acceptodds
Under review as a conference paper at ICLR 2027

Intervention-Aware Compression of Neural Networks via Causal Abstraction

Abstract

Model distillation and mechanistic interpretability share a common objective: replacing a pre-trained model (the teacher) with a more compact algorithm (the student) that preserves relevant aspects of its behavior. Distillation typically targets input-output agreement, while mechanistic interpretability instead focuses on creating compact representations that are causally faithful to a model's internal computation, often on a restricted set of inputs. Prior work suggests that constraining a student to reproduce aspects of a teacher's internal computation can improve distillation. We pursue this direction and build on recent advances in causal abstraction that provide differentiable objectives for measuring whether a smaller model constitutes a causally faithful abstraction of a larger one. We use these objectives to develop Causal Abstraction Distillation (CAD) and its practical variant CAB, which trains student Transformers to match not only the teacher's input-output behavior, but also its responses to interventions and perturbations applied to inputs and intermediate activations. This interventional training encourages the student to preserve the teacher's computational structure rather than merely imitate its predictions. Across encoder and decoder-only models, we find that CAB improves intervention-response robustness in several settings while maintaining competitive task performance. CAB can also improve generalization to related datasets, with benefits extending to 14B-to-3B Qwen2.5 distillation.

Primary area:
interpretability and explainable AI
Keywords:
Interpretability, Distillation, Causality

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_Qd9AOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a method to make teacher and student models share the same computational behaviour under interventions and perturbations of inputs and representations. Perturbations of the teacher models representations are "transported" to the student model using learned layer wise abstraction maps. They alternate between learning these maps with fixed student and teacher representations, and learning the student with a fixed map. They compare this with other KD methods on a variety of LLMs on SQuaD, GLUE and yelp review classification.

Critical strengths
  1. The paper's framing of model distillation as aligning student and teacher responses to interventions on matched representations is interesting and I think may lead to interesting further work.

I would very much like to see the implications of this work on mechanistic interpretability worked through more thoroughly.

  1. I found this paper very easy to follow, it's experiments section was especially clear and reasoning and assumptions underlying choice of evaluation metric were laid out and reasoned through.
Critical weaknesses
  1. The most interesting and consequential ideas in this paper are left to further work.

The framing of this paper, at least in the abstract, leans heavily on the potential use of this for mechanistic interpretability, but this actually only makes up a few lines at the start and end.

  1. Figure 2 shows improved matching on input states doesn't necessarily result in better output mapping, and quite a few of your results have overlapping error bars with the other KD methods. Is there a good reason why someone who doesn't care about the computational/causal fidelity of their student model to the teacher should change their mind and start caring about it, if their only aim is to replace a larger model with a smaller one? Does your method result in distilled models that generalise better for instance?
Questions for authors

In section 4, how does one select a partition? How does the selection of the partition effect downstream results?

4 comments

We thank the reviewer for the positive assessment of our causal-abstraction framing, the clarity of the presentation, and the experimental evaluation. We also appreciate the constructive suggestions regarding mechanistic interpretability, practical benefits beyond causal fidelity, and partition selection.

W1: The most interesting and consequential ideas in this paper are left to further work. [...]

We agree that the distinction between our demonstrated contributions and potential applications to mechanistic interpretability should be clearer. We have revised the Introduction (Section 1) to clarify that mechanistic interpretability motivates our approach, while the present work focuses on causal abstraction as a training and evaluation principle for model distillation. Constructing human-interpretable mechanistic explanations remains future work.

The present contributions are:

  • Unified framework: A causal-abstraction error formulation connecting existing distillation objectives, with formal derivations and assumptions (Section 4; Appendix N).
  • Training: Learned abstraction maps that transport teacher interventions and enforce consistency across multiple downstream hidden boundaries and the output (Section 4).
  • Evaluation: Complementary intervention-response metrics, CAD_IO and CAD, that reveal differences not captured by conventional task performance or clean teacher agreement. For example, on MNLI-m, CAB-FULL and TED achieve comparable accuracy and agreement, while CAB-FULL reduces CAD_IO by 9.9% and CAD by 22.3% (Table 1).

These clarifications distinguish the contributions established in this paper from their longer-term implications for mechanistic interpretability.

W2: Figure 2 shows improved matching on input states doesn't necessarily result in better output mapping, and quite a few of your results have overlapping error bars with the other KD methods. Is there a good reason why someone who doesn't care about the computational/causal fidelity of their student model to the teacher should change their mind and start caring about it, if their only aim is to replace a larger model with a smaller one? Does your method result in distilled models that generalise better for instance?

We agree that improved internal consistency does not necessarily translate into better output behavior, as illustrated by Figure 2. However, our results show that intervention-aware distillation can also provide practical benefits beyond internal causal fidelity.

