- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper formulates activation steering as matching distributions of activation trajectories, motivated by the fact that layerwise marginals do not determine cross-layer dependencies. It constructs a weighted MMD objective over contiguous layer groups, including the full trajectory, and establishes conditional identification, weak-convergence, and functional-agreement results. It also characterizes conditions under which difference-in-means is the optimal additive shift. The implementation fits steering maps to cached activations without repeated backbone execution during fitting. Experiments evaluate rotational, projection, and affine maps on jailbreak and truthfulness tasks, with capability-retention measurements and matching-scope ablations.
Critical strengths
-
A useful theoretical framework with carefully stated boundaries. The analysis distinguishes path-law identification, functional agreement under a shared input, and optimality within additive maps. Theorems 4.3, 4.5, and 4.7 make these distinctions explicit. Appendix C.5 analyzes why transformed cached trajectories are insufficient for the functional guarantee and gives a conditional discrepancy bound. This helps explain what distribution matching can establish without presenting the LLM experiments as verification of all theoretical assumptions.
-
A reusable fitting procedure with meaningful practical results. The cached objective supports several intervention families while avoiding repeated backbone passes during map fitting. Table 1 includes substantial improvements, such as the HarmBench ASR on Qwen3.5-27B increasing from 8.65% with DirAbl to 73.08% with Kolmogorov DirAbl. In Table 2, the Kolmogorov variants achieve higher primary T × I scores than their corresponding affine baselines on all three models. Together with the capability-retention measurements, these results support the practical value of the framework, even though the contribution of cross-layer matching requires more precise attribution.
Critical weaknesses
-
The contribution of cross-layer matching is only partially isolated. Table 1 compares fitted maps with heuristic constructions. In Table 2, MiMiC and Kolmogorov Low Rank use different parameterizations, and no same-parameterization marginal-MMD control is reported. Table A4 provides a relevant scope ablation, but its Gaussian all-group gains over marginal matching are 4.81 and 1.92 ASR points, corresponding to differences of five and two in the aggregate numbers of successful prompts out of 104, without paired uncertainty or variability across fitted runs. Figures A4–A6 on p.30 report cached training losses, rather than held-out marginal/joint discrepancies on actual intervened trajectories. The authors explicitly acknowledge the cached–actual gap and missing shared-input correspondence; these are limits on the empirical interpretation, not errors in the conditional theorems.
-
The matching-scope ablation protocol is missing from its stated location. The reproducibility statement on p.10 and discussion on p.37 refer to Appendix G for the exploratory selection protocol. However, pp.34–35 contain only a short introduction and Table A4. It remains unclear which map, activation sites, optimization settings, and hyperparameters were fixed across scopes, which were reselected, and on which data. Appendix D describes the different models used in Table 1. The missing configuration and selection details limit reproducibility and interpretation of A4; they do not establish test-set tuning.
Questions for authors
-
Could you provide the Table A4 selection protocol, or identify the exact scripts and configurations in the accompanying repository? Please specify the map family, activation/intervention sites, optimization budget, fixed versus reselected hyperparameters, and selection data. Were all matching scopes compared under the same validation rule?
-
Can existing paired per-prompt outcomes or appropriate uncertainty estimates clarify the A4 gains? If repeated fits are available, does the marginal/full/all-group ordering persist across seeds? These results would help determine how much weight to place on the differences in aggregate success counts.
-
Is there a truthfulness comparison fitting the same diagonal or low-rank maps with marginal MMD under the same budget? This would directly clarify objective attribution. As a complementary diagnostic, existing held-out marginal/joint discrepancies on actual intervened trajectories would help interpret the mechanism, while not by themselves validating the functional theorem’s assumptions.
7 comments
Rebuttal by Authors (2/4)
Q1 (Part 2). Figures A4–A6 on p.30 report cached training losses, rather than held-out marginal/joint discrepancies on actual intervened trajectories. The authors explicitly acknowledge the cached–actual gap and missing shared-input correspondence; these are limits on the empirical interpretation, not errors in the conditional theorems.
Answer. First, regarding the cached activation concern. Please allow us to elaborate on this both theoretically and empirically.
Theoretically, in Proposition C.1, we have established that the square root of the intervened population loss can be upper bounded by the square root of the empirical cached loss and other residual terms.
