acceptodds
Under review as a conference paper at ICLR 2027

Interpreting Reasoning of Large Language Models via Partial Information Decomposition

Abstract

Large reasoning models (LRMs) have achieved substantial improvements in solving complex mathematical problems, but often produce lengthy, repetitive, or erroneous reasoning trajectories. In this work, we introduce a new interpretability framework, SLIDER, to evaluate the quality of the reasoning process. SLIDER leverages an emerging body of work from information theory called Partial Information Decomposition to disentangle the information about the final answer between two consecutive reasoning steps into non-negative components: unique information (in preceding steps or current step), redundant information, and synergistic information. Building on this decomposition, we propose the *Step-wise Repetitive Reasoning Index (Step-RRI)*, a theoretically grounded measure that assesses whether the answer-relevant information in the current step is predominantly redundant with the past steps , relative to its unique and synergistic contributions. To evaluate the effectiveness of Step-RRI in detecting repetitiveness, we apply SLIDER to the redundancy class of the PRMBench dataset where Step-RRI improves step-level redundancy identification accuracy by over points compared to embedding-similarity and information-gain baselines. Next, we define *Trajectory-RRI*, an aggregate measure of repetitiveness for an individual reasoning trajectory. To demonstrate its practical relevance, we show that average Trajectory-RRI strongly correlates with actual reasoning length across QwQ-32B, DeepSeek-R1-Distill-Qwen-32B, and GPT-4.1, motivating its use as a signal for improving reasoning efficiency. Finally, we introduce *Trajectory-RRI-guided data selection for fine-tuning*, demonstrating that selecting training data based on Trajectory-RRI can improve a fine-tuned model's reasoning efficiency while largely preserving its task performance. Our results establish RRI measures as both a step-level diagnostic of repetitive reasoning and a practical signal for improving reasoning efficiency.

Primary area:
interpretability and explainable AI
Keywords:
Large Reasoning Models, Partial Information Decomposition, Interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SCTUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SLIDER, an interpretability framework based on Partial Information Decomposition (PID), to analyze the contributions of individual reasoning steps to the final answer in large language models. Specifically, the proposed method decomposes the answer-relevant information shared between the current reasoning step and preceding steps into redundant, unique, and synergistic information. Furthermore, it introduces step-level and trajectory-level metrics to quantify repetitive reasoning. Experimental results demonstrate the effectiveness of the proposed method in identifying repetitive reasoning steps and improving reasoning efficiency.

Critical strengths
  1. Compared with traditional mutual information or conditional mutual information, PID distinguishes redundant, unique, and synergistic information, providing a different perspective for understanding the contributions of individual reasoning steps.
  2. The paper evaluates the proposed method from multiple perspectives through theoretical analysis, synthetic data experiments, and evaluations of reasoning trajectories generated by real-world language models.
Critical weaknesses
  1. Although the authors demonstrate a strong correlation between Trajectory-RRI and reasoning length, they have not sufficiently established whether this metric can accurately measure reasoning redundancy independently of reasoning length.

  2. The reliability of PID estimation requires further validation. SLIDER heavily relies on accurate estimation of the PID components. However, the proposed method involves a relatively complex estimation pipeline, including generating up to 50 numerical variants using an LLM, encoding them with all-MiniLM-L6-v2, discretizing the continuous representations into 10 clusters using K-means, and estimating the joint probability distribution. The rationale behind each component of this pipeline requires further justification.

  3. The effectiveness of Step-RRI in detecting repetitive reasoning requires more comprehensive evaluation. Additional evaluation metrics and stronger baselines should be included to better demonstrate the effectiveness of the proposed method.

  4. The validity of the structure-preserving sample generation strategy requires further investigation. Since a single reasoning trajectory is insufficient to estimate the joint probability distribution, the authors use GPT-4o-mini or GPT-5-mini to generate multiple numerical variants while preserving the same reasoning structure. However, a concern is whether these generated trajectories can accurately reflect the actual reasoning behavior of language models across different problems. In practice, even when two problems share the same structure and differ only in numerical values, language models may follow different reasoning paths, perform different intermediate computations, and exhibit different repetitive reasoning behaviors.

Questions for authors

See Weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer SCTU: Part 1/2

We thank the reviewer for their thoughtful comments.