  • Cross-dataset generalization: We evaluate transfer to related datasets not used during distillation. Relative to DIITO, CAB improves accuracy by up to 3.84 percentage points on MRPC, 0.35 points on SNLI, and 1.79 points on QNLI. These gains depend on the alignment strategy (Section 5.3, Table 2; Appendix G, Table 33).
  • Adversarial robustness: We additionally evaluated SST-2, where CAB did not consistently outperform DIITO on the original task, across three adversarial benchmarks. On AdvGLUE-SST2, CAB achieves slightly higher accuracy (30.81% vs. 30.14%), higher teacher agreement (78.51% vs. 77.03%), and 5.3% lower adversarial CAD_IO. On TF-B and TF-R, CAB also reduces adversarial CAD_IO by 1.9% and 3.0%, respectively, although its accuracy is slightly lower than DIITO on both benchmarks (Appendix G.1, Table 34).

Thus, CAB's potential practical value extends beyond internal causal fidelity: it can improve transfer performance and preserve teacher behavior under distribution shifts. However, these benefits are not universal, and we do not claim that CAB consistently outperforms conventional distillation when clean-task accuracy is the sole objective.

Q1: In section 4, how does one select a partition? How does the selection of the partition effect downstream results?

We have clarified the partitioning procedure in Section 4 and Appendix B.

  • Partition selection: Given a teacher with L layers and a student with H layers, FULL alignment divides the teacher into H approximately equal-sized contiguous blocks, each corresponding to one student layer (Appendix B, Table 6). Thus, the student architecture determines the number of blocks, while their boundaries are fixed by the alignment scheme.
  • Effect of partitioning: The partition determines how much teacher computation each student layer must approximate. Coarser partitions compress more computation into each student layer, potentially making intervention consistency harder to preserve. Different alignment choices also affect downstream performance, as demonstrated by the FULL, MIDDLE, and LATE comparisons (Tables 1–2).

Our layerwise reconstruction analysis shows higher errors at S4–S5 than at S1, suggesting that some parts of the teacher may be harder to approximate (Appendix E, Figure 3). While this does not establish an optimal partitioning strategy, it motivates exploring non-uniform or automatically selected partitions in future work.

We thank you again for your time. Please let us know if we have answered your concerns, or if you need further clarifications from our side.

Thank you for your comprehensive and enlightening response. The updated manuscript has improved it on several fronts:

  1. The contribution of the work is much clearer
  2. The layer-wise analysis analysis answers my question of how partition selection is performed and impacts performance
  3. Several parts of the procedure which were not totally clear (fitting of abstraction maps) have now been clarified.

I think this work will be of interest, especially as to how it may inspire further mechanistic interpretability work. As such I raise my score and recommend acceptance.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for your positive assessment of our revisions, and for increasing your score.​

We also appreciate your encouraging comments.

Reviewer_qJyMOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to compress a neural network while preserving the teacher's responses to interventions, as well as its predictions. The authors propose Causal Abstraction Distillation (CAD), which learns layerwise maps from teacher to student representations, uses them to transfer interventions, and matches downstream hidden states and outputs. Its practical variant CAB combines this objective with a DIITO-style activation-swap loss. Experiments with BERT, DeBERTa, and Qwen2.5 show better teacher–student consistency under interventions in several settings, while keeping task performance competitive. Some cross-dataset transfer results also improve.

Critical strengths
  1. Preserving a teacher's responses to interventions is a useful goal for model compression. The paper gives a clear formulation of how to compare these responses at hidden layers and at the output.

  2. The experiments include output-, feature-, and intervention-based distillation baselines, several model families and tasks, and cross-dataset transfer. Results are reported over multiple seeds.

Critical weaknesses
  1. The benefit over existing intervention-based distillation needs more evidence. The original DIITO paper already includes a variant that matches final hidden representations after interventions. On SST-2 and MRPC, adding CAB's hidden-state loss improves internal consistency, but not mean CAD_IO. The same ablation on MNLI or QNLI, where CAB performs better, would help show whether this loss explains the gains.

  2. CAD_IO compares teacher and student outputs after an input perturbation. It does not directly compare how much their outputs change from the clean input. Figure 2 examines response changes on SST-2 using internal interventions, but the corresponding comparison under the input perturbations used by CAD_IO is missing on MNLI and QNLI.

Questions for authors
  1. Could the authors repeat the bridge-only, output-only, and full-CAB ablations from Appendix H on MNLI or QNLI? Does the DIITO baseline include the original variant with hidden-state matching after interventions?

  2. For the same input perturbations used by CAD_IO on MNLI or QNLI, how closely do teacher and student agree in their output-probability changes from the clean input? Could the authors also extend the internal-intervention diagnostics in Figure 2 to these tasks?

