acceptodds
Under review as a conference paper at ICLR 2027

Activation Steering by Path Law Matching

Abstract

We develop a theoretical framework for activation steering based on matching activation path laws. Distinct path laws can share identical activation marginals at every layer, so marginal alignment alone cannot identify the intended path law. Using Kolmogorov's extension theorem, we characterize this law through consistent finite dimensional joint distributions. For finite depth networks, we construct the Kolmogorov loss to match joint activation distributions over contiguous groups of layers. Under the stated kernel conditions, vanishing population loss is equivalent to weak convergence toward the reference path law. With shared inputs and additional regularity assumptions, sufficiently accurate matching of actual intervened trajectories improves output agreement over exact marginal matching that retains positive functional error. We also establish conditions under which the difference-in-means vector is the unique optimal additive map. For practical fitting, we learn steering maps from cached activations without additional forward or backward passes through the pretrained model. Experiments on jailbreak and truthfulness across language models with 7B to 31B parameters demonstrate the framework's utility across several steering map families.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
activation steering, mechanistic interpretability, stochastic process, analysis

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_g7mrOpenReview ↗
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
  1. 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.

  2. 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
  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. 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.

  2. 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
  1. 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?

  2. 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.

  3. 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
AuthorsOpenReview ↗

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 22 more experiments. First, we run and report the Jailbreak ablation experiment with 55 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 ±\pm standard deviation of AdvBench ASR (%) and TinyBench scores for the jailbreak experiment.

Matching scope\text{Matching scope}Kernel\text{Kernel}wa,b∝w_{a,b}\proptoQwen3.5-9B ASR↑\text{Qwen3.5-9B ASR}\uparrowQwen3.5-9B TinyBench Avg.↑\text{Qwen3.5-9B TinyBench Avg.}\uparrowLlama-3.1-8B-IT ASR↑\text{Llama-3.1-8B-IT ASR}\uparrowLlama-3.1-8B-IT TinyBench Avg.↑\text{Llama-3.1-8B-IT TinyBench Avg.}\uparrow
Baseline (no steer)\text{Baseline (no steer)}—\text{---}—\text{---}0.960.9667.9667.964.814.8167.5767.57
AAS\text{AAS}—\text{---}—\text{---}72.1272.1263.1063.1089.4289.4266.6766.67
Marginal matching\text{Marginal matching}Gaussian\text{Gaussian}1a=b\mathbb{1}_{\\{a=b\\}}90.58±0.7290.58 \pm 0.7266.19±0.1466.19 \pm 0.1494.04±0.4394.04 \pm 0.4367.55±0.0767.55 \pm 0.07
Full trajectory\text{Full trajectory}Gaussian\text{Gaussian}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}93.27±1.2293.27 \pm 1.2266.70±0.03‾\underline{66.70 \pm 0.03}95.19±0.6895.19 \pm 0.6867.75±0.07\mathbf{67.75 \pm 0.07}
All contiguous groups\text{All contiguous groups}Gaussian\text{Gaussian}1195.19±0.61\mathbf{95.19 \pm 0.61}66.04±0.1066.04 \pm 0.1096.35±0.80‾\underline{96.35 \pm 0.80}67.55±0.0767.55 \pm 0.07
All contiguous groups\text{All contiguous groups}RQ\text{RQ}1192.88±0.7792.88 \pm 0.7765.28±0.0765.28 \pm 0.0796.54±1.10\mathbf{96.54 \pm 1.10}67.59±0.02‾\underline{67.59 \pm 0.02}
All contiguous groups\text{All contiguous groups}Gaussian\text{Gaussian}1b−a+1\frac{1}{b-a+1}93.46±0.38‾\underline{93.46 \pm 0.38}66.78±0.01\mathbf{66.78 \pm 0.01}96.15±1.1896.15 \pm 1.1867.42±0.0967.42 \pm 0.09
All contiguous groups\text{All contiguous groups}RQ\text{RQ}1b−a+1\frac{1}{b-a+1}91.92±1.3091.92 \pm 1.3066.46±0.2266.46 \pm 0.2295.96±1.5895.96 \pm 1.5867.50±0.1867.50 \pm 0.18

Second, we record and compare the learned angular steering planes for all (52)=10\binom{5}{2}=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 1010 seed pairs for each setting, with an minimum of 0.9999170.999917 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⁡)\cos(\phi_{\max}), where ϕmax⁡\phi_{\max} is the largest principal angle between the learned two dimensional subspaces. Values closer to one indicate more similar planes.