We have made several edits to the manuscript (marked in blue), including an analysis of SLIDER on PERL traces with natural errors (Section 5 and Appendix A.1), experiments on removing repetitive reasoning steps based on Step-RRI (Appendix A.2), a comparison of redundant information alone with Step-RRI (Appendix A.3), and experiments on how Trajectory-RRI contributes beyond reasoning length (Appendix A.4). We also report additional evaluation metrics (Appendix A.5), present additional sensitivity analysis (Appendix A.6), and expand the discussion of SLIDER’s limitations (Appendix B).

On weakness 1: Trajectory-RRI and reasoning length

We agree that the correlation between Trajectory-RRI and token count alone leaves open how much information Trajectory-RRI provides about redundancy beyond reasoning length. Though average Trajectory-RRI is positively correlated with average reasoning length, we argue that Trajectory-RRI can also capture differences in repetitiveness that length alone misses. A trajectory could be long due to the genuine demands of the problem, e.g., complexity (desirable) or due to unnecessary steps (undesirable), and these two aspects cannot be distinguished by looking at the total length alone.
On the other hand, Trajectory-RRI systematically captures how high the redundancy Red(A:S<i,Si)Red(A:S_{<i},S_i) of the current step SiS_i is in comparison to its informativeness (Uni(A:Si∣S<i)Uni(A:S_{i}|S_{<i}) or Syn(A:S<i,Si)Syn(A:S_{<i},S_i)) about the final answer AA.

We illustrate this with two examples in Appendix A.4.1 from the redundancy subset of PRMBench in which Step-RRI identifies the annotated redundant steps. Problem 1 requires more intermediate calculations and has a longer reasoning trajectory with only one annotated redundant step. Problem 2 has a shorter trajectory with more annotated redundant steps. In this pair, ranking the trajectories by Trajectory-RRI agrees with the annotated redundancy counts. Thus, a longer reasoning trace can contain fewer redundant steps than a shorter one.

On weakness 2: On PID estimation

We acknowledge that estimating PID involves a complex pipeline. Estimating PID terms is computationally challenging and requires nontrivial optimization. This task becomes even more difficult in the continuous high-dimensional domain. However, PID’s ability to quantify the unique, redundant, and synergistic contributions about a target has motivated its use in several machine learning applications involving high-dimensional real-valued data, including analyzing multimodal interactions [1], measuring the information spectrum of large vision-language models [2], and interpreting diffusion models [3]. In this work, we leverage PID to interpret reasoning traces and to quantify repetitiveness using a new measure, Step-RRI.

Each component in SLIDER makes the PID estimation tractable at the individual problem level. The sample generator produces variants that share the same reasoning structure. Next, we use an encoder to map text to embeddings so that semantically similar passages can be grouped even when their wording differs. Finally, we cluster the embeddings and estimate a joint probability distribution for each problem, which serves as the input to the PID optimization. We conduct two sensitivity analyses to assess how the number of generated variants and the number of clusters affect PID estimates.

  • Number of variants (Table 4, Appendix A.6): In the revised manuscript, we now include a sensitivity analysis of PID values for different numbers of variants N∈{50,100,500}N\in\{50,100,500\}. We observe that the dominant PID component remains the same at all sample sizes.

  • [From Previous Draft] Number of clusters (Table 5, Appendix D.2): We vary the number of clusters to assess the stability of the results. We observe that the patterns of redundancy, uniqueness, and synergy are largely preserved across different cluster choices.

In conclusion, SLIDER provides a practical framework for estimating PID on textual data at an individual problem level and can be extended to incorporate other PID estimators.

References:

[1]Liang, Paul, et al. "Multimodal learning without labeled multimodal data: Guarantees and applications." International Conference on Learning Representations. Vol. 2024. 2024.

[2]Xiu, Lixin, Xufang Luo, and Hideki Nakayama. "A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models." arXiv preprint arXiv:2603.29676 (2026).

[3]Dewan, Shaurya, et al. "Diffusion pid: Interpreting diffusion via partial information decomposition." Advances in Neural Information Processing Systems 37 (2024): 2045-2079.

AuthorsOpenReview ↗

Response to Reviewer SCTU: Part 2/2

On weakness 3: Additional evaluation metrics

To further validate the effectiveness of Step-RRI in detecting repetitive reasoning, we incorporated two additional evaluation metrics: (1) an LLM-as-a-judge repetitiveness score and (2) a human-annotated repetitiveness score. We evaluated example reasoning traces from both QwQ-32B and DeepSeek-R1-Distill-Qwen-32B.