  3. How much extra time and memory do map fitting and interventional computations require? Could the authors compare total training time, including map fitting, and peak memory with KD and DIITO on the same hardware and with comparable batch sizes?

5 comments

We thank the reviewer for the constructive suggestions regarding comparisons with DIITO, additional intervention-response analyses, and computational cost. We have conducted new experiments and expanded the appendix to address these concerns, as detailed below.

Q1 (W1): Could the authors repeat the bridge-only, output-only, and full-CAB ablations from Appendix H on MNLI or QNLI? Does the DIITO baseline include the original variant with hidden-state matching after interventions?

We conducted additional experiments addressing both points.

  • Objective ablations: We extended the bridge-only, output-only, and full-CAB ablations to QNLI and both MNLI validation sets. Adding the hidden CAD term reduces hidden-response error by 28.8% on QNLI, 33.8% on MNLI-m, and 33.7% on MNLI-mm relative to output-only CAB, while maintaining comparable task accuracy. These results support the contribution of hidden-boundary supervision to internal intervention consistency (Appendix K, Tables 41–42).
  • DIITO with hidden-state matching: Our original DIITO baseline uses output matching under interchange interventions. We acknowledge that Wu et al. (2022) also introduce a variant that matches final BERT-layer representations after interventions (Appendix A.2). Unlike this variant, CAB enforces consistency across multiple downstream hidden boundaries using learned abstraction maps.

To address the reviewer's concern, we additionally evaluated DIITO with post-intervention hidden-state matching on QNLI. Although this improves DIITO, full CAB achieves higher accuracy (89.57% vs. 89.27%) and lower CAD_IO (0.0411 vs. 0.0487) and hidden-response error (0.2865 vs. 0.3171). These results are reported in Appendix I, Table 37.

Q2 (W2): For the same input perturbations used by CAD_IO on MNLI or QNLI, how closely do teacher and student agree in their output-probability changes from the clean input? Could the authors also extend the internal-intervention diagnostics in Figure 2 to these tasks?

Thank you for this suggestion. We extended both analyses to QNLI using three seeds (Appendix J, Table 38 and Figure 5).

  • Input perturbations: Under the same perturbations used for CAD_IO, full CAB reduces mean output-probability-change MSE by 16.1% relative to output-only DIITO (0.01244 vs. 0.01483) and by 10.7% relative to DIITO with hidden-state matching (0.01244 vs. 0.01393). The reductions hold across all three paired seeds (Table 38).
  • Internal interventions: Extending Figure 2 to QNLI, we compare CAB and output-only DIITO using matched interventions at five aligned sites. CAB achieves lower mean hidden-response error and output-response MSE at every site (Figure 5). Unlike SST-2, the improvement in internal consistency is accompanied by improved output-response consistency across all evaluated sites.

Q3: How much extra time and memory do map fitting and interventional computations require? Could the authors compare total training time, including map fitting, and peak memory with KD and DIITO on the same hardware and with comparable batch sizes?

We added a computational-cost analysis on MNLI and SST-2 using NVIDIA H100 GPUs (Appendix D.4, Tables 28–29).

  • End-to-end cost: Table 28 reports total training time and peak GPU memory for KD, LWD, TED, DIITO, and CAB. CAB-FULL requires approximately 1.8× the training time and 1.6× the peak memory of DIITO in these runs, which also include diagnostic computations not required for training.
  • Matched-batch comparison: To isolate training overhead, we benchmarked methods at batch sizes of 16 and 32. CAB-FULL requires 1.22–1.24× the time per update and 1.15–1.20× the peak memory of DIITO. Abstraction-map fitting adds approximately 23–27 seconds for FULL alignment and 46–53 seconds for MIDDLE/LATE alignment (Table 29).

These results quantify the additional cost of map fitting and interventional training, while distinguishing the end-to-end measurements from the matched-batch comparison.

Thank the authors for the detailed replies and additional experiments.

  1. Contribution beyond DIITO: The stronger baseline and ablations are helpful. CAD clearly improves hidden-state consistency, but its added output benefit remains unclear. On QNLI, bridge-only achieves nearly the same accuracy as full CAB (89.54 vs. 89.57) and the same reported CAD_IO (0.0411). If I understand correctly, this suggests that most of the output gains come from the DIITO-style bridge term, leaving limited evidence of added output benefits beyond DIITO.

  2. Response matching: The new response-change analysis and internal-intervention experiments address my concern about the missing direct evaluation. They show better response matching than the evaluated DIITO baselines. However, CAB's response-change error remains higher than the zero-response reference. I appreciate that the revised appendix acknowledges this limitation. Please keep the main-text claims equally cautious: the results support improved response matching, but do not yet establish strong preservation of the teacher's responses.