Matching scope\text{Matching scope}Kernel\text{Kernel}wa,b∝w_{a,b}\proptoQwen3.5-9B Cosine similarity\text{Qwen3.5-9B Cosine similarity}Llama-3.1-8B-IT Cosine similarity\text{Llama-3.1-8B-IT Cosine similarity}
Marginal matching\text{Marginal matching}Gaussian\text{Gaussian}1a=b\mathbb{1}_{\\{a=b\\}}0.9999340.9999340.9999360.999936
Full trajectory\text{Full trajectory}Gaussian\text{Gaussian}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}0.9999180.9999180.9999450.999945
All contiguous groups\text{All contiguous groups}Gaussian\text{Gaussian}110.9999490.9999490.9999470.999947
All contiguous groups\text{All contiguous groups}RQ\text{RQ}110.9999540.9999540.9999480.999948
All contiguous groups\text{All contiguous groups}Gaussian\text{Gaussian}1b−a+1\frac{1}{b-a+1}0.9999170.9999170.9999380.999938
All contiguous groups\text{All contiguous groups}RQ\text{RQ}1b−a+1\frac{1}{b-a+1}0.9999320.9999320.9999410.999941

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.

AuthorsOpenReview ↗

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.

MethodQwen3.5-27B CachedQwen3.5-27B IntervenedGPT-OSS-20B CachedGPT-OSS-20B IntervenedGemma-4-31B-IT CachedGemma-4-31B-IT Intervened
Angular steering29.41%58.79%27.51%61.49%25.14%19.69%
Directional ablation39.85%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 α=1\alpha = 1. 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.

AuthorsOpenReview ↗

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 r=2r=2 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 T×IT\times I, the percentage of responses judged both truthful and informative.

Table 4 shows that all-contiguous-group matching achieves the highest T×IT\times I 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. T×IT \times I measures responses judged both. All scores are percentages, with higher values being better.

Model\text{Model}Matching scope\text{Matching scope}wa,b∝w_{a,b}\proptoTruth↑\text{Truth}\uparrowInfo↑\text{Info}\uparrowT×I↑T\times I\uparrow
Llama-3.1-8B\text{Llama-3.1-8B}Marginal\text{Marginal}1a=b\mathbb{1}_{\\{a=b\\}}61.161.192.492.454.154.1
Full trajectory\text{Full trajectory}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}65.765.791.491.457.4‾\underline{57.4}
Contiguous\text{Contiguous}1164.164.194.694.659.0\mathbf{59.0}
Qwen2.5-7B\text{Qwen2.5-7B}Marginal\text{Marginal}1a=b\mathbb{1}_{\\{a=b\\}}80.780.785.285.266.066.0
Full trajectory\text{Full trajectory}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}81.281.285.985.967.2‾\underline{67.2}
Contiguous\text{Contiguous}1183.283.285.885.869.4\mathbf{69.4}
Mistral-7B\text{Mistral-7B}Marginal\text{Marginal}1a=b\mathbb{1}_{\\{a=b\\}}67.667.690.590.558.458.4
Full trajectory\text{Full trajectory}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}67.967.991.991.959.9‾\underline{59.9}
Contiguous\text{Contiguous}1169.869.890.090.060.0\mathbf{60.0}

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.

