acceptodds
Under review as a conference paper at ICLR 2027

GRAD School: Teaching a Student Its Teacher's Circuit

Abstract

Knowledge distillation supervises a student on its teacher's answers, not on the computation behind them. We introduce GRaph Attribution Distillation (GRAD), which makes the teacher's circuit a training target: differentiable attribution graphs over MLP neurons are coarsened into supernodes grouped by function, so both models share labelled nodes, and the student is trained, alongside standard KD, to route influence between them as the teacher does. Distilling Llama-3-8B into Llama-3.2-1B on two-digit addition, GRAD improves on standard KD on every family, beats it and three other baselines on the held-out mean (0.55 against at most 0.50), lifts never-trained multiplication from 0.51 to 0.95, and raises a 3B student's held-out mean from 0.83 to 0.88; a scrambled teacher target does not reproduce the gain. Mechanistically, GRAD's student develops features as specific as a label-trained student's without gold labels, and on multiplication it corrects a failure of standard KD, whose students hedge instead of answering in the evaluated input format, the format in which the graph term is computed; a loss that rewards answering in that format produces the same correction. These analyses yield a checklist for objectives that aim to transfer a mechanism.

Primary area:
interpretability and explainable AI
Keywords:
Distillation, Attribution Graph

Reviews

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

The paper proposes GRaph Attribution Distillation (GRAD), which supplements standard knowledge distillation with a Jensen–Shannon loss matching attribution-routing profiles between functionally aligned groups of MLP neurons. Task-specific ANOVA probes define shared supernodes across an 8B teacher and 1B/3B students. Experiments train on two-digit addition and evaluate several held-out arithmetic families, with representation-matching, scrambled-target, and compute controls. A detailed audit shows that the largest 1B multiplication gain arises from correcting a tendency to hedge under a BOS-token format mismatch; a simple commitment loss reproduces the improvement. The paper therefore contributes both a graph-based objective and a concrete investigation of what an apparent mechanism-transfer gain actually measures.

Critical strengths

The strongest contribution is the explicit causal investigation of the apparent generalization gain. Section 5 separates commitment from correctness, evaluates both BOS conventions, transplants layers, patches residual states, and tests a targeted commitment loss. Table 3 shows that this simple loss reproduces the multiplication improvement, and Table 8 reports the absence of the advantage under consistent BOS training. These controls expose a concrete alternative explanation and make the paper informative even where the original mechanism-transfer interpretation is weakened.

Critical weaknesses

The practical advantage of GRAD itself remains unestablished once the identified format confound is controlled. Under the standard BOS convention, Table 8 does not show a benefit over KD; under the mismatched convention, the much simpler commitment objective matches or exceeds GRAD in Table 3. Meanwhile, a warm-cache GRAD step costs about 14 times a KD step, with substantial additional teacher-cache construction cost. The 3B result is promising, but has not received the same commitment/format audit. Thus, the evidence currently supports the methodological audit more strongly than a broadly useful circuit-distillation method, and the title and conclusion should reflect that balance.

Questions for authors
  1. For the 3B student, do the operand-count gains persist when training and evaluation use the same BOS convention and when KD receives the commitment loss? This comparison would help establish whether any remaining benefit is specific to the graph objective.
  2. How do feature specificity and routing differ between GRAD and the commitment-only control? Do aligned supernode interventions show a teacher-like computational change beyond the commitment effect, and which exact arithmetic labels or masks are required to construct the training target?
Reviewer_kvC1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In short, the authors proposed a new distillation method, GRAD, to transfer a teacher’s computational circuit to a student, but failed to justify why circuit distillation is necessary, given that GRAD offers no performance improvement over standard KD under consistent BOS settings while incurring substantially higher computational costs. More specifically, the authors proposed GRAD to distill the teacher's computational circuit into a student model. Both the teacher and student were represented as graphs, where supernodes grouped neurons according to a predefined set of task-specific functions. The edge weights represented the influence of source supernodes on target supernodes. The student was then trained using a combination of conventional knowledge distillation (KD) loss and a graph loss that encourages the student to reproduce the teacher's edge weights. The authors reported a substantial performance improvement on a multiplication task involving a two-digit number and a one-digit number. However, their own analysis revealed that this improvement primarily originates from an unconventional handling of the beginning-of-sequence (BOS) token. Specifically, BOS was omitted when computing the KD loss but included when computing the graph loss and during evaluation. Under consistent BOS settings, GRAD failed to outperform standard KD. Moreover, GRAD incured approximately 14 times the computational cost. Consequently, the paper failed to establish why circuit distillation is necessary or practically beneficial, particularly given its substantially higher computational cost.

Critical strengths

Idea of representing neural networks as graphs with corresponding functional supernodes is interesting.