We observed statistically significant positive Spearman correlations (ρ>0.5\rho>0.5) between Step-RRI and both the LLM and human repetitiveness scores, outperforming the baseline InfoGain. These positive correlations further support the effectiveness of Step-RRI in identifying repetitiveness in reasoning trajectories. Appendix A.5 of the revised manuscript describes the shared scoring criteria used by the LLM judge and human annotators to rate step-level repetitiveness.

Table 3. Correlation between Step-RRI and step-level repetitiveness ratings.

MethodHuman Spearman’s ρHuman p-valueLLM Spearman’s ρLLM p-value
−InfoGain0.4240.0550.3880.082
Step-RRI (ours)0.6520.0010.5620.008

On weakness 4: Validity of structure-preserving sample generation

We agree that a model can follow different reasoning paths when the numbers in a problem change. However, our sampling strategy is designed to estimate PID for a fixed reference trajectory. The resulting PID estimates characterize how answer-relevant information is distributed across steps within the reference reasoning structure. If a model takes a different path when solving a variant independently, that path falls outside the scope of this estimate.

We use GPT-4o-mini or GPT-5-mini to generate numerical variants of each reasoning trajectory. The prompt templates in Appendix D.1 STRUCTURE-PRESERVING SAMPLE GENERATION instruct the models to change only numerical values, preserve the structure and order of the reasoning steps, and recompute intermediate results for each variant. We manually inspected a subset of the generated samples and refined the prompts to ensure that the resulting variants satisfied these constraints. We then discard variants whose step count differs from that of the original trajectory. Thus, we restrict each variant to numerical changes while preserving the reference trajectory’s reasoning structure.

Dear reviewer,

As the review period is coming to an end, we would really appreciate it if you could let us know whether our rebuttal has addressed your concerns. We are happy to respond to any additional questions. Thank you so much for your time and consideration.

Thank you for your response. The authors have largely addressed my concerns, so I have decided to raise my score.

Reviewer_q9bNOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SLIDER, a framework for analyzing repetitive reasoning through Partial Information Decomposition (PID). It separates answer-relevant information into redundant, unique, and synergistic contributions from the current reasoning step and its preceding context. Based on this decomposition, the authors propose Step-RRI for identifying repetitive steps and Trajectory-RRI for quantifying repetition across a reasoning trace. These quantities are estimated using aligned numerical variants, text embeddings, and clustering. Experiments on synthetic arithmetic problems, PRMBench, GSM8K, and AIME2025 demonstrate improved redundancy detection and associations with reasoning length. The paper also explores RRI-guided data selection for supervised fine-tuning, obtaining shorter outputs with broadly comparable task accuracy.

Critical strengths
  1. A Principled Application of PID to Reasoning Analysis. I appreciate the authors' use of Partial Information Decomposition (PID) to characterize answer-relevant information in reasoning steps through redundancy, unique information, and synergy, rather than relying on heuristic measures such as embedding similarity. I find the consideration of synergy particularly valuable. In multi-step mathematical reasoning, a step may not directly reveal information about the final answer but can contribute meaningfully when combined with previous results. This distinction is, in my view, one of the most compelling conceptual contributions of the paper. Although PID itself is well established, adapting it to Chain-of-Thought (CoT) reasoning is nontrivial. Estimating mutual information and PID requires probability distributions that cannot be reliably inferred from a single reasoning trajectory. The authors address this challenge through structure-preserving numerical variations, which generate reasoning-aligned samples by modifying numerical values in mathematical problems. These samples enable the estimation of joint distributions between reasoning steps and final answers. I find this a thoughtful and interesting approach that provides a practical connection between PID and reasoning efficiency analysis.

  2. A Promising Connection to Downstream Training. Another strength is that the paper goes beyond post hoc interpretability analysis and explores practical applications in model training. The authors establish a coherent framework that progresses from PID-based analysis of reasoning information to Step-RRI for identifying repetitive steps, Trajectory-RRI for measuring overall redundancy, and RRI-guided data selection for supervised fine-tuning (SFT). Importantly, the paper also provides downstream experimental evidence. The results suggest a strong correlation between the RRI of selected training trajectories and the reasoning length and efficiency scores of fine-tuned models, while largely preserving problem-solving accuracy. I appreciate this effort to connect information-theoretic analysis with practical improvements in reasoning efficiency, which makes the proposed framework more convincing and highlights its potential utility for training data selection.

