.

Distinctions from prior work on CoT faithfulness and efficiency. Our framework simultaneously characterizes CoT faithfulness and efficiency, and moreover, each dimension also offers a distinct contribution relative to its respective literature. These differences are discussed in detail in Section 2 (Table 1) and Appendix C. Here, we briefly elaborate on these distinctions.

CoT faithfulness. We want to clarify that CoT faithfulness remains an open research question without a universally accepted operationalization. Prior studies have approached faithfulness from different perspectives, including semantic sufficiency and logical support (Lyu et al., 2023; Yu et al., 2025), behavioral or causal dependence (Lanham et al., 2023; Paul et al., 2024), disclosure of influential information (Turpin et al., 2023; Chen et al., 2025), and information flow (Jia et al., 2026). Please refer to the literature section for more details. Our framework addresses a distinct question: Does the LLM's predictive dependence throughout a reasoning trajectory agree with the semantic dependencies expressed among its steps? Rather than assessing logical validity or answer-level dependence alone, we characterize behaviorally revealed predictive dependence and semantically inferred support. This enables a unified decomposition of predictive information into faithful versus unfaithful and answer-reaching versus off-path components, yielding compatible global and step-level measures of both faithfulness and efficiency.

CoT efficiency. Unlike most prior work that assesses CoT efficiency primarily through token consumption, our framework measures efficiency through the allocation of predictive dependence. Efficient reasoning should concentrate its predictive dependence on answer-reaching steps and minimize its allocation to off-path reasoning. While most prior approaches to identifying unnecessary reasoning emphasize either structural redundancy (which steps appear semantically unnecessary) or behavioral importance (how strongly the model relies on particular reasoning components), our formulation, to our knowledge, is the first to connect these two perspectives by quantifying the allocation of predictive dependence according to the semantic answer-reaching structure. Moreover, our inefficiency metric shows changes in the same direction as token consumption (Tables 4 and 8), providing complementary empirical support for our interpretation.

AuthorsOpenReview ↗

General Response – Continued

Finally, most of the remaining concerns relate to additional validation and empirical generalization. We appreciate the reviewers’ constructive suggestions and are making every effort to provide further evidence during the discussion period. Meanwhile, we believe that a new framework naturally opens avenues for extensions across models, tasks, and domains, not all of which can reasonably be addressed within a single study. We hope the reviewers can understand and respectfully ask that requests for broader empirical coverage do not overshadow the scope of the paper, the novelty of the theoretical framework, or the experimental evidence provided to support our central claims.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.