  3. Computational cost: The cost analysis addresses my request and usefully separates training from additional diagnostics. However, full CAB requires about 22–24% more time per update and 15–20% more peak memory than DIITO under matched conditions. Given the small accuracy gains and limited added output benefit over bridge-only, it remains unclear whether the gains justify the extra cost.

Overall, the new experiments clarify the results, but my concerns about the added contribution and the trade-off between cost and benefit remain. I will keep my rating at 2 (weak rejection).

AuthorsOpenReview ↗

Response to Reviewer qJyM

We thank the reviewer for their quick follow-up and thoughtful assessment. We would like to emphasize that our primary contribution is a causal-abstraction framework for distillation that improves internal intervention consistency, while maintaining competitive task performance.

Objective ablations. As also suggested by reviewer vGZQ, we added objective ablations on QNLI and MNLI (Appendix D.5, Table 30). Compared with bridge-only, CAD-only reduces hidden-response error by approximately 32–33%, with slightly lower but comparable accuracy:

DatasetCAD-onlyBridge-only
QNLI88.89%89.54%
MNLI-m82.06%82.37%
MNLI-mm82.37%82.80%

We agree that the response-matching results show improvements relative to the evaluated baselines under the tested perturbations. We have clarified in Section 5.2 that these results establish relative improvement.

Computational cost. We acknowledge the trade-off. Full CAB combines the bridge and CAD objectives. This improves internal consistency while keeping output-level performance competitive, but it costs more to compute. CAD-only is a lower-cost alternative: on the matched MNLI benchmark, it needs about 12% less time per update and 21% less peak GPU memory than DIITO (Appendix D.5, Table 32).

Reviewer_LCXSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper explores causal abstraction distillation with the newer techniques in the field. Paper mirrors the set up of DIITO and has done a reasonable comparison over the baseline methods. Authors have shown that the improved technique would produce a lower CADIO over the student model, and a better cross-dataset transferability over NLP tasks.

Critical strengths
  • The paper is well-motivated and well executed, the experiments are done in good faith to show a clean comparison with the prior works. I appreciate the clean and honest presentation of the result, including cases where CAB underperforms.
  • The intuition behind the use of CAD objective, which intended to enforce the commutativity is reasonable to me.
  • Clean paper writing with few AI-slops.
Critical weaknesses
  • Though the author has correctly acknowledge works like DIITO (Wu et al) as the prior work, I found the intro slightly overstated its novelty on top of the existing line of works, and .
  • The strength of CAB over the other methods do not seem very obvious to me from table 1.
Questions for authors

I found this paper to be informative and should be accepted as it explored a direction that is reasonable and essential for the line of work over using mech-interp for model distillation. The gain to me has caveat and marginal for a clear accept, tho. The only major question I have:

  • A head-to-head comparison on the training cost with the baselines?
4 comments

We thank the reviewer for the positive assessment and constructive feedback. We address each point below.

W1: Though the author has correctly acknowledged works like DIITO (Wu et al) as the prior work, I found the intro slightly overstated its novelty on top of the existing line of works, and .

Thank you for highlighting the need to clarify our contributions. We have revised the Introduction (Section 1) to acknowledge more explicitly our relationship to prior causal-distillation work and distinguish three contributions:

  • Distillation as causal abstraction: Building on DIITO, we formulate distillation through a general causal-abstraction error objective that recovers DIITO and several other existing distillation methods as special cases under explicit assumptions. Following reviewer vGZQ's suggestion, we added formal derivations establishing these connections and clarifying the conditions under which they hold (Section 4; Appendix N).
  • Causal Abstraction Distillation: Unlike DIITO's output-oriented interchange objective, CAD uses learned abstraction maps to transport teacher interventions into the student and enforce consistency across multiple downstream hidden representations and the output. Although the framework supports arbitrary admissible teacher interventions, our experiments use sampled interchange interventions to have fair comparison.
  • Interventional evaluation of compressed models: We evaluate compressed models using both map-independent intervention-response robustness (CAD_IO) and map-dependent causal-abstraction diagnostics (CAD). These reveal differences not captured by task accuracy or clean teacher agreement. For example, on MNLI-m, CAB-FULL and TED achieve comparable accuracy and agreement, while CAB-FULL reduces CAD_IO by 9.9% and CAD by 22.3% (Table 1; Section 5.2).

These revisions clarify that our contribution extends existing causal-distillation methods rather than introducing causal distillation itself. In particular, while DIITO builds on interchange intervention training (IIT), our approach builds on Causal Abstraction Error (CAE; Méloux et al., 2026), a distributional measure of approximate causal abstraction that has demonstrated empirical advantages in evaluations on complex systems. This provides a distinct foundation for our formulation of intervention-aware distillation.

