acceptodds
Under review as a conference paper at ICLR 2027

Decomposing Refusal Steering in Mixture-of-Experts Models

Abstract

Safety alignment in instruction-tuned large language models (LLMs) depends on a model's ability to reliably refuse harmful or disallowed requests. Recent work has shown that a steering vector can be applied to a dense LLM during inference to suppress refusal behavior and induce responses to harmful requests. We extend this refusal steering method to three open-source Mixture-of-Experts (MoE) LLMs to decompose and better understand how refusal mechanisms operate across MoE components. We find that steering performance is uninhibited by the complex routing patterns inherent to the MoE architecture, and that a single expert recovers 78% of the full steering effect on average when free to select its own location. However, expert-level steering only recovers 54% on average when constrained to the location used by full-layer steering, a gap that widens further under safety-related system prompts. Our results also show that refusal signals captured by these steering methods differ from those captured by expert routing behavior. Together, these results suggest that refusal behavior is distributed unevenly across the MoE LLM architecture rather than governed by a single unified mechanism.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, alignment, steering, mixture-of-experts, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rB1HOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3(+1)
Soundness
2(+1)
Read review
Paper summary

This paper investigates whether refusal steering in MoE models can be decomposed into expert-specific contributions. The authors further find that the most steerable layer does not necessarily contain the most steerable individual expert.

Critical strengths
  1. Expert-level analysis of refusal steering.
  2. Interesting single- and all-expert steering settings.
Critical weaknesses
  1. Limitations in experimental design.
  2. Lack of in-depth mechanistic analysis.
Questions for authors

The authors explore two approaches to expert-level refusal steering in MoE models: single-expert and all-expert steering. By comparing fixed and flexible intervention locations, they find that the most steerable layer does not necessarily contain the most steerable expert. While the paper provides interesting insights into the role of individual experts in refusal steering, I have the following concerns:

  1. Impact of post-attention residual activations on refusal steering. It is valuable to investigate how individual experts contribute to refusal behavior in MoE models. However, I think the current experimental design does not sufficiently support the authors' conclusions. In ActAdd, the steering vector is extracted from the full layer output, which includes the post-attention residual stream and the aggregated expert outputs. In contrast, expert-level steering vectors are extracted only from individual expert outputs, without directly incorporating the post-attention residual stream. Therefore, the observed differences in steering effectiveness may also be influenced by refusal-related information in the post-attention residual stream. I suggest that the authors further investigate its contribution to refusal steering before drawing conclusions about the role of individual experts.

  2. The role of KL divergence in experimental comparisons. If I understand correctly, KL divergence is an important factor in ensuring a fair comparison across different steering settings. The authors should ensure that different steering methods result in comparable KL divergence values, so that differences in ASR can be attributed to steering effectiveness rather than varying degrees of disruption to the model's original behavior. Although KL divergence is used during candidate selection, it is unclear whether the final selected interventions across different settings achieve similar KL divergence values. I would appreciate more details and analysis on this aspect.

  3. Readability and presentation. The paper's structure and presentation could be improved for better readability. For example, the introductory paragraph of Section 3 (Methods) contains extensive experimental details, which could be moved to the Experiments section. In addition, some mathematical notations could be more standardized. For instance, instead of using kl and rs to denote the KL-divergence and refusal scores, the authors could consider using a more consistent notation. Finally, some experimental results would benefit from clearer explanations. For example, the reported 54%–78% ASR recovery around line 302 is not immediately clear. Only after carefully examining Table 1 did I realize that these values refer to the averages of the corresponding columns. Explicitly stating this would improve readability.

3 comments
AuthorsOpenReview ↗

rB1H Rebuttal (1/1)

Thank you very much for your time and consideration in reviewing our paper. We have uploaded a revised manuscript that reflects your and other reviewers' feedback and have summarized the changes therein in a general comment to our submission. We will refer to the changes outlined there in our response so as not to repeat content in multiple places.

To briefly address your questions, in order:

  1. We agree that the post-attention residual stream likely plays an important role in refusal behavior, and identify its contribution as the gap between ActAdd, whose direction is computed from the full residual stream, and our expert-specific steering vectors (lines 176-185). In lines 329-331 we describe how the partial ASR recovery of expert-specific methods indicates that a portion of the refusal signal comes from the post-attention residual stream.

