acceptodds
Under review as a conference paper at ICLR 2027

Endpoint-Sensitive Causal Audits of Mixture-of-Experts Safety

Abstract

Expert-routing interventions in autoregressive mixture-of-experts models can alter both the next-token prediction and the attention state retained for later decoding, while the measured safety effect further depends on whether a prefix or the complete answer is evaluated. We introduce a four-cell token/state experiment that isolates these pathways, using prefix and complete-answer endpoints, paired and different-request donors, and a final-layer continuation invariant. In a three-model development audit with 279 prompt pairs per model, changing the evaluated span alters which pairs appear unstable even when aggregate flip counts agree. A frozen 49-request study then shows that, with the next token fixed, paired routing transfer changes decoded text in 32/33 naturally completed OLMoE misinformation comparisons and 20/21 Qwen3-MoE comparisons, yet changes no complete-answer compliance labels; a dense residual comparator similarly alters text in 19/31 comparisons without affecting compliance, and different-request donors produce comparable text changes. Few native MoE answers are classifier-positive, limiting conclusions about repair, while one reserved OLMoE case exhibits a nonzero state contrast under prefix grading and zero under complete-answer grading (with a selected development case showing the complementary pattern), and execution audits identify four auxiliary factuality trials that fail strict position isolation. These findings provide bounded counterexamples to treating token recovery, changed text, and measured safety repair as interchangeable evidence, and support joint audits of decoding pathways, numerical validity, and evaluation endpoints.

Primary area:
interpretability and explainable AI
Keywords:
mixture-of-experts, causal interventions, mechanistic interpretability, AI safety, activation patching, evaluation endpoints

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_p2rLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates how single-step expert-routing interventions in sparse Mixture-of-Experts based autoregressive large language models affect subsequent generated tokens and measured safety compliance. To disentangle decoding pathways, the paper proposes a four-cell decoder audit that separates the effect of the immediate next token generated and the retained KV Cache for the experiments. Experiments show that state-only routing transfers do change the subsequent generation, but they produce little change in the complete-answer compliance. The paper also identifies isolated cases where evaluating response prefixes versus complete answers can lead to different conclusions about intervention effects.

Critical strengths

The paper investigates a relevant question about whether changes in expert routing affect safety behavior through a systematic four-cell audit. The experiments include several useful controls, such as different-request donors, a dense-model residual comparator, and final-layer controls.

Critical weaknesses

The novelty of the proposed approach is limited, as separating the generated token from the KV Cache at a single decoding step into a four-cell audit is a relatively straightforward decomposition, and it is unclear about what new insights it provides specifically about MoE safety. The empirical evaluations is very limited, e.g. it contains only 17 HarmBench misinformation requests and 32 TruthfulQA questions in the held-out cohort. Also, as Qwen3-MoE and OLMoE only comply with 0/21 and 2/33 requests natively, the setup cannot meaningfully test safety repair. TruthfulQA evaluation relies on automated LLM graders without human evaluation conducted to validate the results. On writing, the paper is difficult to follow, with excessive procedural details, repetitive qualifications, and unclear organization that obscure the main contributions, probably due to use of LLM in writing.

Questions for authors
  • Can the authors provide a more comprehensive evaluation, for example, with a larger set of base prompts, more cases where models initially comply with harmful requests, and human validation for the TruthfulQA grading? Currently, having only 2/33 OLMoE and 0/21 Qwen3-MoE native positive comparisons provides little insight into safety repair.
  • Can the authors clarify what new insights the proposed four-cell audit provides specifically about MoE safety, beyond the existing activation-patching approaches?
Reviewer_UYvTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I had a difficult time parsing the text of this paper. I believe my understanding to be incomplete due to how dense the text was. From my understanding, the authors seek to ask whether choice of experts in MOE has an impact on overall harmfulness of a model, and how this perceived effect changes depending on if you look at the entire generated model output. They design interventions that, like in activation steering, augment the model during generation. They then try to measure the difference between a changed model state and a changed decoded token, although as stated below I believe their approach has confounders. They also see if they can "repair" poor model behavior with this intervention, although their sample sizes for some models is n=2. Resultantly, I think that the question is interesting, but I find it hard to glean any convincing takeaways from this paper.

Critical strengths

None

Critical weaknesses

Clarity and terminology

A lot of terms are extremely loaded/undefined which makes the text illegible.

