acceptodds
Withdrawn from ICLR 2027

Distilling Model-Inherent Reasoning Circuit for Large Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Knowledge Distillation (KD) is widely adopted to compress Large Language Models (LLMs) by aligning the output distributions of teacher model and student model. However, output matching alone does not explicitly characterize the internal computations reflecting the teacher's reasoning capability, and directly aligning the representations of paired Transformer components risks disrupting the student's intrinsic reasoning pattern. To address these limitations, we propose model-inherent Reasoning Circuit Distillation (RCD). RCD characterizes model-inherent reasoning patterns as critical circuits formed by Transformer edges, with a separate circuit identified for each model using Edge Attribution Patching with Integrated Gradients. Subsequently, to transfer the teacher's reasoning capability, RCD aligns the predictive effects of both circuits by matching the logit variations induced by circuit attenuation. Theoretically, we establish guarantees for recovering the ground-truth reasoning circuits and show that the empirical RCD loss controls the population discrepancy in predictive effects between the ground-truth circuits of the teacher model and the student model. Experiments on mathematical reasoning and code generation show that RCD consistently outperforms state-of-the-art KD methods.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Knowledge Distillation, Large Language Models, Mechanistic Interpretability, Circuit Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_scoUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The job try to propose an alternative paradigm to conventional knowledge distillation: instead of enforcing identical internal architectures between the teacher and student models, it identifies the critical computational circuits for each model separately, then aligns the impact of these circuits on model predictions. Nevertheless, the paper suffers from a fundamental flaw that persists throughout the manuscript: the authors prove and measure prediction shifts under specific interventions, yet repeatedly interpret such shifts as the models’ inherent reasoning mechanisms. They further equate the alignment of prediction shifts with the transfer of reasoning capabilities.

Critical strengths

S1: Well-organized layout and thoughtfully chosen color schemes for figures.

S2: The proposed idea is moderately novel: task-relevant circuits of teacher and student are identified independently, and distillation is achieved via aligning their predictive effects rather than direct matching of internal components.

S3: The method is compatible: RCD serves as an auxiliary loss and can be integrated with existing distillation methods (Naive KD, ABKD, etc.) with no architecture modification.

S4: Experimental validation is fairly comprehensive: comparisons against baselines over multiple model families on mathematical reasoning and code generation tasks, supplemented by circuit analysis and ablation studies.

Critical weaknesses

W1: The claim that we can perfectly recover the circuit depends on Assumption 2. This assumption directly states that all circuits found by EAP together equal the true faithfulness-optimal circuit. In other words, the assumption already takes for granted the core identifiability relation that the theorem is supposed to prove. The authors need to clarify: the theorem only guarantees consistent estimation if identifiability holds. It does not prove that identifiability is true in the first place.

W2: (Theorem 2,Lines 317–323; C.3,Lines 1208–1218) The generalization analysis assumes fixed model checkpoints and alignment samples that do not depend on training or model selection. In practice, though, the RCD loss is optimized over training data, and student parameters keep changing throughout training. Because of this, the bound they derive cannot directly prove generalization for the training objective or the learned student model. The authors need to clearly explain this difference. They also need to align the per-token selection for each example assumed in the proof with the batch-aware allocation used in Algorithm 1.

W3: A central motivation of RCD is to preserve the student’s intrinsic reasoning pattern, a claim that remains unverified in the paper (Lines 173–178, Lines 245–249, Appendix B, Algorithm 1, Line 937). While the method fixes selected edge indices during training, this does not guarantee that their computational functions stay invariant as the student parameters get updated. Direct circuit comparison before and after distillation, alongside baselines such as standard KD, is required to substantiate this claim.

W4: Interpreting Logit Variation Alignment as Reasoning Capability Transfer constitutes a conceptual leap (Lines 255–259, Page 5, Equation (6), Page 9, Conclusion, Lines 478–483). Though the proposed loss aligns circuit-attenuation-induced logit changes, it is unclear whether matching these predictive effects corresponds to reasoning capability transfer instead of general intervention sensitivity. The paper does not build a direct connection between effect alignment and the preservation or acquisition of specific reasoning operations. The authors are advised to provide targeted mechanistic evidence or limit their claim of RCD to intervention-based functional distillation.

Questions for authors

I suggest the authors conduct a unified ablation study on the underlying mechanism: compare the distillation performance of EAP Circuit, Random Circuit, and Bottom-K Circuit, and recover the student circuits after distillation. This experiment can simultaneously verify the two core claims: whether the identified circuits are genuinely important, and whether RCD indeed preserves the student’s internal computational patterns.

Reviewer_w2ftOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method for extracting and distilling model-inherent reasoning circuits from large language models. The authors identify internal attention patterns and MLP activations that correspond to specific reasoning steps, then distill these circuits into a compact representation that can be transferred to smaller models. The method involves analyzing model internals during correct reasoning, identifying circuit components through causal intervention, and distilling the most transferable patterns. Experiments on mathematical and logical reasoning tasks show modest improvements (2-5% absolute gains) over baselines, with the authors claiming that model-inherent circuits provide a more faithful and interpretable alternative to input-output distillation.

Critical strengths

S1. The focus on model-inherent reasoning circuits is a valuable contribution to interpretability research. By analyzing internal activations rather than just input-output behavior, the method provides insights into how models actually reason, which could inform future architecture design or training strategies.