W2: The strength of CAB over the other methods do not seem very obvious to me from table 1.

We agree that some differences in Table 1 are small relative to seed variability. To clarify the comparisons, we added seed-paired differences with 95% confidence intervals for MNLI and QNLI (Appendix D.3, Tables 26–27).

On both MNLI evaluation sets, CAB-FULL reduces CAD_IO and CAD relative to DIITO and TED, with confidence intervals excluding zero. For example, on MNLI-mm, compared with DIITO, CAB-FULL improves accuracy by 0.36 percentage points [0.13, 0.59] and reduces CAD_IO by 0.0111 [−0.0152, −0.0071]. On QNLI, CAB-FULL reduces CAD_IO by 0.0080 [−0.0091, −0.0068] and CAD by 0.1182 [−0.1297, −0.1068] relative to DIITO.

These results clarify that CAB's primary advantage is improved intervention-response robustness and internal consistency, rather than consistently higher task accuracy. The gains remain task- and alignment-dependent, as discussed in Section 5.2 and Appendix D.3.

Q1: A head-to-head comparison on the training cost with the baselines?

Thank you for this suggestion. We added a computational-cost comparison across KD, LWD, TED, DIITO, and CAB on MNLI and SST-2 using NVIDIA H100 GPUs (Appendix D.4, Tables 28–29).

  • In the end-to-end measurements, CAB-FULL requires approximately 1.8× the training time and 1.6× the peak GPU memory of DIITO. However, these runs include additional forward passes for clean-reference and mapped clean-control diagnostics, used to analyze differences between clean and intervened computations. These passes are not required to compute the training loss.
  • Under matched batch sizes of 16 and 32, CAB-FULL requires 1.22–1.24× the time per update and 1.15–1.20× the peak memory of DIITO. The additional abstraction-map fitting time is also reported in Table 29.

We further isolate the computational cost of individual objectives in Appendix D.5 (Table 32). Compared with DIITO output-only, CAD-only requires approximately 12% less time per update and 21% less peak GPU memory, while full CAB requires approximately 23% more time per update and 20% more peak GPU memory.

These measurements quantify CAB's additional computational overhead and clarify the trade-off between training cost and improved interventional consistency.

We thank you again for your time. Please let us know if we have answered your concerns, or if you need further clarifications from our side.

Reviewer_LCXSOpenReview ↗

Thank you for clarifying

The authors response has partially addressed my concerns. Although the strength of the method is still not obvious given the extra cost on the run time and memory, but the method shows potential and paper is clean. Adjusted my soundness score to 3 while maintaining my overall assessment of the paper.

AuthorsOpenReview ↗

Official Comment by Authors

Thank you for your feedback and for increasing your soundness score. We appreciate your recognition of the potential of our method.

Reviewer_vGZQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper frames knowledge distillation (KD) as a causal abstraction (CA) task. Unlike interpretability, where an abstraction must be discovered in a fixed network, distillation lets one choose the student-teacher alignment and learn both the student and the abstraction maps. Based on this, and on the thesis that preserving some teacher’s internal structure improves distillation, the authors propose to train the student model to be an approximate causal abstraction of the teacher. The authors claim three main contributions. (i) A framework for distillation as causal abstraction (Sec. 4; l.149-161, l.211-222), where teacher and student are viewed as low- and high-level SCMs, respectively. Each student layer is aligned with a contiguous block of teacher layers, and a learned abstraction map τj\tau_j translates the teacher representation at the end of the block into the student's representation space (Eq.3). To enforce commutativity, the authors define a new penalty term called Causal Abstraction Distillation (CAD, Eq.7): an intervention on a teacher block boundary is transported to the student through the abstraction map (Eq.6), both models are run forward, and CAD is a weighted sum of discrepancies between the mapped teacher states and the corresponding student states, measured at the aligned boundaries downstream of the intervention. Since both the type of intervention and where consistency is checked can vary, the authors argue that CAD recovers standard KD, intermediate-feature distillation, distillation on augmented or adversarial inputs, and DIITO (Wu et al., 2022) as special cases. (ii) In practice, they propose to train the student model using a loss combining a CAD term, a DIITO-style "bridge" term (which matches teacher and student outputs under independent interventions) together with a task loss and a KD term (Eq.9). They name the combination of CAD and DIITO-style terms CAB (Eq. 8). (iii) Besides task accuracy and teacher-student agreement, the authors evaluate students with two interventional metrics. CADIOCAD_{IO}: a map-independent divergence between teacher and student outputs when the same perturbation is applied to their inputs (masking for the encoders; synonym, deletion and antonym edits for Qwen). And CADCAD: the map-dependent quantity of Eq. 7.