Critical weaknesses
  1. Redundancy Does Not Necessarily Imply Inefficiency. The paper primarily identifies repetitive reasoning based on the redundancy of answer-relevant information between the current step and previous steps. However, information redundancy does not necessarily imply reasoning inefficiency. For example, recomputing previous results, cross-checking solutions through alternative approaches, or revisiting important constraints may introduce little new information while still helping detect errors and improve the reliability of the final answer. This distinction is particularly important for complex mathematical reasoning and long-horizon tasks that require verification and self-correction. A discussion of existing related work [1] could also help contextualize the trade-off between reducing redundant reasoning and preserving necessary deliberation. While Step-RRI captures statistical relationships between reasoning steps, the current experiments do not sufficiently establish whether steps identified as repetitive are indeed unnecessary for successful reasoning. It therefore remains unclear whether RRI can reliably identify genuinely unnecessary reasoning rather than merely measuring information redundancy.

  2. Confounding Between Trajectory-RRI and Reasoning Length. The paper defines Trajectory-RRI as the absolute number of repetitive reasoning steps, making it inherently dependent on the total number of reasoning steps. Even when two trajectories have the same proportion of repetitive steps, the longer trajectory may receive a higher Trajectory-RRI. Consequently, the reported strong correlation between Trajectory-RRI and reasoning length does not necessarily demonstrate that the metric captures redundancy independently of trajectory length. Furthermore, the correlation analysis is conducted using group-level averages after grouping trajectories by reasoning length, which further limits the interpretation of the results. It remains unclear whether Trajectory-RRI provides meaningful information beyond simple length-based metrics, such as token count and step count, and whether this additional information translates into downstream training benefits.

[1] Efficient Reasoning with Balanced Thinking. ICLR 2026.

Questions for authors

I am currently leaning toward Weak Acceptance. I appreciate the well-motivated problem formulation, the interesting application of PID to reasoning analysis, and the promising experimental results. However, I still have some concerns regarding the practical validity of RRI and its contribution beyond reasoning length. Clarifying the following questions, ideally with additional experimental evidence, would help strengthen my confidence in the work.

  1. Practical Utility of Redundancy Detection. Could the authors consider conducting step-level intervention experiments in which reasoning steps with high and low RRI scores are selectively removed or skipped, allowing the model to continue reasoning? Comparing the resulting accuracy and token consumption would help assess whether RRI identifies steps that can be safely omitted. In particular, can RRI distinguish unnecessary repetition from verification or self-correction steps that contribute to answer reliability?

  2. Independent Contribution of Trajectory-RRI. Could the authors further investigate the contribution of Trajectory-RRI while controlling for reasoning length and step count? For example, in the SFT data selection experiments, it would be informative to compare selection strategies based on token count, step count, normalized Trajectory-RRI, and the original Trajectory-RRI, while controlling for potential differences in teacher models. Such comparisons would help establish whether RRI offers practical benefits beyond simpler length-based metrics.

4 comments
AuthorsOpenReview ↗

Response to Reviewer q9bN: Part 1/2

We thank the reviewer for their thoughtful feedback and for highlighting the strengths of our work. We appreciate the comment, ``I appreciate the well-motivated problem formulation, the interesting application of PID to reasoning analysis, and the promising experimental results.’’

We have made several edits to the manuscript (marked in blue), including an analysis of SLIDER on PERL traces with natural errors (Section 5 and Appendix A.1), experiments on removing repetitive reasoning steps based on Step-RRI (Appendix A.2), a comparison of redundant information alone with Step-RRI (Appendix A.3), and experiments on how Trajectory-RRI contributes beyond reasoning length (Appendix A.4). We also report additional evaluation metrics (Appendix A.5), present additional sensitivity analysis (Appendix A.6), and expand the discussion of SLIDER’s limitations (Appendix B). We have also included [1] in the Related Work section of the revised manuscript.