While our work demonstrates that the MoE sublayer holds only partial refusal signal, we leave for future work a decomposition of the remaining signal coming from both the attention output and the residual stream from previous layers. We have edited the language in our contributions section to better reflect the sources in our decomposition:

  • We explore two methods for expert-level refusal steering that isolate the contribution of individual experts to ActAdd’s aggregate signal, with the remainder attributable to other components of the residual stream. Steering based on the output of a single expert recovers 78% of ActAdd ASR on average, demonstrating that individual experts carry substantial refusal signal.
  1. We have added a visualization of "non-response" (degenerate) generations and ASR vs KL-divergence (Figure A1) and a table of KL-divergence values for the selected intervention locations (Table A6) as Appendix E.2. We briefly explain that, while KL-divergence is positively correlated with response degeneration, we account for this by including degenerate responses as steering failures and by balancing KL-divergence with ASR in our selection criteria. Based on this design and the evidence in Table A6, we are confident that differences in ASR reflect steering effectiveness rather than model disruption.

  2. Thank you for these notes on clarity. We have made the following changes in our manuscript based on your and other reviewers' feedback:

  • We have added the subsection title 'Experimental Details' between the first and second paragraphs under 'Methods'
  • We now denote refusal score and KL-divergence as s_ref and D_KL.
  • We have updated the main paper text to clarify that our ASR calculation only counts "full-response" generations.
  • We have updated Table 1 to include column means.

Thank you again for your time and attention in helping us to improve our paper. We hope that these additions address your concerns about experimental design and clarify the scope of our mechanistic claims, and welcome further discussion as needed.

Hello, We wanted to send a gentle reminder as the discussion deadline is approaching. We hope our response and the additional experiments addressed your concerns, and we would be grateful if you could reconsider your assessment in light of our revised response. We would also greatly appreciate any further feedback you may have before the deadline. Thank you again for your time and consideration.

Thanks for your clarification. My second and third concerns have been resolved, but the first remains unresolved. I'd like to increase my Presentation and Soundness scores while keeping my Overall Rating unchanged.

Reviewer_yTM7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies refusal steering in Mixture-of-Experts (MoE) language models by extending ActAdd to expert-level interventions.

The authors propose two approaches: single-expert steering, which uses a steering direction computed from an individual expert, and all-expert steering, which combines directions from the selected top-k experts. Experiments on three MoE models under different system prompt settings show that expert-level steering can recover a substantial portion of ActAdd's effectiveness, but performance depends strongly on the intervention location. The authors also find that safety experts identified through routing frequency differences are not necessarily predictive of effective refusal steering directions

Critical strengths
  1. The experiments cover three MoE architectures and different system prompt settings.
  2. The paper also investigates the relationship between expert routing statistics and steering effectiveness. I think this is important because expert routing statistics are a very common indicator during MoE training, and if it's not related to steering effectiveness, it's good to know for the community, as it's very surprising.
Critical weaknesses
  1. The definition of safety experts may limit the conclusions about routing-based steering. It's mostly because of top-1 vs. top-k. In Section 3.1, safety experts are identified using top-1 routing frequency differences, although MoE models use top-k experts for their actual FFN computation. Why should top-1 routing statistics be sufficient to identify safety-related experts? Other experts in the top-k could still contribute significantly to refusal behavior. The main predictive analysis in Tables 2 and 3 relies on top-1 statistics. More importantly, as one of the key conclusions, the authors argue that refusal signals are distributed across multiple expert outputs and layers rather than concentrated in a single location. However, given that MoE models naturally combine multiple experts through top-k routing, I find this observation somewhat unsurprising. More importantly, I find it hard to understand which experimental evidence in the main body directly supports this claim, as Table 1 mainly compares steering effectiveness across different methods rather than directly analyzing how refusal signals are distributed.

  2. In Table 1, there is considerable variation across models and system prompts, particularly for the fixed-location interventions. I find it hard to extract conclusions from it.

  3. Limited novelty and broader significance. Both refusal steering and MoE mechanistic interpretability have been studied extensively. The methods here mainly extend existing ActAdd-style steering to MoE expert outputs, without introducing substantially new analysis tools. The setting is also relatively narrow, focusing on ActAdd-based refusal steering in MoE models. While the empirical observations are interesting, I am not sure what additional mechanistic understanding we gain from these results, or what the actionable next steps would be. For example, should these findings inform better interpretability methods, more effective steering, or changes to MoE architectures? I do not think an interpretability paper necessarily needs to improve model performance, but I would expect more concrete mechanistic insights given the limited methodological novelty.