The paper compares the proposed method with four baselines: knowledge distillation (KD), layer-wise distillation (LWD), task-aware layer-wise distillation (TED), and a task-specific adaptation of DIITO as the interventional baseline. On encoder models (BERT and DeBERTa), CAB lowers CADIOCAD_{IO} on several tasks at comparable accuracy with exceptions (e.g., SST-2 and RTE on BERT, where DIITO is lowest). CAB sometimes improves transferability more than DIITO to related datasets within the same task category. On decoder-only models (Qwen2.5 14B to 3B), CAB gives marginal improvements over KD. Finally, the paper points out that hidden-level consistency doesn’t always carry over to the output.

Critical strengths
  1. Relevant question and framing. The paper starts from the thesis that making the student reproduce aspects of the teacher’s internal computation improves distillation. It asks whether the teacher’s responses to interventions on internal layers (aligned at several depths) are a good kind of structure to preserve. The question is relevant as neither accuracy nor teacher-student output agreement capture the internal student structure. The proposed answer, training the student to be an approximate causal abstraction of the teacher, is also sensible. In fact, unlike for post-hoc interpretability where alignment and abstraction maps must be discovered, in distillation they can be designed and trained. The direction is therefore of interest for both the neural compression and the causal-interpretability communities.

  2. Broad experimental setup. The experimental suite is broad and covers two kinds of architectures (encoders vs decoder-only), several tasks (GLUE and SQuAD v2), both depth and width compression, and transfer to similar tasks. The authors report seed variation and they are transparent on counter-examples such as SST-2 as well as on baseline's CAD being not informative. On several tasks the proposed method shows CADIOCAD_{IO} reductions (BERT MNLI, QNLI, CoLA) and a clear F1 gain on SQuAD v2, at no accuracy cost.

Critical weaknesses
  1. In comparison to previous work, the contribution is incremental. The paper names DIITO (Wu et al. 2022) as its closest work, which already frames distillation as making the student a causal abstraction of the teacher via interchange interventions. Over DIITO, the paper claims two improvements: "causally informative" teacher interventions, and consistency enforced at multiple depths through learned abstraction maps (the CAD term). The first is never tested since all experiments use the same class of interventions as DIITO, i.e., interchange interventions with randomly sampled source inputs. The second is never evaluated on its own. All experiments use Eq. 9, which contains both DIITO's bridge term and the CAD term, and the paper gives no justification for this practical variant (denoted CAB). This also weakens the unified view claimed in Sec. 4.2: if CAD generalizes DIITO, it should work without the DIITO term, yet CAD alone is never compared. In connection to this, the only isolation of its effect (App. H, Tab. 27) shows that the hidden CAD term lowers hidden-response error, but does not improve CADIOCAD_{IO} nor task score. Moreover, Tab. 27 reports identical rows for two different SST-2 conditions (bridge-only and output-only) without explanation. Elsewhere, gains over DIITO are often small and about a third of the comparisons are within one SD, with DIITO having the lowest CADIOCAD_{IO} on BERT SST-2 and RTE (Tab. 12, 16).

  2. Several aspects need a more precise specification, in particular the unifying framework and the abstraction maps. The authors state that they "derive a unified objective" (l.072-074) of which several KD methods are special cases, but no formal derivation is provided. While one can intuitively see how varying the intervention μ\mu and the evaluation set K(μ)\mathcal{K}(\mu) relates Eq. 7 to these methods, phrased as a main contribution the unification would benefit from a formal proposition or derivation. Most importantly, the family of abstraction maps τj\tau_j is never specified (e.g., linear, MLP, applied per token or pooled). Since sufficiently expressive maps make almost any network an abstraction of any algorithm (Sutter et al., 2025), the abstraction claim, and the map-dependent CAD metric, cannot be interpreted without it. How the maps are trained is also contradictory, Sec. 4.1 describes alternating map fitting and student training (l.183-184), while Sec. 5.1 describes fitting the maps once and then freezing them (l.288-292).


Update after rebuttal: The authors addressed most of my concerns by adding several ablations, a full derivation and proof of their unified-framework claim, and by removing the inconsistent parts. While CAD alone has a small impact on student's output-level accuracy, its practical variant, CAB (which includes a DIITO-style term), improves over DIITO alone. Furthermore, I think this line of research is interesting and might serve as motivation for future work on causality and interpretability. I am raising my score to 3 (weak acceptance).

Questions for authors
  1. Is the main contribution the CAD penalty or the CAB variant, which also includes DIITO’s bridge term? If the former, please add an ablation separating the CAD loss from the bridge term, ideally on tasks where gains are claimed. Relatedly, why is the practical CAB variant needed?