Critical weaknesses

W1) The overall quality of writing and presentation falls well below acceptable academic standards. Many sentences were extremely difficult to understand due to (1) excessive compression of technical information into single sentences, (2) insufficient logical connections and transitions between ideas, and (3) frequent use of ambiguous referencial expressions. Examples included lines 107-125 (pp. 2-3), lines 109-192 (p. 4), and lines 231-243 (p. 5). Furthermore, the paper was not well-organized. (4) Some statements in the earlier sections only become understandable after reading the entire paper, forcing readers to reread the paper several times to reconstruct the authors’ intended arguments. Also, (5) the main text relies excessively on experimental results presented in the appendix, forcing readers to repeatedly switch back and force between the text and the appendix to follow the arguments. On top of that, (6) important mathematical notation, such as (W_U) and (h_{-1}^{(L)}) (lines 176–177, p. 4), was introduced without explicit definitions. Also, (7) figures also lack essential explanations. For example, Figure 2 did not clearly explain the axes of the squares in the figure or the quantities represented by the colors, making the figure difficult to interpret. Overall, these issues substantially hinder the readability and comprehensibility of the paper, making it unnecessarily difficult for readers to follow.

W2) I would like to highlight several logical flaws and insufficiently justified methodological choices. As noted in the summary, (1) the authors failed to establish a compelling motivation for circuit distillation; (2) the experimental findings did not substantiate the paper's original goal. In particular, GRAD provided no performance improvement over standard KD under consistent BOS settings, despite incurring substantially higher computational costs. This raises a fundamental question about the necessity and practical value of GRAD; and (3) the graph construction procedure is highly complex, yet its key design choices, including the selection of neurons for functional supernodes and the computation of edge weights, were insufficiently justified, either conceptually or empirically.

Questions for authors
  1. Justification for the six functional categories: How do the authors justify the assumption that the computational circuits of both the teacher and student can be adequately represented by the six predefined functional categories? Could the authors conduct an ablation study comparing the current six-category setting with a three-category setting (removing the first-operand ones digit, second-operand ones digit, and sum ones digit) and a nine-category setting (additionally including the first-operand tens digit, second-operand tens digit, and sum tens digit)?

  2. Generalization across model families: The authors used teacher and student models from the same Llama family, differing primarily in model size. Could the authors evaluate GRAD on teacher–student pairs from different model families, such as Llama-8B to Gemma-1B? In particular, how does GRAD compare with standard KD in such a setting?

  3. Comparison with the closest prior work: The authors explicitly identified Circuit Distillation (Wadhwa et al., 2025) as the closest prior work. Could the authors include Circuit Distillation as an additional baseline to demonstrate the advantages of GRAD over this closely related approach?

Reviewer_weH8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposed GRAD, an algorithm which utilizes Jenson Shannon Divergence between teacher and student routing profiles over functional “supernodes” in differentiable attrition graphs over raw MLP neurons. The authors evaluate the effectiveness of the framework on Llama models on a multiple datasets

Critical strengths

Leveraging attrition graphs for training is interesting The authors performed extensive experiments on understanding where the gains of the algorithms come from, ranging from layer analysis to when the decision is made

Critical weaknesses
  1. The authors claim novelty based on the fact unlike standard knowledge distillation, which works ONLY on the output of a student, the proposed method leverages internal computation as well. But, there are plenty of papers existing in this domain, like Knowledge distillation with Knowledge transfer with jacobian matching[1], Distilling Knowledge from Graph convolutional networks[2] all leverage the internal working of a teacher to enhance distillation

  2. The “supernode” component of the algorithm is extremely specific to this task( two digit addition and other related ones in the paper). Most Knowledge distillation algorithms present in literature work across datasets in a given domain ( eg vision, graph etc). The authors should have demonstrated how the algorithm can work across any kind of text dataset.

  3. Insufficient baselines- The paper has only compared against a limited number of older baselines like tinyBERT. For a more comprehensive understanding of framework effectiveness, the paper should add more baselines like MiniLLM[4], GKD[5] etc. The authors are requested to compare the proposed framework with the above baselines and the other baselines present in the related work

  4. The authors have primarily done experiments only on Llama models ( 8B and 1B). The framework also needs to evaluated on other LLM model families like qwen, mistral and experiments on cross-model family distillation should also be conducted.

  5. The authors themselves have stated that the largest claim comes from the unconventional tokenization convention, so it seems the gains from the proposed algorithm is minimal

Questions for authors

Some references to add based on points listed in weaknesses

References Srinivas, S., & Fleuret, F. (2018, July). Knowledge transfer with jacobian matching. In International conference on machine learning (pp. 4723-4731). PMLR. Yang, Y., Qiu, J., Song, M., Tao, D., & Wang, X. (2020, June). Distilling knowledge from graph convolutional networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 7072-7081). IEEE. Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., ... & Liu, Q. (2020, November). Tinybert: Distilling bert for natural language understanding. In Findings of the association for computational linguistics: EMNLP 2020 (pp. 4163-4174). Gu, Y., Dong, L., Wei, F., & Huang, M. (2024, May). Minillm: Knowledge distillation of large language models. In International Conference on Learning Representations (Vol. 2024, pp. 32694-32717). Tan, S., Tam, W. L., Wang, Y., Gong, W., Zhao, S., Zhang, P., & Tang, J. (2023, July). Gkd: A general knowledge distillation framework for large-scale pre-trained language model. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 5: Industry Track) (pp. 134-148).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.