MethodQwen3.5-27B CachedQwen3.5-27B IntervenedGPT-OSS-20B CachedGPT-OSS-20B IntervenedGemma-4-31B-IT CachedGemma-4-31B-IT Intervened
Angular steering29.41%58.79%27.51%61.49%25.14%19.69%
Directional ablation39.85%48.84%38.82%56.81%24.82%39.26%
Reviewer_mCtaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a theoretical framework for activation steering based on matching activation path laws rather than individual layer distributions. The authors introduce the Kolmogorov loss, which uses MMD to align joint activation distributions across contiguous groups of layers. The paper provides theoretical guarantees for distributional matching and proposes an efficient optimization procedure using cached activations. Experiments on jailbreak and truthfulness tasks show improvements over several existing steering methods.

Critical strengths

The motivation is clear and well supported: matching layerwise marginals does not necessarily preserve dependencies across layers.

The theoretical analysis is extensive and provides a principled justification for the proposed distributional objective.

The method is computationally attractive, as steering maps can be optimized using cached activations without repeated backbone passes.

The experimental evaluation covers multiple model families and steering parameterizations, showing promising improvements over the corresponding baselines.

Critical weaknesses

Gap between theory and practice: The theoretical guarantees concern actual intervened trajectories with shared inputs, whereas the practical method optimizes transformed cached activations from potentially different inputs. This limits the applicability of the main theoretical results to the proposed implementation.

Limited methodological novelty: Joint distribution matching using MMD has been previously explored, notably in Joint Adaptation Networks (Long et al., Deep Transfer Learning with Joint Adaptation Networks, ICML 2017), which aligns joint distributions of activations across multiple network layers using Joint MMD. More generally, MMD-based distribution matching is well established (Gretton et al., A Kernel Two-Sample Test, JMLR 2012). While applying this principle to activation steering is interesting, the methodological novelty of the proposed objective could be better clarified. In particular, Theorem 4.3 already guarantees path-law identification when the full-trajectory term has positive weight, independently of the additional contiguous-group terms. Therefore, the necessity of including all contiguous groups appears primarily empirical rather than theoretical, and would benefit from stronger experimental justification.

Insufficient isolation of contributions: The ablations comparing marginal and joint matching are relatively limited. Additionally, some baseline comparisons involve different steering parameterizations, making it difficult to attribute improvements specifically to the proposed loss.

Questions for authors

Could the authors provide a more systematic comparison between marginal matching, full-trajectory matching, and the proposed contiguous-group objective under identical steering parameterizations? In particular, it would be valuable to understand whether the additional groups consistently improve performance and how well the cached optimization objective reflects the distribution of actual intervened trajectories.

8 comments
AuthorsOpenReview ↗

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. The theoretical guarantees concern actual intervened trajectories with shared inputs, whereas the practical method optimizes transformed cached activations from potentially different inputs. This limits the applicability of the main theoretical results to the proposed implementation.

Answer. We thank the reviewer for the careful assessment and constructive suggestions. We address the concerns below.

First, 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. It guarantees that, loss between the intervened path law and the reference path law converges to 00 if and only if intervened path law converges weakly to the reference path law. This result does not require the shared input assumption.

The shared input assumption is needed only in Theorem 4.5, which concerns behavior alignment. In this theorem, having the similar behavior is mathematically defined as exhibiting similar input-output mapping. For a given input, the theorem guarantees that our objective can recover the reference network’s intended mapping more effectively than marginal matching when the steering family is sufficiently expressive. Without shared inputs, it becomes nontrivial to determine whether the steered and reference networks have the same behavior, since behavioral equivalence between mappings taking different inputs must first be rigorously formalized. The theorem therefore provides a guarantee in the tractable setting of shared input. Extending it to non shared inputs requires first formally defining behavioral equivalence when the inputs are different. This is an interesting and difficult problem on its own and is beyond the scope of this work.

Second, 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 plus 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 1 provided in 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 1: Percentage decrease in Kolmogorov loss for cached and actual intervened trajectories evaluated on test set after fitting steering maps on cached activations from train set.