  • "Our different-request donors control topic and approximate prompt length, leaving perturbation magnitude as a limitation."
    • It is essential to define what is meant by donor, control topic, and perturbation magnitude.
    • Define what "development cohorts" are and why you are studying them.
    • What are "mechanical newline variants", give an example since this is central to your experiments.
    • "10 exact repeats" How do I interpret this?
    • Sentences like "Excluding repeats gives 379 responses and 279 original-to-variant comparisons clustered within 100 base requests." This is far too dense to comprehend.
    • The 384-token view is extracted from the same generated token sequence. What is this "view". Do the authors mean that you are truncating the final generated output?
    • "Inspecting the pilot" What is the pilot? Is this referencing some preliminary "smoke-test" that didn't make the final paper? This reads as though the authors asked claude to write this paper, and the model is referencing experiments that were initially ran but did not make the main body or appendix.
    • "The new cohort was fixed on September 16 before its target generations." I understand the importance of clarifying the fixing of the cohort, but why is the specific date necessary?
  • Direction in an intervention is not clearly defined.
  • All of 5.3 is very hard to parse. "The case"? Do the authors mean they did a case study on a particular sample? If so they should say this clearly!
  • This is a completely non-exhaustive list.
  • Also the introduction has no citations!

Motivation of key aspects of this project and relevance

  • I am trying to understand why any safety evaluation, especially one that is testing a model's ability to reply in an open ended fashion, would only look at part of the generated output. This is sometimes the norm for multiple choice benchmarks but these are not the evaluations the authors are working with. Why then is a critical part of this project comparing a "view" of the generated output with the full output. Is this a problem the authors generally have with safety evaluations? If so, they need to justify that this is a practice that people do, but again there are no citations in the introduction that motivate this.

Intervention design and interpretation

  • Expert choice vs Routers choice is confusing.
    - From my understanding, the goal was to test whether the selection of experts matters or the relative weighting of the experts chosen. However since the weights are being renormalized anyways this experiment is not cleanly testing this question. Also, can the authors describe in more detail how this question would be of interest for safety? Suppose, the key factor for harmfulness was the weighting of experts, what suggestions would the authors have for future interventions during open-ended inference?
  • Secondly, I do not find the lack of "repair" to be surprising here. I think the interventions they are comparing have the potential of being too brittle to justify any claims of repair. Once an intervention or a shift to the MOE's selected is applied, you are changing the models distribution. This is the same if you force a token that is low probability under the state of the model. Since both mixed conditions could be out of distribution at this point, I think whether or not you observe repair is not informative of the particular export selection or weighting. There are too many interaction impacts. A first check that should be run is recording the forced token’s probability under the intervened state and vice versa.

Evaluation and strength of evidence

  • The authors do not emphasize what the classier they are using to measure repair is in the main text.
  • From my parsing of this section it seems that the judge will rate examples as noncompliant even when the model still cites the false fact? If this is the case shouldn't this discredit how strongly one can improve safety via a MOE intervention? Or at least how strongly one should take the judge results as a signal for safety improvement.
  • If the judges used to grade responses on TruthfulQA (a widely deployed and trusted dataset) seem to produce nonsensical classifications when the authors implement the benchmark, why did they not try to use the graders applied in the original paper? This is the obvious fix. The fact that this was not addressed casts severe doubt on the reliability of the quality of the LLM judges used in the paper, especially the Harmbench results.
  • Another concern is that the sample sizes the authors are working with are too small to carry any robust takeaways. For some models, since they are mostly aligned and do not produce harmful outputs, the sample size winds up being n=0, and for others it is n=2. This is far too few to make a robust scientific claim and thus some of the key results are not convincing.
  • Since the authors are not claiming that MOE edits can repair harmful behavior this is fine, however due to small sample size there is no scientific takeaway from this section at all!

Placement of implementation details

  • 3.3 is an implementation detail that is not central to the narrative. The only purpose of running this intervention at the last layer is to check for bugs in the code. This does not belong in the main text and possibly not even in the appendix.
  • "Each attempt records source hashes, model revision, library versions, and device information. Atomic per-treatment records support restart without overwriting completed outputs. Checks include native next-token replay, bitwise self-intervention replay, complete native continuation replay, restoration after intervention, and final-layer cache and continuation invariants. CPU integration tests exercise all three model implementations, and earlier tests additionally check cached prefixes of different lengths."
  • It feels as though 70% of the text is composed mainly of these extremely dense, and unnecessary, implementation details.
  • Overall I believe that the current presentation of the results and the motivation are too dense to be conference or submission ready.
Questions for authors
  • I believe that the changes required to materially affect my assessment constitute a complete rewrite of the paper, along with some massive changes approach to experimentation. I think that would require a completely new review cycle. However, I do think the substantive question worth answering is how is harmful behavior localized to different experts in MOE modes and can we steer model behavior by altering expert selection. It might be worth thinking about the simplest possible ways to approach this question first.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.