  4. Insufficient discussion of related work. The paper does not have a dedicated related work section, and I think it would benefit from a more thorough discussion of existing MoE mechanistic interpretability research. I'm pretty confident that there are quite a few.

Questions for authors

See Weakness above.

  1. One small comment: for the system prompt part, I do think some refusal behavior may lie in the attention part. See https://arxiv.org/pdf/2505.12025 and https://arxiv.org/abs/2607.26228. So it's good to do some verification of your hypothesis with these tools as follow-ups.
  2. It would be easier to understand the different methods (this is based on my understanding) you used for comparison in the experiments if you include something like this, as the names of the methods don't provide much information now.
MethodSteering DirectionExpert SelectionLayer / Token PositionOptimized Parameters
ActAddFull residual streamN/AFreeℓ, i
SafetySESingle expertExperts with large absolute diff_eFreeℓ, i, e*, c
AllAEWeighted top-k expert directionsNaturally selected top-k expertsFreeℓ, i, c
FixedSESingle expertAll experts at ActAdd's selected layerFixed to ActAdde*, c
FixedAEWeighted top-k expert directionsNaturally selected top-k expertsFixed to ActAddc
5 comments
AuthorsOpenReview ↗

yTM7 Rebuttal (1/2)

Thank you very much for your time and consideration in reviewing our paper. We have uploaded a revised manuscript that reflects your and other reviewers' feedback and have summarized the changes therein in a general comment to our submission. We will refer to the changes outlined there in our response so as not to repeat content in multiple places.

To briefly address your list of critical weaknesses, in order:

  1. We'll break this point into two related questions:
  • Shouldn't the top-k experts be considered when computing steering directions? We did not perform a complete experiment using top-k statistics to select steering directions due to compute limitations, but a single test run (OLMoE with the Llama-2 system prompt) resulted in a decrease of 20 percentage points (81% to 61%) in recovered ASR compared to using the top-1 statistics. It is possible that this decreased steering performance reflects a codependency of the top-k experts, such that steering with the single top expert cannot recover the combined refusal signal of the top-k experts. While we do not directly test steering based on the combination of a static set of top-k safety experts, the all-experts methods (AllAE and FixedAE) steer with the top-k experts for each token and do not show substantial improvement over single-expert methods.

We also added a table to Appendix E.3 comparing top-1 and top-k routing statistics. The two statistics identify largely different experts (8–31% overlap among the top 5%), but neither reliably predicts srefs_{ref}, and top-1 more often shows the expected negative correlation (6 of 9 settings vs. 3 of 9). Our finding that routing-identified experts don't predict steerability therefore isn't an artifact of using top-1 statistics.

  • Is it surprising that refusal is distributed across multiple experts? And what evidence points to this distribution? We have updated the language in the main text describing this result to more accurately reflect our intended conclusion, which is not that our results demonstrate refusal signal distribution across multiple experts within a layer (e.g. the top-k "safety experts"), but instead that substantial refusal signal available for ActAdd-style steering exists at several different locations within the model. The evidence for this comes from two sources: 1) over almost every model and system prompt combination, the three free-location methods (ActAdd, SafetySE, and AllAE) select different intervention locations (Appendix E.1); 2) the ASR gap between the ActAdd and fixed-location methods indicates that some portion of the refusal signal bypasses the MoE sublayer (i.e. it exists entirely outside of the MoE sublayer).
  1. We have updated Table 1 to include the column means, which are what we most often cite to support our analyses in the main text. As for the variation across models and system prompts, we agree that our results do not provide a simple picture of how refusal behavior is distributed within MoE models. We provide some discussion about our hypotheses for the apparently inconsistent results in Appendix D, in which we discuss the non-monotonic effect of the steering coefficient cc. We suggest that our results reflect nonlinearity in the model's refusal subspace, which, combined with the limited grid search over cc necessitated by our compute budget, resulted in a less clear picture of how experts contribute to refusal behavior. We have found that many recent works from larger labs that rely on ActAdd-style steering mention a tuned steering coefficient, without detailing how that coefficient was optimized or selected for the most effective interventions (Steering Llama 2 via Contrastive Activation Addition Rimsky et al. 2024, On the Biology of an LLM Lindsey et al. 2025).