MethodQwen3.5-27B CachedQwen3.5-27B IntervenedGPT-OSS-20B CachedGPT-OSS-20B IntervenedGemma-4-31B-IT CachedGemma-4-31B-IT Intervened
Angular steering29.41%58.79%27.51%61.49%25.14%19.69%
Directional ablation39.85%48.84%38.82%56.81%24.82%39.26%
AuthorsOpenReview ↗

Rebuttal by Authors (2/4)

Q2 part 1. Limited methodological novelty: Joint distribution matching using MMD has been previously explored, notably in Joint Adaptation Networks (Long et al., Deep Transfer Learning with Joint Adaptation Networks, ICML 2017), which aligns joint distributions of activations across multiple network layers using Joint MMD. More generally, MMD-based distribution matching is well established (Gretton et al., A Kernel Two-Sample Test, JMLR 2012). While applying this principle to activation steering is interesting, the methodological novelty of the proposed objective could be better clarified.

Answer. We would like to clarify that our main methodological contribution is providing a novel viewpoint of seeing activation steering as matching activation path laws. Conceptually, steering should account for how activations evolve across layers, and Kolmogorov Extension Theorem provides the foundation for characterizing these laws through joint distributions. This connection may seem intuitive in hindsight once the formulation has been worked out, but recognizing and developing it into a principled alignment method is by no means trivial. To the best of our knowledge, there is no prior activation steering work that has made this connection. Building on this novel viewpoint, our work develops a rigorous theoretical framework and an efficient empirical objective. We use MMD loss because it is well suited to theoretically analyze and has efficient implementation, but our framework is by no mean restricted to this choice. Other distribution matching methods, including Sinkhorn divergences and variational estimates of ff-divergences, can also be used, as discussed in Remark 3.3 of Section 3.2.

Q2 part 2. In particular, Theorem 4.3 already guarantees path law identification when the full trajectory term has positive weight, independently of the additional contiguous group terms. Therefore, the necessity of including all contiguous groups appears primarily empirical rather than theoretical, and would benefit from stronger experimental justification.

Answer. We agree with the reviewer that Theorem 4.3 establishes that full trajectory matching is sufficient for theoretical identification, but this result assumes an ideal data setting. As discussed in Section 3.2, limited steering data can leave distinct joint distributions in the high dimensional trajectory space statistically indistinguishable under MMD loss. Including shorter contiguous groups can provide useful learning signals at little additional computational cost, in which we have also provided an efficient cumulative implementation in Appendix B, where all kernel evaluations used for the full trajectory loss are reused for other contiguous groups. Empirically, matching all contiguous groups outperforms full trajectory matching in most settings across the Jailbreak and Truthfulness ablations, as shown in Table 5 in the manuscript and Table A5 in Appendix G.2.

AuthorsOpenReview ↗

Rebuttal by Authors (3/4)

Q3. Insufficient isolation of contributions: The ablations comparing marginal and joint matching are relatively limited. Additionally, some baseline comparisons involve different steering parameterizations, making it difficult to attribute improvements specifically to the proposed loss.

Answer. We appreciate the reviewer’s suggestion. We address the concern with two additional experiments, whose detailed setups are provided in Section G of the updated manuscript.

First, we run and report the Jailbreak ablation experiment with 55 random seeds and report the mean and standard deviation. The result in Table 2 demonstrates that the cross-layer variants consistently outperform marginal MMD in mean attack successful rate (ASR) across multiple seeds.

Table 2. Mean ±\pm standard deviation of AdvBench ASR (%) and TinyBench scores for the jailbreak experiment.