S2. The causal intervention approach to identifying circuit components is principled and goes beyond correlation-based methods. By systematically perturbing internal activations and measuring the effect on outputs, the authors can distinguish between circuits that are necessary for reasoning and those that are merely correlated with it.

S3. The distillation of transferable circuits is a practical contribution: if reasoning patterns can be extracted from large models and injected into smaller ones, this could significantly reduce the computational cost of deploying reasoning-capable models. The method is more targeted than full knowledge distillation and preserves interpretability.

Critical weaknesses

W1. The reported gains (2-5% absolute) are within the noise range for most reasoning benchmarks, and there are no statistical significance tests or confidence intervals. Given that the baseline already uses strong models, the incremental benefit of circuit distillation is modest. The paper does not report ablation on circuit size, intervention threshold, or distillation strategy, making it impossible to assess whether the gains are robust or due to hyperparameter tuning on the test set.

W2. The theoretical assumptions are unverified. The paper assumes that reasoning circuits are modular and transferable across models, but provides no evidence for this claim. There is no analysis of whether the extracted circuits are actually reused across different problems, whether they generalize to out-of-distribution tasks, or whether they are specific to the training distribution. Without this validation, the "model-inherent" claim is speculative.

W3. The per-pair tuning requirement is heavy: the method requires analyzing the teacher model's internals for each problem-solution pair, which is computationally expensive and limits scalability. The paper does not report the computational cost of circuit extraction or compare it with simpler distillation methods. For a method claiming to reduce deployment costs, the extraction overhead is a significant practical limitation.

W4. The circuits are fixed after extraction and do not adapt to the student model's architecture or capabilities. This rigid transfer may fail when the teacher and student have significantly different internal representations. The paper does not investigate how circuit compatibility affects transfer success or whether circuits can be fine-tuned for the student model.

Questions for authors
  1. Can the authors provide statistical significance tests for the 2-5% gains and demonstrate that they exceed run-to-run variance across multiple seeds?

  2. What evidence do the authors have that reasoning circuits are modular and transferable across models? Is there analysis of circuit reuse across different problems or generalization to out-of-distribution tasks?

  3. It is not entirely clear what the term "circuit-level reasoning pattern" signifies in essence. One must ask whether it can truly describe the reasoning process; if so, does the possibility of heterogeneity—that is, different paths leading to the same conclusion—arise within such complex chains? Furthermore, how might such an issue be addressed?

Reviewer_PdBZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes model-inherent Reasoning Circuit Distillation (RCD). It uses EAP-IG to identify separate teacher and student circuits under three reasoning-transition modes. At response positions with high predictive entropy under the teacher, it aligns the logit changes induced by attenuating each model's own circuit. The resulting loss can be combined with a base distillation objective and does not require correspondence between individual teacher and student components. The paper also provides theoretical analyses of circuit recovery and effect alignment under stated assumptions, and evaluates RCD on mathematical reasoning and code generation.

Critical strengths
  1. Aligning circuit effects rather than individual components is a reasonable design choice. It allows the teacher and student to use their own circuits without requiring a component-wise mapping. The RCD loss can also be added to existing distillation objectives.
  2. The experiments cover mathematical reasoning and code generation, with several teacher-student pairs and base distillation objectives. Tables 1 and 2 show improvements in average performance across multiple settings. The analyses of circuit faithfulness, discovery stability, and random-edge interventions in Figure 2 provide useful supporting evidence.
Critical weaknesses
  1. The respective contributions of circuit discovery and effect alignment are not sufficiently clear. RCD discovers separate circuits and matches the logit changes caused by circuit attenuation, without component matching. The paper discusses both methods and includes them as baselines, but these comparisons alone do not isolate the contribution of each design choice. Figure 2(c) shows that the selected edges matter for prediction; it does not test whether these edges improve distillation more than alternative selections under the same loss. An explanation based on the existing analyses, or a representative controlled comparison, would help clarify the roles of edge selection, the three discovery modes, and high-entropy token selection.
  2. I find the meaning of preserving the student's intrinsic reasoning pattern unclear. In a different distillation setting, DistilGPT2 retains aspects of GPT2-small's behavior while reorganizing, compressing, or discarding internal components. Similar behavior therefore need not imply the same internal mechanism. Algorithm 1 fixes the discovered edge sets while the student's parameters are updated. Keeping these sets fixed does not by itself establish that their computational roles are preserved. Figure 2(b) measures stability across circuit-discovery runs, and Figure 3 measures teacher-student effect alignment. These results do not directly establish which aspects of the student's pre-distillation computation are preserved. Could the authors clarify what is intended to be preserved, what changes are allowed, and how the existing evidence or a concrete example supports this claim?
  3. The conditions under which RCD is most effective need further discussion. Appendix A compares circuit structures across model families, whereas the distillation experiments in Tables 1 and 2 pair teachers and students within the same model family. Based on the existing results, I suggest discussing how the teacher-student capability gap, the representativeness of the circuit-discovery data, and their match to the target task may affect the gains or limit the method's applicability.
Questions for authors
  1. Compared with Causal Distillation [1] and Circuit Distillation [2], which design choices account for RCD's gains? Can the existing analyses or a representative controlled comparison clarify the roles of circuit selection and effect alignment?
  2. Based on the current results, which teacher-student pairs and task settings are most suitable for RCD? How might the representativeness and task relevance of the circuit-discovery data affect its performance?

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

Related papers

Discussion (0)

Sign in to comment.