While we would have liked to perform a more complete grid-search over possible values for cc to produce a smoother table of results, we believe that the overall patterns in our results support our analyses and conclusions, despite noise seen in individual cells.

AuthorsOpenReview ↗

yTM7 Rebuttal (2/2)

  1. While we acknowledge that there is quite a bit of existing work around refusal steering and a smaller body of existing work around MoE mechanistic interpretability, we would argue that the novelty of our contribution lies in the decomposition of that behavior among model components. To the best of our knowledge, at the time of writing there were no published works that explored the specific contribution of model components to refusal steering--some feature-related work, such as Scaling Monosemanticity (Templeton et al. 2024), mentions incomplete steering signal from the FFN as a caveat, but not as an object of study. There have since been two works accepted to COLM 2026 (Component and Dimension Sparsity in Transformer Refusal Mechanisms, Siu et al. 2026) and EMNLP 2026 (What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal, Cheng et al. 2026) that decompose refusal behavior within transformer LLMs, though they run their experiments on dense models and, thus, focus their efforts on other ways of isolating components of the models' refusal mechanisms.

On the applicability of our results, we suggest three actionable effects or future directions that we have integrated into the Discussion of the updated manuscript:

  • Our demonstration that routing statistics do not capture the same refusal signal used in ActAdd-style steering indicates that a routing-based safety audit may miss a substantial component of the model's refusal mechanism.
  • Our system prompt experiment suggests that contextual refusal behavior exists outside of the FFN, which indicates future work to explore the contribution of the attention sublayer to prompt-based safety guardrails.
  • The finding that the full ActAdd method selects a different intervention location than expert-specific methods also suggests that there may be more to learn about the balance of refusal signals coming from different model components at any given layer. Building on the work exploring the attention sublayer's contributions, future work could trace the balance of refusal steering signal coming from different components as it moves through the model.
  1. We removed quite a bit of our discussion of related work from the Introduction and Preliminaries sections to meet the ICLR page limit, but we agree that we should have moved this content to an Appendix instead of removing altogether. We have added a short discussion of related work that is not central to our own methods and analyses, some of which has been published since writing our original manuscript, in Appendix F.

In short, we provide context from several existing bodies of work: those that demonstrate the effectiveness of linear steering methods, those that decompose model behavior across model components, those that explore the role of experts in behavioral specialization, and those that use specialized experts as a potential attack surface. We then briefly discuss the two contemporaneous works described above, and differentiate our methods and results from their related work on dense LLMs.

To address your questions:

  1. Thank you for sharing these papers. We have included them in our new Related Work Appendix F and appreciate their relevance to future work in examining the role of attention in refusal behavior.
  2. Thank you for the suggested methods table, which we agree does a nice job of outlining the distinguishing characteristics of each steering method. We have included a slightly modified version of this table in our Methods section.

Thank you again for your time and attention in helping us to improve our paper. We welcome any further feedback or discussion that you are able to provide.

Thanks for the detailed response! I appreciate the additional top-k analysis and the clarification of the claims about distributed refusal signals. I understand the distributed refusal signals claim better now.

For the top-k analysis, I think the new routing statistics partially address my concern. Also I feel "single test run (OLMoE with the Llama-2 system prompt) resulted in a decrease of 20 percentage points (81% to 61%) in recovered ASR compared to using the top-1 statistics." such a big performance gap is very surprising for MoE and might be an advantage of your method? How should I interpret this result correctly, and what are the roles for top-k>1 routers?

I also appreciate the discussion of potential implications, especially the points about routing-based safety audits and the role of attention in system-prompt-based refusal. I think both are interesting directions, but I would like to see some more direct evidence supporting them. Given the compute limitations and the remaining rebuttal time, I think even one small experiment could help:

For routing-based safety audits, could the authors test how well top-1/top-k routing statistics predict actual refusal behavior, rather than steering effectiveness? For example, on one model, they could compare a simple classifier trained on prompt-level routing statistics against a residual-stream refusal score (e.g., projection onto a direction computed from the mean activations of naturally refused vs. answered prompts). Comparing held-out AUROC would help establish whether routing statistics capture the same refusal-related information as the residual stream. If routing is substantially less predictive, I think this would provide a stronger motivation for why routing-based safety analysis alone may be insufficient.