Matching scope\text{Matching scope}Kernel\text{Kernel}wa,b∝w_{a,b}\proptoQwen3.5-9B ASR↑\text{Qwen3.5-9B ASR}\uparrowQwen3.5-9B TinyBench Avg.↑\text{Qwen3.5-9B TinyBench Avg.}\uparrowLlama-3.1-8B-IT ASR↑\text{Llama-3.1-8B-IT ASR}\uparrowLlama-3.1-8B-IT TinyBench Avg.↑\text{Llama-3.1-8B-IT TinyBench Avg.}\uparrow
Baseline (no steer)\text{Baseline (no steer)}—\text{---}—\text{---}0.960.9667.9667.964.814.8167.5767.57
AAS\text{AAS}—\text{---}—\text{---}72.1272.1263.1063.1089.4289.4266.6766.67
Marginal matching\text{Marginal matching}Gaussian\text{Gaussian}1a=b\mathbb{1}_{\\{a=b\\}}90.58±0.7290.58 \pm 0.7266.19±0.1466.19 \pm 0.1494.04±0.4394.04 \pm 0.4367.55±0.0767.55 \pm 0.07
Full trajectory\text{Full trajectory}Gaussian\text{Gaussian}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}93.27±1.2293.27 \pm 1.2266.70±0.03‾\underline{66.70 \pm 0.03}95.19±0.6895.19 \pm 0.6867.75±0.07\mathbf{67.75 \pm 0.07}
All contiguous groups\text{All contiguous groups}Gaussian\text{Gaussian}1195.19±0.61\mathbf{95.19 \pm 0.61}66.04±0.1066.04 \pm 0.1096.35±0.80‾\underline{96.35 \pm 0.80}67.55±0.0767.55 \pm 0.07
All contiguous groups\text{All contiguous groups}RQ\text{RQ}1192.88±0.7792.88 \pm 0.7765.28±0.0765.28 \pm 0.0796.54±1.10\mathbf{96.54 \pm 1.10}67.59±0.02‾\underline{67.59 \pm 0.02}
All contiguous groups\text{All contiguous groups}Gaussian\text{Gaussian}1b−a+1\frac{1}{b-a+1}93.46±0.38‾\underline{93.46 \pm 0.38}66.78±0.01\mathbf{66.78 \pm 0.01}96.15±1.1896.15 \pm 1.1867.42±0.0967.42 \pm 0.09
All contiguous groups\text{All contiguous groups}RQ\text{RQ}1b−a+1\frac{1}{b-a+1}91.92±1.3091.92 \pm 1.3066.46±0.2266.46 \pm 0.2295.96±1.5895.96 \pm 1.5867.50±0.1867.50 \pm 0.18

Second, we have added a truthfulness ablation of the same Kolmogorov low-rank map parameterization with Gaussian kernel, comparing three matching scopes: individual-layer marginals, all contiguous layer groups, and the full trajectory.

The result in Table 3 shows that contiguous matching achieves the highest T×IT \times I score on all three models. Full-length matching also improves over marginal matching, but remains slightly below contiguous matching in most settings. Consistent with the jailbreak ablation, these results support matching across layer groups rather than relying on individual-layer marginals alone.

Table 3. Truthfulness ablation of matching scope for Kolmogorov low-rank steering. Truth and Info measure truthful and informative responses, respectively. T×IT \times I measures responses judged both. All scores are percentages, with higher values being better.