Empirically, we conduct an additional experiment on the jailbreak task using Qwen3.5-27B, GPT-OSS-20B, and Gemma-4-31B-IT to demonstrate that the cached loss can serve as an effective surrogate for the actual intervened loss. We fit steering maps on cached activations and evaluate the loss on held-out test sets using both transformed cached trajectories and actual intervened trajectories obtained by applying the maps layer by layer. Table 3 from the reply shows that, from initialization, the actual intervened loss surprisingly decreases by an even larger percentage than the cached loss in most settings, suggesting that reducing the cached loss has strong effect on reducing the actual intervened loss. Figures A7, A8, and A9 in Appendix H further demonstrate that both losses share the same trend during training.
Table 3: Percentage decrease in Kolmogorov loss for cached and actual intervened trajectories evaluated on the test set after fitting steering maps on cached activations from train set.
| Method | Qwen3.5-27B Cached | Qwen3.5-27B Intervened | GPT-OSS-20B Cached | GPT-OSS-20B Intervened | Gemma-4-31B-IT Cached | Gemma-4-31B-IT Intervened |
|---|---|---|---|---|---|---|
| Angular steering | 29.41% | 58.79% | 27.51% | 61.49% | 25.14% | 19.69% |
| Directional ablation | 39.85% | 48.84% | 38.82% | 56.81% | 24.82% | 39.26% |
Second, we would like to clarify the shared input assumption.
Theorem 4.3 does not assume shared inputs and serves as the theoretical foundation for our Kolmogorov loss. The shared input assumption is needed only in Theorem 4.5, which concerns behavior alignment. Without shared inputs, it becomes nontrivial to determine whether the steered and reference networks have the same behavior, since behavioral equivalence between their mappings must first be rigorously formalized. The theorem therefore provides a guarantee in the tractable shared input setting.
Q2. The matching-scope ablation protocol is missing from its stated location. The reproducibility statement on p.10 and discussion on p.37 refer to Appendix G for the exploratory selection protocol. However, pp.34–35 contain only a short introduction and Table A4. It remains unclear which map, activation sites, optimization settings, and hyperparameters were fixed across scopes, which were reselected, and on which data. Appendix D describes the different models used in Table 1. The missing configuration and selection details limit reproducibility and interpretation of A4; they do not establish test-set tuning.
Answer. We have detailed the experimental setup of our ablation in Section G.1 in the updated manuscript. For convenience, we repeat them here. The experiment uses the same dataset and source code described in Appendix D. We strictly follow the settings in Appendix D for initialization, the activation intervention site, the optimizer, and the evaluation protocol. We then adapt the experimental setup to Qwen3.5-9B and Llama-3.1-8B-Instruct, use the angular adaptive map, and learn the steering plane. Each reported result is based on five seeds, and the RQ kernel uses . For each seed, we vary the inference angle while keeping the fitted steering plane fixed, and select the angle that achieves the best result on the validation set. We then use the selected angles to evaluate performance on the test set. Similarly, for the AAS baseline, we follow the same protocol, except that we use the authors’s heuristic parameter selection.
Q3. Could you provide the Table A4 selection protocol, or identify the exact scripts and configurations in the accompanying repository? Please specify the map family, activation/intervention sites, optimization budget, fixed versus reselected hyperparameters, and selection data. Were all matching scopes compared under the same validation rule?
Answer. We refer the answer of this question to Q2's answer.
Q4. Can existing paired per-prompt outcomes or appropriate uncertainty estimates clarify the A4 gains? If repeated fits are available, does the marginal/full/all-group ordering persist across seeds? These results would help determine how much weight to place on the differences in aggregate success counts.
Answer. We refer the answer of this question to Q1 Part1's answer.
Rebuttal by Authors (3/4)
Q5. Is there a truthfulness comparison fitting the same diagonal or low-rank maps with marginal MMD under the same budget? This would directly clarify objective attribution. As a complementary diagnostic, existing held-out marginal/joint discrepancies on actual intervened trajectories would help interpret the mechanism, while not by themselves validating the functional theorem’s assumptions.