I think any experiments to support your listed three actionable next steps could provide a stronger motivation for the paper in the Intro, depending on the results. I only give an example for your first bullet point, and feel free to pick other directions. I don't think an exhaustive new study is necessary, but one targeted experiment would help connect the current observations to a more concrete mechanistic finding. Otherwise, I think the three implications would be better presented as hypotheses for future work rather than conclusions directly supported by the current experiments. I wanted to see at least part of them in the current paper to strengthen its contribution.

By the way, I recommend authors to highlight their change in another color in the updated manuscript if possible.

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

This paper applies the ActAdd steering method to isolated experts instead of the full residual stream and evaluates safety refusal behavior at the expert level using this method. They also search across layers and token positions to apply this method and show that a single expert intervention can recover 78% of the full steering effect. They conclude that refusal is distributed unevenly across the MoE LLM architecture.

Critical strengths
  1. The idea of using ActAdd on single experts is interesting, and the writing is easy to follow, which I really appreciate.
Critical weaknesses
  1. Although reliance on smaller LLMs is mentioned in the limitations, in the context of MoE LLMs, the number and size of experts, as well as the quality of training data, training recipe, and safety post-training, can potentially have a significant effect on expert specialization. The fact that Table 3 supports the paper's conclusions only for the older Mixtral and OLMoE models, and not for GPT-OSS 20B, which is more recent and likely benefits from more advanced training data and techniques, is concerning. This may suggest that routing-based methods work better in more modern and better-trained models. Therefore, the conclusion that "routing-based safety experts are not reliable predictors" seems too broad given the limited models tested.

  2. Moreover, based on my understanding, their diffediff_e in Eq. 9 for detecting routing-based safety experts differs from the settings used in related work. They seem to only record top-1 instead of top-k, ignoring all other activated experts in a layer (Table A6 reports entropy to measure the concentration of routing differences, but this does not establish whether top-1 or top-k routing reliably identifies effective safety experts. A more meaningful test of the identified experts' importance could be to iteratively deactivate the top-n experts and measure changes in refusal behavior, as explored in prior work). They only use the arbitrary last five tokens of the prompt instead of all prompt tokens or response tokens, even though input prompt expert activations may be related to harmfulness detection and not necessarily refusal generation. They also seem to use separate harmful/harmless sets instead of matched pairs, with only 200 examples in each class. All these choices could change which experts are selected by their routing-frequency method, yet no alternative identification methods or rankings from related work are tested for comparison.

  3. I see two important contributions: (1) "Individual experts carry substantial refusal signal," which, although demonstrated here using ActAdd, is closely related to findings established in previous work and therefore has limited novelty. (2) "The refusal signal driving effective steering is distinct from that identified by expert routing statistics," which relies on a limited set of models and questionable methodological choices. This makes the claim too broad, and the conclusion that routing-based methods are unreliable seems insufficiently supported.

Questions for authors

None

4 comments
AuthorsOpenReview ↗

eSbF Rebuttal (1/2)

Thank you very much for your time and consideration in reviewing our paper. We have uploaded a revised manuscript that reflects your and other reviewers' feedback and have included those changes in a general comment to our submission. We will refer to the changes outlined there in our response so as not to repeat content in multiple places.

To briefly address your list of critical weaknesses, in order:

  1. While we agree that our limitation to smaller LLMs is not ideal, we respectfully disagree with your concern about the results in Table 3 (now Table 4). While these results from our random expert pool experiment suggest that, in the OSS model, routing-based safety experts produce stronger steering vectors on average than a random selection of non-safety experts, the results in Table 2 (now Table 3) still support the conclusion that routing statistics are not reliable predictors of steering performance. Here, we find that the OSS model's routing-based diff_e is not strongly correlated with the refusal steering score srefs_{ref} (ρ\rho = -0.21, 0.32, -0.02 over our three system prompts), that only 3-7 of the top-10 safety experts are retained after the first selection filter, and that the selected expert is typically not highly ranked according to diffe\text{diff}_e (R = 25, 5, 10 over the three system prompts).

However, we agree that the results in Table 3 (now Table 4) could indicate that routing-based refusal signal is stronger in more modern/mature models. As such, we have added the following footnote to our discussion of these results:

One possible explanation for the OSS model results in Table 3 (now Table 4) is that a model with more mature or frontier-style training holds stronger signal in its routing patterns. This could also explain the OSS model's high susceptibility to the routing-based jailbreaking methods in Lintelo et al. (2026), though a targeted experiment comparing the strength of safety experts on larger or frontier models would be required to support this hypothesis.