Model\text{Model}Matching scope\text{Matching scope}wa,b∝w_{a,b}\proptoTruth↑\text{Truth}\uparrowInfo↑\text{Info}\uparrowT×I↑T\times I\uparrow
Llama-3.1-8B\text{Llama-3.1-8B}Marginal\text{Marginal}1a=b\mathbb{1}_{\\{a=b\\}}61.161.192.492.454.154.1
Full trajectory\text{Full trajectory}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}65.765.791.491.457.4‾\underline{57.4}
Contiguous\text{Contiguous}1164.164.194.694.659.0\mathbf{59.0}
Qwen2.5-7B\text{Qwen2.5-7B}Marginal\text{Marginal}1a=b\mathbb{1}_{\\{a=b\\}}80.780.785.285.266.066.0
Full trajectory\text{Full trajectory}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}81.281.285.985.967.2‾\underline{67.2}
Contiguous\text{Contiguous}1183.283.285.885.869.4\mathbf{69.4}
Mistral-7B\text{Mistral-7B}Marginal\text{Marginal}1a=b\mathbb{1}_{\\{a=b\\}}67.667.690.590.558.458.4
Full trajectory\text{Full trajectory}1a=0, b=L\mathbb{1}_{\\{a=0,\,b=L\\}}67.967.991.991.959.9‾\underline{59.9}
Contiguous\text{Contiguous}1169.869.890.090.060.0\mathbf{60.0}
Reviewer_EMkcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Typical steering concerns changing a model's outputs to some other, more desirable output. However, focusing only on the input and output of the model leaves the intermediate processing of the model possibly still enacting undesirable behavior internally. Furthermore, even matching the marginals at each layer does not guarantee that the desired path law is being performed. This work formulates steering through path-law matching, rather than only matching the output distribution. This is done by considering continuous subsets of layers in tandem. By using the activations of the model when given both the prompt and the desired answer, we are given a collection of paths we desire the model to be steered towards. Using the difference in means in kernel space, averaged across each of these continuous strings of layers, we obtain vectors that steer the entire path law across layers to the desired behavior.

Critical strengths

This method is based on standard probability-theoretic concepts, namely Kolmogorov's Extension Theorem. Furthermore, toy situations describe well the failures of only matching the marginals rather than the path laws.

This method also works to fix a very real issue in steering: many steering methods result in a model that is more aligned, yet they fail to be safe from future attacks that extract undesirable behaviors from the model. Experiments show that in at least some situations, this method results in steered models that are consistently less harmful than other tested methods.

Critical weaknesses

The empirical comparison points do not include standard methods such as LoRA or full parameter fine-tuning. These would be helpful baselines to compare against for each of these settings. Other representation alignment methods, such as ReFT, are not tested either.

Questions for authors
  1. Is there a reason why the few steering methods selected for Table 1 and Table 2 were chosen? Or are there experiments on other methods that were missed when reading?
  2. The appendix appears to suggest that uniform weighting of each of these strings of layers, regardless of length, leads to the best performance. Is there some intuition for this?
  3. Does using the aligned responses as the path-law target introduce any issues with aligning early layers and late layers in the model? Early layers typically process grammar and other structure, which seems like it would not be desirable to align these layers if the output follows a different, yet still aligned, structure.

Minor Notes: Line 100: Needs a period.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

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.


W1. The empirical comparison points do not include standard methods such as LoRA or full parameter fine-tuning. These would be helpful baselines to compare against for each of these settings. Other representation alignment methods, such as ReFT, are not tested either.

Answer. In this work, we focus on efficiently finding activation steering map parameters entirely offline, with no further computation through the base LLM after activation extraction. Full fine-tuning requires backbone forward and backward passes and requires large memory for model gradients and optimizer states. Efficient finetuning method such as LoRA and ReFT require repeated forward computation, gradient propagation through the relevant backbone layers, and saving intermediate activations for backpropagation. Our offline fitting eliminates these computational and memory requirements, completing in less than 8080 seconds and using less than 55 GB of memory on models up to 3131B parameters, as shown in Table 1.

Table 1. Kolmogorov map fitting time and peak allocated GPU memory during optimization.

ModelMethodTime (s)Peak GPU memory (GiB)
Qwen3.5-27BKolmogorov AAS77.324.74
Qwen3.5-27BKolmogorov DirAbl69.324.66
Qwen3.5-27BKolmogorov CAA53.452.33
GPT-OSS-20BKolmogorov AAS53.821.00
GPT-OSS-20BKolmogorov DirAbl44.420.98
GPT-OSS-20BKolmogorov CAA42.150.54
Gemma-4-31B-ITKolmogorov AAS79.514.58
Gemma-4-31B-ITKolmogorov DirAbl45.102.97
Gemma-4-31B-ITKolmogorov CAA53.202.30

Q1. Is there a reason why the few steering methods selected for Table 1 and Table 2 were chosen? Or are there experiments on other methods that were missed when reading?