  2. What is the family of the abstraction maps and why is it restrictive enough to avoid vacuous abstractions (Sutter et al., 2025)? Are the maps fitted once and frozen (l.288-292) or refit alternately during training (l.183-184)? Also, which maps produce CAB’s reported CAD?

  3. Provide a formal statement or derivation of the unified view: for each recovered method, the intervention, evaluation set, divergence, map and (possibly) conditions under which equality holds. Otherwise, soften the claim to "related to".

  4. Since all experiments use interchange interventions with randomly sampled sources, as in DIITO, and optimized interventions are left to future work (Sec. 4.2), should Sec. 2 (l.111-114) still list them as an advance over DIITO?

  5. No alignment is consistently best across tasks and metrics. Why do you think this is, and how should a practitioner choose among FULL, MIDDLE and LATE?

  6. Why are the SST-2 "Bridge only" and "CAB output-only" rows of Tab. 27 identical in every column?

  7. Why wasn't DIITO run under the MIDDLE and LATE alignments?

5 comments
AuthorsOpenReview ↗

Response 1/2

We thank the reviewer for the detailed feedback on novelty, objective ablations, abstraction maps, and experimental design.

Q1: Is the main contribution the CAD penalty or the CAB variant, which also includes DIITO’s bridge term? If the former, please add an ablation separating the CAD loss from the bridge term, ideally on tasks where gains are claimed. Relatedly, why is the practical CAB variant needed?

The main methodological contribution is CAD, which enforces consistency between teacher and student intervention responses across aligned hidden boundaries and the output. CAB is our practical training variant, combining CAD with DIITO's bridge term.

To isolate their contributions, we added ablations comparing KD-only, CAD-only, bridge-only, and full CAB on QNLI and MNLI (Appendix D.5, Table 30). We additionally evaluated KD-only, CAD-only, and full CAB in Qwen2.5 14B→3B distillation on Yelp (Table 31).

On QNLI and both MNLI validation sets, CAD-only reduces hidden-response error by approximately 40–42% relative to KD-only, but yields limited improvements in output-level metrics. Conversely, bridge-only primarily improves output-level consistency and accuracy. Full CAB combines these complementary effects, achieving output performance comparable to bridge-only while substantially reducing hidden-response error.

The Qwen results further support this combination: full CAB achieves the highest accuracy (64.20%), teacher agreement (85.53%), and lowest CAD_IO (0.1158) among the evaluated ablations.

These results explain the practical motivation for CAB: CAD provides internal intervention consistency, while the bridge term helps translate this supervision into competitive output behavior.

Q2: What is the family of the abstraction maps and why is it restrictive enough to avoid vacuous abstractions (Sutter et al., 2025)? Are the maps fitted once and frozen (l.288-292) or refit alternately during training (l.183-184)? Also, which maps produce CAB’s reported CAD?

We use layer-specific affine maps, τj(h)=Wjh+bj\tau_j(h) = W_j h + b_j, applied independently to token representations. Their restricted form limits expressivity compared with unrestricted nonlinear mappings. Also, the problem of non-identifiability mentioned by Sutter et al. and others matters much less for our work since we are learning (constructing) the abstraction: we only need to find one good abstraction (with low causal abstraction error) within the model class spanned by the student architecture. Nevertheless, the restriction to affine maps follows the assumptions required by Sutter et al.

  • During training, maps are fitted once to the initial shared student checkpoint by minimizing clean teacher–student representation MSE, then frozen throughout student optimization. We corrected the earlier wording suggesting alternating optimization.
  • For evaluation, the reported CAD values use independently refitted affine maps for each final student checkpoint, including the baselines. We have clarified this distinction and the limitations of comparing map-dependent CAD values across methods (Section 4.1; Appendix D).

Q3: Provide a formal statement or derivation of the unified view: for each recovered method, the intervention, evaluation set, divergence, map and (possibly) conditions under which equality holds. Otherwise, soften the claim to "related to".

Thank you for this suggestion. We added formal derivations in Appendix N, specifying the interventions, evaluation sets, divergences and assumptions under which existing distillation objectives are recovered as special cases of CAD.

These derivations clarify the scope of the unifying formulation and distinguish exact reductions under stated conditions from broader conceptual relationships (Section 4; Appendix N).

Q4: Since all experiments use interchange interventions with randomly sampled sources, as in DIITO, and optimized interventions are left to future work (Sec. 4.2), should Sec. 2 (l.111-114) still list them as an advance over DIITO?

We agree that support for broader interventions should not be presented as an experimentally demonstrated advantage over DIITO.