AuthorsOpenReview ↗

eSbF rebuttal (2/2)

  1. We will break this point into two separate discussions:

Safety Expert Identification.

We did not perform a complete experiment using top-k statistics to select steering directions due to compute limitations, but a single test run (OLMoE with the Llama-2 system prompt) resulted in a decrease of 20 percentage points (81% to 61%) in recovered ASR compared to using the top-1 statistics. We have added a table comparing top-1 and top-k routing statistics to Appendix E.3. In this table, we see that the two statistics identify largely different experts (8–31% overlap among the top 5%), but neither reliably predicts srefs_{ref}, and top-1 more often shows the expected negative correlation (6 of 9 settings vs. 3 of 9). Our finding that routing-identified experts don't predict steerability therefore isn't an artifact of using top-1 statistics.

While iterative expert ablation, as performed in Lintelo et al. 2026, is quite effective in identifying the right combination of routing-based safety experts needed to bypass refusal behavior, this answers a different question than what we target with the SafetySE method in our paper: whether routing-based safety experts control the same component of the model's refusal mechanism behind ActAdd-style refusal steering.

We concede that our methodological choice for the token position, while supported by the previous work listed below, may have inadvertently identified a different type of safety expert than those that would produce effective steering vectors. We have added that consideration to our proposed future work to explore harmfulness detection vs. refusal generation in our updated Discussion section.

Grounding for Methods.

We acknowledge that there are several common methodological choices from previous works regarding the isolation and identification of refusal behavior in generative LLMs. Our methods are grounded as follows:

  1. We agree with the contributions that you have identified, but suggest that a reframing would address your concerns about their novelty, significance, and validity. Namely, we would like to highlight that our methods, results, and analysis are focused on the component of a model's refusal behavior that is used in ActAdd-style refusal steering.
  • While we agree that it has been established via routing-based methods that individual experts carry refusal signal, to the best of our knowledge no one has attempted to use these or other individual experts to compute refusal steering directions. We believe that the relative success of our single-expert steering intervention contributes a meaningfully novel perspective to the well-established routing-based result.
  • It was not our intention to claim that routing-based jailbreaking methods are unreliable, as is clearly contradicted by the results of papers such as Lintelo et al. 2026 and Wu et al. 2026. The claim that our results support is that routing-based methods are unreliable predictors of effective steering directions, indicating that they might identify different components of the model's refusal mechanism. We briefly discuss the implications of this result on expert-related safety audit methods and propose future work to further explore the roles that these different types of safety experts might play in our updated Discussion section.

Thank you again for your time and attention in helping us to improve our paper. We welcome any further feedback or discussion that you are able to provide.

Hello, We'd like to send a gentle reminder as the discussion deadline is approaching. We hope our response and the additional experiments addressed your concerns, and we would be grateful if you could reconsider your assessment in light of our revised response. We would also greatly appreciate any further feedback you may have before the deadline. Thank you again for your time and consideration.

Hello, thank you for the detailed response. I appreciate the effort put into the revisions, particularly the blue-highlighted changes, which made them easier to follow.

I normally understand resource limitations, but in this case, making claims about "routing-based methods" in general requires testing various methodologies, including top-k routing, prompt vs. response tokens, different comparison metrics, dataset variations, and larger, more specialized models. Routing-based expert identification is a delicate methodology that heavily depends on these choices. Therefore, I believe the paper should either cover these variations to support its general claims or significantly narrow its conclusions to one specific method from prior work, using its exact methodology, models, and datasets, and limit the claims to that setting alone, rather than the entire category of routing-based methods. For this reason, my main concern remains. But I acknowledge the authors' efforts and am raising my rating from 1 to 2. (minor formatting issue: the captions of Figure 3 and Tables 3 and 4 overlap on page 9)

Comments on the paper

6 comments
AuthorsOpenReview ↗

Rebuttal Revision Changes (1/5)

We thank our reviewers and the ACs for your time in consideration of our paper. In response to the reviewers' feedback, we have made the following edits to our manuscript and have uploaded a revised copy. The full edits are included as general comments for all reviewers' convenience.