Answer. Our reason for selecting these methods is to test whether the Kolmogorov loss can improve established steering maps by replacing their original parameter constructions with parameters learned through activation path law matching. These methods are chosen to test whether the Kolmogorov loss provides a common fitting objective across representative steering map families, including translation (CAA), rotation (AAS), projection (DirAbl), and affine transformations (Linear-AcT and MiMiC). Since each method’s design and parameter selection are task specific, we evaluate them on their corresponding tasks. Together, these comparisons assess the effectiveness of our objective across several intervention map families and tasks.

Q2. The appendix appears to suggest that uniform weighting of each of these strings of layers, regardless of length, leads to the best performance. Is there some intuition for this?

Answer. We hypothesize that uniform weighting works well because it retains the contribution of longer layer groups, which capture dependencies that shorter groups may miss. Meanwhile, the weighting allows shorter groups contribute more overall because there are more shorter groups. For LL intervened layers, there are L−l+1L-l+1 contiguous groups of length ll. Thus, uniform weighting naturally emphasizes shorter groups while preserving longer groups contribution to path law matching.

Q3. Does using the aligned responses as the path-law target introduce any issues with aligning early layers and late layers in the model? Early layers typically process grammar and other structure, which seems like it would not be desirable to align these layers if the output follows a different, yet still aligned, structure.

Answer. We agree with the reviewer that early layer alignment may be less desirable for some tasks. Especially when target behavior emerges in later layers, earlier layer representations may already be similar, with differences only reflecting grammatical structure. Our formulation can account for this by assigning smaller weights to early layer groups and larger weights to later layer groups when necessary. Therefore, we hypothesize tuning these weights or developing a heuristic to select them automatically for each task could even further improve performance over uniform weighting. To this end, investigating how to choose these weights based on the task and the role of different layers is an interesting research direction and we leave this for future work.

Minor note. Line 100: Needs a period.

Answer. We thanks the reviewer for the careful review, we have updated the manuscript accordingly.


We hope we have cleared your concerns about our work. We have also revised our manuscript according to your comments, and we would appreciate it if we can get your further feedback at your earliest convenience.

I sincerely thank the authors for their attention and care in their rebuttal. Of the 4 papers in my batch that provided rebuttals, this is the only one that is not AI-generated.

My concerns have been addressed. The revisions in the draft have added useful ablations and details. Reading through the other reviewers' comments, I agree with the authors: while many of the techniques have been established before, such as MMD, their application to steering is novel and will be interesting to the steering community. As a result, I maintain my positive assessment and recommend acceptance.

AuthorsOpenReview ↗

Thank you for your endorsement

We sincerely thank the reviewer for the positive evaluation and thoughtful assessment of our work. We greatly appreciate the recognition of our contributions and the acknowledgment that the earlier concerns have been addressed.

Thank you again for your endorsement and careful reading of our paper. We remain open to any further questions from the reviewer in the rest of the discussion phase of the rebuttal.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revision

Incorporating the reviewers’ comments and suggestions, we summarize below the main changes to the revised manuscript. The revised text, tables, and figure captions are highlighted in blue.

  1. We expand the jailbreak ablation study to five random seeds. Table 5 now reports mean and standard deviation for the comparisons across matching scopes, kernels, and interval weights. With a Gaussian kernel and uniform weighting, contiguous-group matching improves mean ASR over marginal matching by 4.61 and 2.31 percentage points on Qwen3.5-9B and Llama-3.1-8B-Instruct, respectively. We additionally assess the stability of the learned steering planes in Appendix G, finding nearly identical planes across seeds, with a minimum similarity of 0.999917.

  2. We add a controlled truthfulness ablation. Using the same rank-2 low-rank map, data splits, intervention layers, and optimization settings, we compare marginal, full-trajectory, and contiguous-group matching. Table A5 in Appendix G.2 shows that contiguous-group matching achieves the highest T×IT\times I across all three models, improving over marginal matching.