On W1: Redundant information alone vs Step-RRI

Step-RRI does not contain redundancy alone, but essentially redundancy relative to informativeness. Redundant information (Red(A:Si,S<i)Red(A:S_i, S_{<i})) measures whether the current reasoning step SiS_i repeats answer-relevant information already present in the preceding steps S<iS_{<i}. Step-RRI quantifies redundant information Red(A:Si,S<i)Red(A:S_i, S_{<i}) relative to its informativeness (unique information $Uni(A:S_i|S_{<i}) \ \text{or}\ $ synergy Syn(A:S<i,Si)Syn(A:S_{<i},S_{i})). It subtracts η\eta times the larger of two PID terms Uni(A:Si∣S<i) or Syn(A:S<i,Si)Uni(A:S_i|S_{<i}) \ \text{or}\ Syn(A:S_{<i},S_{i}) from the redundancy term. Therefore, when two steps have the same redundancy, the more informative step gets a lower Step-RRI score. This distinction helps Step-RRI flag potentially unnecessary reasoning. In Fig. 12 in Appendix A.3, we observe that steps S3S_3 and S4S_4 have similar PID redundancy, but S3S_3 carries substantially more unique information and receives a negative Step-RRI score, while S4S_4 receives a positive score. Thus, a step can repeat information from earlier reasoning and still receive a low repetitiveness score as measured by Step-RRI because it adds more answer-relevant information.

On W1 & Q1: Step-level Intervention of repetitive steps

To test whether high-Step-RRI steps can be removed with minimal impact on final-answer accuracy, we use 1,000 reasoning traces generated by QwQ-32B for problems in the GSM8K training set. For each trace, we remove the k steps with the highest estimated Step-RRI scores, retain the remaining steps in their original order, and resume generation from the resulting prefix. Table 2 in Appendix A.2 shows that, relative to the k=0 baseline, mean token counts decrease by approximately 14%, 27%, and 32% for k=1,2,3, respectively. Accuracy remains comparable to the baseline across all three interventions.

top-kAccuracy (%)Token Counts
0100.00990.08
199.86848.67
2100.00726.49
399.73669.81

[1] Efficient Reasoning with Balanced Thinking. ICLR 2026.

AuthorsOpenReview ↗

Response to Reviewer q9bN: Part 2/2

On W2 & Q2: Trajectory-RRI vs reasoning length

Reasoning length alone is not a good replacement for Trajectory-RRI. This is because Trajectory-RRI systematically captures how high the redundancy (Red(A:Si,S<i)Red(A:S_i, S_{<i})) of the current step SiS_i is in comparison to its informativeness (Uni(A:Si∣S<i) or Syn(A:S<i,Si)Uni(A:S_i|S_{<i}) \ \text{or}\ Syn(A:S_{<i},S_{i})) about the final answer AA, while length is more heuristic. Length captures both the genuine demands of the problem, e.g., complexity and how the model approaches it, e.g., repetitiveness and unnecessary steps, and these two aspects cannot be distinguished by looking at the total length alone.

For example, a problem that requires several intermediate calculations may warrant a longer solution even without redundant steps. We illustrate this with two examples in Appendix A.4.1 (WHY LENGTH IS NOT ENOUGH TO CAPTURE REPETITIVENESS?) from the redundancy subset of PRMBench in which Step-RRI identifies the annotated redundant steps. Problem 1 requires more intermediate calculations and has a longer reasoning trajectory with only one annotated redundant step. Problem 2 has a shorter trajectory with more annotated redundant steps. In this pair, ranking the trajectories by Trajectory-RRI agrees with the annotated redundancy counts. Thus, a longer reasoning trace can contain fewer redundant steps than a shorter one.

Comparison of SFT Data Selection Methods. To investigate the practical benefits of Trajectory-RRI beyond simpler length-based metrics for SFT data selection, we compare SFT data selection by Trajectory-RRI with selection based on length (thinking-token count). Selecting the trace with the lowest Trajectory-RRI per problem improves test accuracy by 0.68 percentage points over selecting the shortest trace, at similar average test token counts (Fig.13, Appendix A.4.2). Trajectory-RRI can retain useful steps that the shortest traces may skip. Selecting the longest traces gives the highest accuracy, but uses 2.3 as many test tokens as selecting by lowest Trajectory-RRI (Fig.13, Appendix A.4.2). Since length counts both useful reasoning and undesirable repetition, it offers limited guidance for balancing accuracy and efficiency.