Answer. We thank the reviewer for the suggestion. We have included the truthfulness ablation for Kolmogorov low-rank with rank with three variants: layerwise marginals, full trajectory, and all contiguous layer groups. Across all runs, we preserve the same data splits, intervention layers, optimization settings, and evaluation protocol as the main truthfulness experiment (Appendix E). We report Truth, Info, and , the percentage of responses judged both truthful and informative.
Table 4 shows that all-contiguous-group matching achieves the highest on all three models. Full-trajectory matching also outperforms marginal matching in all settings. These controlled comparisons provide direct evidence that the improvement comes from the matching objective, rather than a change in steering parameterization
Table 4. Truthfulness ablation of matching scope for Kolmogorov low-rank steering. Truth and Info measure truthful and informative responses, respectively. measures responses judged both. All scores are percentages, with higher values being better.
Regarding the question about interpreting the mechanism by analyzing discrepancies in the actual intervened loss, we evaluate whether the cached objective reflects actual intervened trajectories. We do so by testing whether the cached loss serves as a surrogate for the actual intervened loss on the jailbreak task across Qwen3.5-27B, GPT-OSS-20B, and Gemma-4-31B-IT. To this end, we fit steering maps on cached activations from train set using the all contiguous groups variant of the Kolmogorov loss with a Gaussian kernel, then evaluate Kolmogorov loss of both the transformed cached trajectories and the actual intervened trajectories obtained by applying the maps layer by layer on test sets.
Table 5 shows that, from initialization, the actual intervened loss surprisingly decreases by an even larger percentage than the cached loss in most settings, suggesting that reducing the cached loss has strong effect on reducing the actual intervened loss. Figures A7, A8, and A9 of Appendix H in the updated manuscript further demonstrate that both losses share the same trend during training.
Table 5: Percentage decrease in Kolmogorov loss for cached and actual intervened trajectories evaluated on the test set after fitting steering maps on cached activations from train set.
| Method | Qwen3.5-27B Cached | Qwen3.5-27B Intervened | GPT-OSS-20B Cached | GPT-OSS-20B Intervened | Gemma-4-31B-IT Cached | Gemma-4-31B-IT Intervened |
|---|---|---|---|---|---|---|
| Angular steering | 29.41% | 58.79% | 27.51% | 61.49% | 25.14% | 19.69% |
| Directional ablation | 39.85% | 48.84% | 38.82% | 56.81% | 24.82% | 39.26% |
Rebuttal by Authors (1/4)
Thank you for your thoughtful review and constructive feedback. We have revised our manuscript in response to your suggestions, with all changes highlighted in blue. Below, we address each of your concerns in detail.
Q1 (Part 1). The contribution of cross-layer matching is only partially isolated. Table 1 compares fitted maps with heuristic constructions. In Table 2, MiMiC and Kolmogorov Low Rank use different parameterizations, and no same-parameterization marginal-MMD control is reported. Table A4 provides a relevant scope ablation, but its Gaussian all-group gains over marginal matching are 4.81 and 1.92 ASR points, corresponding to differences of five and two in the aggregate numbers of successful prompts out of 104, without paired uncertainty or variability across fitted runs.
Answer. We address the Reviewer's concern by conducting 2 more experiments. First, we run and report the Jailbreak ablation experiment with 5 random seeds and report mean along with standard deviation. The result in Table 1 demonstrates that the cross layer variants consistently outperform in mean over marginal MMD when running multiple seeds.
Table 1. Mean ± standard deviation of AdvBench ASR (%) and TinyBench scores for the jailbreak experiment.
Second, we record and compare the learned angular steering planes for all (25)=10 pairs of seeds within each ablation setting to show learnt steering plane are consistent across random seeds. For each pair, we compute the cosine of the largest principal angle between these planes. Table 2 reports the minimum similarity over all 10 seed pairs for each setting, with an minimum of 0.999917 indicating that the learned steering planes are nearly identical across seeds.
Table 2. Cross-seed stability of the learned angular-steering planes across five fitting seeds. Each entry is the minimum over the ten seed pairs of cos(ϕmax), where ϕmax is the largest principal angle between the learned two dimensional subspaces. Values closer to one indicate more similar planes.
For the question about marginal MMD control experiment with the same parameterization in the truthfulness task, we refer the reviewer to our response to Q5, which includes a new experiment for this setting.