  3. We add an empirical comparison of cached and actual intervened losses. Section 5.5 evaluates both losses on held-out data across three models and two steering methods. Table 6 shows that both losses decrease in all six settings after fitting exclusively on cached activations. Figures A7–A9 in Appendix H provide the corresponding training curves.

  4. We update the experimental setups for the ablation studies. Appendices G.1 and G.2 now describe the steering maps, initialization, intervention settings, optimization, and evaluation procedures, including angle selection for the jailbreak ablation and the settings held fixed across truthfulness matching scopes.

  5. We expand the computational resource reporting. The fitting-time table in Appendix D now includes peak allocated GPU memory during optimization, ranging from 0.54 to 4.74 GiB.

  6. We update the conclusion and corrected a minor typographical issue. The conclusion now reflects the empirical evidence relating cached fitting to actual intervened loss, and we have added the missing period identified by Reviewer EMkc.

AuthorsOpenReview ↗

Global Response

Dear AC and reviewers,

We thank the reviewers for their thoughtful reviews and constructive comments, which have helped us improve the paper. We are encouraged by the endorsement that: 1) the motivation for moving beyond marginal matching is clear and well supported (Reviewers EMkc, mCta), 2) our theoretical framework provides a principled analysis with carefully stated boundaries (Reviewers mCta, g7mr), 3) our method delivers meaningful practical improvements over existing steering baselines (Reviewers mCta, g7mr), and 4) our fitting procedure has broad applicability across multiple model families and steering parameterizations (Reviewers mCta, g7mr).

We address three main concerns below and provide detailed responses to each reviewer separately.

  1. Contribution, scope, and baseline selection [Reviewers EMkc, mCta]. We clarify the novelty of our framework and the rationale for our experimental scope and baseline selection. Our main contribution is the formulation of activation steering as matching activation path laws, together with its theoretical analysis and practical fitting objective. MMD is only one choice of discrepancy within the framework, as discussed in Remark 3.3. Our experiments examine whether this objective improves representative steering map families: translation, rotation, projection, and affine transformations. We focus on offline fitting from cached activations without additional backbone passes, whereas LoRA, ReFT, and full fine-tuning require computation through the backbone during training. The selected baselines therefore reflect the scope of our work.

  2. Connection between theoretical guarantees and cached fitting [Reviewers mCta, g7mr]. We clarify that Theorem 4.3 about path law identification does not require shared inputs. The shared input assumption applies to the functional guarantee in Theorem 4.5. Regarding the relationship between the cached loss and the actual intervened loss, we provide both theoretical and empirical evidence. Theoretically, Proposition C.1 upper bounded the square root of the actual intervened loss by the square root of the cached loss plus residual terms. Empirically, we have conducted an additional experiment on the jailbreak task using Qwen3.5-27B, GPT-OSS-20B, and Gemma-4-31B-IT to show that the cached loss serves as an effective surrogate for the actual intervened loss, which is showed in Table 6 in the revised manuscript and the corresponding curves provided in Appendix H.

  3. Benefits of cross-layer matching and stability across random seeds [Reviewers EMkc, mCta, g7mr]. We have expanded the ablations comparing marginal and cross-layer matching under the same steering parameterization on both jailbreak and truthfulness tasks. Across both tasks, full-trajectory and contiguous-group matching outperform marginal matching, supporting the benefit of capturing dependencies across layers. Although full trajectory matching suffices theoretically, we empirically show that including shorter contiguous groups provides additional learning signals with limited data and improves performance in our ablations. Regarding stability across random seeds, we have expanded the jailbreak ablations to five seeds and report means and standard deviations. We also compare the learned steering planes across seeds, finding nearly identical planes with a minimum similarity of 0.999917. These results, together with the consistent benefits across both tasks, strengthen the empirical support for cross-layer matching.

We hope that our rebuttal has helped to clear concerns about our work. We are glad to answer any further questions you have on our submission and we would appreciate it if we could get your further feedback at your earliest convenience.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.