Summary of Changes

  • We have updated Table 1 (our main results) to include column means which are most often cited to support the analyses in our main text.
  • We have included a new table, suggested by yTM7, that clearly outlines our steering intervention methods.
  • We have updated our Discussion section to include more concrete applications and suggestions for future work stemming from our results.
  • We have added a new Appendix section (E.2) outlining the relationship between KL-divergence and our steering interventions.
  • We have added a table to the end of Appendix E.3 (previously E.2) comparing top-1 and top-k routing statistics.
  • We have added a full Related Work section to a new Appendix F, including context that we had previously removed due to the submission page limit and two highly-related contemporaneous publications.
  • We have made minor changes throughout the text to address issues of clarity and precision, with the goal of more clearly representing our intended conclusions and analyses.

Full Edits

To address rB1H's note about clarifying the use of column means and yTM7's note about difficulty reading the Table 1 results, we have updated Table 1 (now Table 2, lines 362-377) to include column means:

ASR% ActAdd ASR Recovery
ModelSPNoneActAddSafetySEAllAEFixedSEFixedAE
OSSN0.100.9478636987
LW0.000.66112775371
L20.000.8334661652
MixtralN0.460.9296984968
LW0.080.772777931
L20.060.779473110
OLMoEN0.640.90101929788
LW0.350.94846710062
L20.520.9381949583
Mean0.250.8578785461


Thank you to yTM7 for the suggested methods table, which we agree will improve clarity. We have included it in Section 3.2 with minor changes (Table 1, lines 271-280):

MethodSteering DirectionExpert SelectionLocationOptimized
ActAddFull residual stream, vℓ,iv_{\ell,i}N/AFreeℓ\ell, ii
SafetySESingle expert, ve∗,iv_{e^*,i}Experts with large |diff_e|Freeℓ\ell, ii, e∗e^*, cc
AllAETop-k weighted, {ve,i}e∈Eℓ\{v_{e,i}\}_{e \in E_{\ell}}Router top-k per tokenFreeℓ\ell, ii, cc
FixedSESingle expert, ve∗,iv_{e^*,i}All experts e∈Eℓe \in E_{\ell}ActAdd (ℓ,i\ell, i)e∗e^*, cc
FixedAETop-k weighted, {ve,i}e∈Eℓ\{v_{e,i}\}_{e \in E_{\ell}}Router top-k per tokenActAdd (ℓ,i\ell, i)cc
AuthorsOpenReview ↗

Rebuttal Revision Changes (2/5)

To address yTM7's concerns about the applicability of our results and eSbF's concern about using input text instead of generated text, we have updated our Discussion section (Section 5, lines 479-524):

The specific mechanisms behind LLM safety alignment remain poorly understood, and we believe that decomposing refusal steering at the level of individual architectural components offers a tractable direction for future work on model safety and interpretability. Our results point toward three areas of future work worth pursuing: How does the distinction between routing-based refusal signal and steering-based refusal signal affect potential safety audits in MoE models? Can routing-based safety expert identification be improved to better align with the signal that drives effective steering? And what role does the attention sublayer play relative to the FFN sublayer in mediating context-dependent refusal?

MoE Safety Audits. A natural approach to auditing safety in MoE models is to inspect routing behavior: identify the experts that are selected more often for harmful prompts than for harmless ones, and then monitor, harden, or ablate them (Lai et al., 2025; Fayyaz et al., 2026; Lintelo et al., 2026). Our results suggest that such an audit would capture only part of the picture. Since expert routing patterns do not reliably predict which experts yield effective refusal steering directions, an expert that looks unremarkable in its routing statistics may still carry a strong refusal signal that ActAdd-style steering can exploit. Further, the gap between ActAdd and fixed-location expert steering indicates that some portion of the refusal signal bypasses the MoE sublayer entirely, and that this portion grows under safety-related system prompts. An audit restricted to expert routing would therefore miss both the activation-level vulnerability within the experts and the signal carried outside of them. We suggest that safety audits of MoE models pair routing-based analysis with activation-level probes, such as the expert-specific steering vectors used here, since they appear to capture different aspects of the model's refusal mechanism.