To examine whether Trajectory-RRI helps select SFT data when average training reasoning length is held fixed, we construct two subsets with matched mean trace lengths and different mean Trajectory-RRI scores. We use traces generated by QwQ-32B for both subsets to control for differences between teacher models. We fine-tune Qwen2.5-7B-instruct separately on each subset using LoRA and evaluate both models on the GSM8K test set. As shown in Fig.14, Appendix A.4.2 the model trained on the subset with lower mean Trajectory-RRI achieves about one percentage point higher test accuracy. In essence, Step-RRI does not capture redundancy alone, but redundancy relative to informativeness. Thus, two steps can have the same redundancy, but the more informative step gets a lower Step-RRI score. Therefore, training on more repetitive traces may encourage the model to revisit steps it has already completed, creating more opportunities for errors.

Dear reviewer,

As the review period is coming to an end, we would really appreciate it if you could let us know whether our rebuttal has addressed your concerns. We are happy to respond to any additional questions. Thank you so much for your time and consideration.

Thank the authors for their additional experiments and analysis. Most of my concerns are addressed. I maintain my positive score.

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

This paper proposes SLIDER, a framework that uses Partial Information Decomposition (PID) to analyze repetitive reasoning in large language models. The authors introduce Step-RRI to identify repetitive reasoning steps and Trajectory-RRI to quantify repetitiveness across an entire reasoning trajectory. Experiments show that Step-RRI outperforms cosine similarity and InfoGain baselines on PRMBench, while Trajectory-RRI strongly correlates with reasoning length and can be used for SFT data selection.

Critical strengths

(1) A potentially interesting methodological direction. Using PID to distinguish redundant, unique, and synergistic information in reasoning steps provides a perspective on reasoning analysis that differs from conventional similarity-based metrics.

(2) Quantitative validation. Step-RRI achieves 61.91% accuracy on the redundancy class of PRMBench, outperforming the evaluated cosine similarity and InfoGain baselines.

(3) Exploration of practical applications. The authors apply Trajectory-RRI to SFT data selection. The results suggest that this approach may improve reasoning efficiency while largely preserving task accuracy.

Critical weaknesses

(1) The paper primarily studies reasoning errors through synthetic numerical perturbations. Although Appendix D.1 analyzes a model-generated GSM8K trajectory, the errors are artificially introduced rather than naturally produced by the model. Furthermore, the trajectory-length experiments explicitly select trajectories with correct final answers. Consequently, the paper does not establish whether SLIDER remains effective on naturally generated trajectories with incorrect final answers, leaving an important gap in its empirical validation.

(2) SLIDER uses the generated final answer as the target variable for PID. When the answer is incorrect, the decomposition still characterizes the information relationships between reasoning steps and the generated answer, but does not necessarily reflect their contribution to the correct solution. The paper does not clarify the interpretation and limitations of its metrics in this setting or empirically establish whether they remain reliable for detecting repetitive reasoning in erroneous trajectories.

(3) The theoretical and empirical analyses of errors primarily focus on controlled numerical perturbations, while the quantitative evaluation on PRMBench focuses on redundancy detection. These results do not establish that the proposed metrics can assess other aspects of reasoning quality, such as incorrect assumptions or invalid logical deductions. The paper should more clearly distinguish its demonstrated capability for redundancy detection from its broader claims about reasoning evaluation.

Questions for authors

(1) Can the authors evaluate SLIDER on naturally generated reasoning trajectories with incorrect final answers, without artificially injecting numerical noise? Does Step-RRI reliably identify repetitive steps in these trajectories?

(2) When the generated final answer is incorrect, how should the resulting PID components be interpreted? Can the authors clarify whether the metrics characterize information relevant to the generated answer, the correct answer, or both, and what limitations follow from this distinction?

(3) To what extent do the theoretical and empirical findings obtained under synthetic numerical perturbations generalize to naturally occurring reasoning errors, such as incorrect intermediate conclusions or invalid logical deductions? What evidence supports the applicability of SLIDER beyond the controlled error settings considered in the paper?

(4) Since the PRMBench evaluation focuses on redundancy detection, what evidence supports extending the proposed framework to broader reasoning-quality assessment? Can the authors clarify which types of reasoning errors the current method is designed to detect and which remain outside its scope?

3 comments
AuthorsOpenReview ↗

Response to Reviewer CbSS: Part 1/2

We thank the reviewer for their review. We appreciate the comment: ``A potentially interesting methodological direction. Using PID to distinguish redundant, unique, and synergistic information in reasoning steps provides a perspective on reasoning analysis that differs from conventional similarity-based metrics.’’