We revised Section 2 to distinguish the generality of the framework from the improvements evaluated in this paper. Although CAD permits interventions beyond donor-activation swaps, all reported experiments use randomly sampled interchange interventions, as in DIITO for fair comparison. We clarified this in Section 4.2 and leave evaluation of broader intervention families to future work.

(continued in the next comment)

AuthorsOpenReview ↗

Response 2/2

(continued from the previous comment)

Q5: No alignment is consistently best across tasks and metrics. Why do you think this is, and how should a practitioner choose among FULL, MIDDLE and LATE?

Each alignment specifies which teacher and student representations are supervised for intervention consistency. Different tasks may depend on different stages of the teacher's computation, so no single alignment necessarily performs best across all tasks and metrics.

We recommend FULL as a default because it supervises the entire aligned network without requiring prior knowledge of which layers are most relevant. When validation resources permit, practitioners can also evaluate MIDDLE and LATE and select according to their primary objective, such as task accuracy or CAD_IO.

The three alignments have similar training costs, so computational overhead is not a major differentiating factor (Appendix D.4, Table 29). Our results demonstrate the importance of alignment selection, while automatically identifying suitable alignments remains future work.

Q6: Why are the SST-2 "Bridge only" and "CAB output-only" rows of Tab. 27 identical in every column?

Thank you for identifying this ambiguity. The rows appear identical because the reported values were rounded. We added higher-precision results in Appendix K, Table 40, showing small differences in the diagnostic metrics. The four diagnostic means differ by only 0.00000537–0.00001319 in absolute terms.

The similarity is consistent with both objectives supervising output-level behavior under interventions. Under our experimental setup, adding the output CAD term to the bridge loss has little measurable effect on SST-2. We now clarify this observation in Appendix K.

Q7: Why wasn't DIITO run under the MIDDLE and LATE alignments?

We added direct comparisons of DIITO and CAB under FULL, MIDDLE, and LATE alignments on SST-2 and QNLI (Appendix H, Tables 35–36).

On QNLI, CAB achieves higher accuracy and lower CAD_IO under all three alignments, reducing CAD_IO by 21.7%, 14.9%, and 12.9%, respectively.

On SST-2, the results are mixed: DIITO achieves slightly higher accuracy under FULL and MIDDLE, while CAB performs better across the reported metrics under LATE. CAB also achieves lower hidden-response error and map-dependent CAD across the evaluated alignments, although these improvements are partly expected from its training objective.

These results show that CAB's advantages on QNLI are not solely attributable to alignment selection, while confirming that improvements remain task-dependent.

We thank you again for your time. Please let us know if we have answered your concerns, or if you need further clarifications from our side.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Reponse

We thank the reviewers for their constructive feedback. Their comments helped us strengthen the experimental evidence, and better characterize the benefits and limitations of CAB. We have provided an updated version of the paper addressing reviewer comments, new experiments and improvements. The changes are marked in blue in the revised version. For the sake of readability, we do not include the full results in the rebuttal but we provide references to the paper.

The main concerns centered on novelty, empirical improvements, and computational cost. We addressed these through several additions:

  • Objective ablations (vGZQ, qJyM): We isolated CAD-only, bridge-only, and full CAB on MNLI and QNLI, with additional Qwen2.5 14B→3B experiments (Appendix D.5). CAD-only reduces hidden-response error by approximately 40–42% relative to KD-only but yields limited output-level gains. The bridge term primarily improves output-level performance, while full CAB combines both benefits, substantially improving internal consistency while maintaining competitive task accuracy.

  • Stronger comparisons (vGZQ, qJyM, LCXS): We added DIITO comparisons across alignment schemes, a hidden-state matching variant, seed-paired confidence intervals, and intervention-response diagnostics (Appendix D; Appendices H-J.). CAB outperforms the DIITO variant on QNLI, while the confidence intervals and diagnostics support improvements in intervention-response consistency on MNLI and QNLI. These results suggest that CAB's benefits extend beyond alignment selection and conventional hidden-state matching.

  • Practical benefits and costs (LCXS, Qd9A): We expanded cross-dataset and adversarial evaluations and measured training overhead. CAB improves intervention-response consistency in several settings, although gains in task-level metrics remain task-dependent. Compared with DIITO, CAB-FULL requires 22–24% more time per training update and 15–20% more peak GPU memory at matched batch sizes (Appendices D and G).

  • Methodological clarifications (vGZQ, Qd9A): We clarified abstraction-map fitting, the scope of supported versus evaluated interventions, the relationship to DIITO, and the distinction between our demonstrated results and future mechanistic-interpretability applications (Sections 1–4; Appendix N).

We address specific reviewer comments in their respective threads.

Overall, the revisions provide stronger evidence for CAD as a distinct source of internal intervention-consistency supervision and for CAB as a practical combination with output-level distillation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.