Harmful detection vs. refusal production. Lai et al. (2025) suggest that there are two separate classes of safety experts: those that detect harmful prompts, and those that safely respond to (refuse) harmful prompts. It may be that our diffe\text{diff}_e metric more readily identifies the harmfulness detector group, and that the experts that were found to be most effective in refusal steering would fall more under the response control group. Future work using routing-based metrics computed on generated responses instead of input prompts or more specifically targeting refusal response patterns, such as those proposed in Lai et al. (2025) or Lintelo et al. (2026), may find stronger alignment with ActAdd-style steering directions.

Two refusal pathways. Our results showed that expert-level steering performance degraded in the presence of system prompts, which suggests that refusal behavior that is driven by system prompts may be more concentrated outside of the MoE feed-forward sublayer. One explanation for this comes from evidence that the attention sublayer is primarily responsible for processing input context, like system prompts, during generation (Olsson et al., 2022). Geva et al. (2023) extend this concept and find that the attention sublayer is responsible for "retrieving" information from the input, while the FFN sublayer is responsible for "storing" information learned during training. This mechanism would explain the pattern that we observe: the FFN sublayer controls the internal, learned refusal behavior, while the attention sublayer controls the refusal behavior as driven by the system prompt. A direct test of this theory would require isolation of the attention sublayer's contribution to refusal, analogous to the expert-level decomposition that we perform here. Once isolated, we could compare the balance of refusal signal coming from the attention and FFN sublayers throughout the model, in hopes of explaining the different intervention locations selected by the steering methods in our experiments.

AuthorsOpenReview ↗

Rebuttal Revision Changes (3/5)

To address rB1H's question about KL-divergence, we have added the following content to a new Appendix section E.2 (lines 1026-1073):

Figure A1 shows the relationship between KL-divergence and response degeneration for all candidate intervention locations below the model-specific KL-threshold during our second selection stage. These values should only be compared within each model, since KL-divergence will differ according to vocabulary size and the model's baseline logit distribution. As expected, response degeneration (responses that are labeled "non-response") increases as KL-divergence increases over all three models. We account for this expected relationship by selecting the intervention location with the highest ASR under this KL-threshold, and by counting "non-response" generations as failures, such that ASR does not increase due to degeneration (also evident in Figure A1).

fig:kl_div_compare is a three-panel line chart with KL-divergence on the x-axis and lines showing the non-response rate and ASR for each model

As shown in Table A6, the KL-divergence values (DKLD_{KL} from Equation 4) of the selected intervention location are not uniform across models, system prompts, or steering interventions. However, due to the combination of our three-label judgment protocol and the grid-search over both large and small steering coefficients, the selected interventions represent each method's best trade-off between steering strength and coherence, so differences in ASR reflect steering effectiveness rather than differing degrees of disruption. For example, free-location methods on Mixtral were selected at higher DKLD_{KL} than fixed-location methods, but produced a similar proportion of degenerate responses (Table A4).

Model (KL threshold)SPSafetySEAllAEFixedSEFixedAE
OSS (1.0)N0.230.310.880.46
LW0.640.920.940.58
L20.510.520.940.59
Mixtral (2.0)N1.801.480.310.27
LW2.001.640.260.33
L21.581.731.810.63
OLMoE (2.0)N1.811.620.930.13
LW1.741.431.250.57
L20.951.310.560.26


To address yTM7 and eSbF's questions about top-1 vs. top-k routing statistics, we have added the following table and discussion to the end of Appendix E.3 (lines 1126-1151):

ModelSP% Overlapρ top-1ρ top-k
OSS (k=4)N21−0.180.05
LW29−0.30−0.03
L2180.00−0.25
Mixtral (k=2)N15−0.76−0.52
LW310.430.55
L215−0.100.67
OLMoE (k=8)N100.250.09
LW12−0.040.38
L28−0.010.16

Table A8 compares the experts identified by top-1 and top-k routing statistics and how well they predict srefs_{ref}. Here we see that there is minimal overlap in the top-5% of experts between the two groups (8-31% across models and system prompts), supporting the result from Table A7 that the routing statistics change substantially when including the top-k experts. However, while neither set of routing statistics presents a clear correlation with srefs_{ref}, the top-1 ∣diffe∣|\text{diff}_e| more often has the expected negative correlation than the top-k ∣diffe∣|\text{diff}_e|. This is a somewhat surprising result, since we might expect that the top-k statistics (which more accurately reflect the natural behavior of the model) would provide a stronger refusal signal than isolating the top-1 expert alone, and supports our decision to use the top-1 routing statistics in our SafetySE experiment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.