We have made several edits to the manuscript (marked in blue), including an analysis of SLIDER on PERL traces with natural errors (Section 5 and Appendix A.1), experiments on removing repetitive reasoning steps based on Step-RRI (Appendix A.2), a comparison of redundant information alone with Step-RRI (Appendix A.3), and experiments on how Trajectory-RRI contributes beyond reasoning length (Appendix A.4). We also report additional evaluation metrics (Appendix A.5), present additional sensitivity analysis (Appendix A.6), and expand the discussion of SLIDER’s limitations (Appendix B).

On W1,3 and Q1,3: SLIDER under natural errors

To determine naturally occurring reasoning errors, we examine an annotated trace containing logical inconsistencies and accumulation errors from PERL[1], and a DeepSeek-R1-Distill-Qwen-32B reasoning trajectory for a GSM8K word problem with logical errors (Appendix A.1).

To use SLIDER for this error determination, we need the ground truth answer as the target YY. The relevant PID measures for error detection are: (i) the unique information (Uni(A:Si∣S<i)Uni(A:S_i|S_{<i})) which measures what the current step contributes on its own; and (ii) synergy (Syn(A:S<i,Si)Syn(A:S_{<i},S_{i})) measures what it contributes jointly with the preceding steps. Changes in these terms can help flag arithmetic mistakes, invalid deductions, or steps that carry forward an earlier error.

Fig. 10 in Appendix A.1(SLIDER UNDER NATURAL ERRORS) shows reasoning trajectories under logical and accumulation error and their corresponding PID values. In Problem W1, step S3S_3 correctly computes the earnings and has the highest unique information in the current step S3S_3. At S4S_4, the solver assumes without justification that the peach Lilia kept is worth \2.Thislogicalinconsistencycoincideswithadropincurrent−stepuniqueinformationto0.137.Step2. This logical inconsistency coincides with a drop in current-step unique information to 0.137. Step S_5thenaddstheassumedvaluetothesalessubtotalandreports then adds the assumed value to the sales subtotal and reports \\27. Although the addition is arithmetically correct, it propagates the earlier error. The unique information in S5S_5 rises to 0.262 but remains below its value at S3S_3. This example illustrates how tracking changes in the unique information in the current reasoning step can help identify departures from valid reasoning. In Problem W2, the model misinterprets the problem and never reaches the correct answer. This pattern is also identifiable from the trends in the PID values. The preceding steps S<3S_{<3} and the current step S3S_3 jointly provide more information about the target variable, resulting in high synergistic information; however, in subsequent steps, there is no significant increase in the unique information provided by the current step. Thus, by tracking PID values across reasoning steps, we can identify logical errors.

However, a step with a systematic (non-random/fixed) error may remain statistically dependent on the correct final answer, and thus undetectable. Its unique information can remain substantial if the step itself conveys answer-relevant information absent from the preceding steps. In this scenario, our framework SLIDER might not work as expected for error detection. We include this limitation in Appendix A.1 and B in the revised paper.

[1] Mukherjee, Sagnik, et al. "Premise-augmented reasoning chains improve error identification in math reasoning with llms." arXiv preprint arXiv:2502.02362 (2025).

AuthorsOpenReview ↗

Response to Reviewer CbSS: Part 2/2

On W 1,2, & Q1,2,4: Step-RRI under errors

For repetitiveness, SLIDER with Step-RRI still remains useful even under errors. Instead of using the ground-truth answer, we can use the model-generated final answer (could be incorrect) as the target YY. Step-RRI (Red(A:Si,S<i)−η⋅maxUni(A:Si∣S<i),Syn(A:S<i,Si)Red(A:S_i,S_{<i}) - \eta \cdot max\\{Uni(A:S_i|S_{<i}), Syn(A:S_{<i},S_{i})\\}) computed with the ground truth answer might understate its repetitiveness since the mistakes carry little information about the ground truth answer, and hence the redundancy/uniqueness/synergy about the ground truth answer that is shared across steps tends to 0. However, using the model-generated answer as the target Y allows Step-RRI to characterize repetition relative to the conclusion the model actually reached, even when that conclusion is incorrect. For example, a step that restates an earlier mistake may receive a high score since the step can have high redundancy about the incorrect final answer and low informativeness (since the mistake is repeated). This innovative use of Step-RRI with the model-generated answer makes it applicable to situations where the ground truth label may be unavailable.

To investigate how the Step-RRI behaves under model-generated logical inconsistency, we consider the following problem: `` Jen decides to travel to 3 different countries. He has to pay \124forthesuppliesheneeds,intotal.Theticketsfortravelcost,intotal,50%morethanthesupplies.Howmuchdoestravelcost?′′Ideally,wecansummarizethereasoningflowforthiswordproblemasfollows:124 for the supplies he needs, in total. The tickets for travel cost, in total, 50\% more than the supplies. How much does travel cost?'' Ideally, we can summarize the reasoning flow for this word problem as follows: S_1:compute: compute 50%ofthesupplycostof the supply costx;; S_2: compute \0.5x+x $ to obtain the ticket cost; S3S_3: compute 1.5x+x1.5x+x to determine the total travel cost. However, as shown in Fig. 11, Appendix A.1(SLIDER UNDER NATURAL ERRORS), the reasoning demonstrates a misinterpretation of the problem objective, where the model computes only the ticket cost rather than the complete travel cost required by the task. Fig. 11 also shows the corresponding PID values and Step-RRI. The model correctly calculates that the tickets cost \186,butomitsthe186, but omits the \\124 spent on supplies, yielding \186insteadofthecorrecttotalof186 instead of the correct total of \\310. Using the model-generated answer as the proxy PID target, we compute Step-RRI for each step. Step-RRI is negative at S2S_2 and S3S_3, where the model computes the \62markupandusesittoobtaintheticketcost.Itbecomespositiveat62 markup and uses it to obtain the ticket cost. It becomes positive at S_4andandS_5,wherethemodelrecheckstheadditionbyplacevalueandthenverifiesthesameticketcostbymultiplying, where the model rechecks the addition by place value and then verifies the same ticket cost by multiplying \\124 by 1.5. Although the two checks use different calculations, both revisit a result already established at S3S_3. This example illustrates how Step-RRI can identify repetitiveness with respect to the model’s answer even when that answer is incorrect.

Dear reviewer,

As the review period is coming to an end, we would really appreciate it if you could let us know whether our rebuttal has addressed your concerns. We are happy to respond to any additional questions. Thank you so much for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Author Response

We thank the AC and reviewers for reviewing this paper. We appreciate that Reviewer CbSS has found our methodology interesting, and Reviewer q9bN has appreciated our well-motivated problem formulation, the interesting application of PID to reasoning analysis, and the promising experimental results.

In response to their comments, we have made several edits to the manuscript (marked in blue) and made the following changes:

Appendix A.1: SLIDER under natural errors. To determine naturally occurring reasoning errors, we examine an annotated trace containing logical inconsistencies and accumulation errors from PERL, and a DeepSeek-R1-Distill-Qwen-32B reasoning trajectory for a GSM8K word problem with logical errors. We also illustrate how, for repetitiveness, SLIDER with Step-RRI still remains useful even under errors.

Appendix A.2: Step-level intervention of repetitive steps. To test whether high-Step-RRI steps can be removed with minimal impact on final-answer accuracy, we perform additional experiments on GSM8K.

Appendix A.3: Redundant Information Alone vs Step-RRI. We examine how Step-RRI differs from redundant information using a model-generated reasoning trajectory.

Appendix A.4: Trajectory-RRI vs Reasoning Length. We perform additional experiments to compare SFT data selection using Trajectory-RRI and a length-based heuristic to investigate the practical benefits of Trajectory-RRI beyond simpler length-based metrics.

Appendix A.5: Additional Evaluation Metrics. To further validate the effectiveness of Step-RRI to identify repetitive reasoning, we add additional evaluation metrics including LLM-as-a-judge and a human-annotated repetitiveness score.

Appendix A.6: Additional Sensitivity Analysis. We include a sensitivity analysis of how PID values change with the number of generated variants.

We have also included the additional references, expanded the discussion of SLIDER’s limitations in Appendix B, and are committed to making any other edits or answering any other questions. We look forward to the responses to our rebuttal.

Dear Reviewers,

As the discussion period is coming to an end, I kindly encourage you to review and respond to the authors' rebuttal as soon as possible.

Thank you very much for your time and valuable contributions to the review process.

Best regards,
Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.