acceptodds
Under review as a conference paper at ICLR 2027

From Capability to Alignment: Repurposing MoE Routing for Safety

Abstract

Mixture-of-Experts (MoE) models route each input to a sparse subset of experts, re- alizing capability through per-input expert specialization. Can this input-dependent routing also be used to improve safety? On a given input, some experts contribute to unsafe behavior while others support safe responses, suggesting that suppress- ing the former could improve safety. Realizing this idea raises two challenges. First, the identity of risk-amplifying experts shifts across queries, so the experts to suppress must be identified per input. Second, queries vary in risk level, so suppression strength must also be calibrated per input. Risky inputs may require stronger intervention, while benign ones should remain largely unaffected. To address these challenges, we introduce PIRA (Per-Input Router Alignment), an alignment-enhancing technique that adapts both the experts suppressed and the sup- pression strength at inference time, without updating model weights. PIRA derives two complementary signals from the same input representation: gradients of an output-safety score with respect to router logits identify active experts that amplify risk for the current input, while an input-risk score controls intervention strength. For queries above a calibrated risk threshold, a single forward-backward probe computes a per-query negative router bias that remains fixed during generation; other queries receive no suppression. Across three MoE backbones and six selected PandaGuard JailbreakBench attacks, PIRA improves average safe-response rates by over 20 percentage points relative to the original models. XSTest answer rates decrease by at most 2 percentage points, and MMLU accuracy by at most 0.3 percentage points. Additional evaluations further support the generality of PIRA across model backbones and safety settings. These results suggest that MoE’s per-input routing mechanism can be repurposed to enhance safety.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Mixture-of-Experts, LLM, Security

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_UQCVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper starts with two challenges. 1. The identity of risk-amplifying experts shifts across queries, so the experts to suppress must be identified per input. 2. Queries vary in risk level, so suppression strength must also be calibrated per input. To this end, the authors propose an alignment-enhancing technique that adapts both the experts suppressed and the suppression strength at inference time, without updating model weights. The results show the advantage of the proposed methods.

Critical strengths
  1. The writing is good, with clear figures.
  2. The analysis of 1. The identity of risk-amplifying experts shifts across queries, so the experts to suppress must be identified per input. 2. Queries vary in risk level, so suppression strength must also be calibrated per input is good and clear.
Critical weaknesses
  1. How can we guarantee an external judge’s verdict on the resulting continuation is reliable?
  2. An expert is risk-amplifying on this specific query if a small upward push on its router bias at this layer would lower the safety score. This claim lacks evidence.
  3. The evaluated model is not sufficient.
Questions for authors
  1. How can we guarantee an external judge’s verdict on the resulting continuation is reliable?
  2. An expert is risk-amplifying on this specific query if a small upward push on its router bias at this layer would lower the safety score. This claim lacks evidence.
  3. The evaluated model is not sufficient.
Reviewer_2bkgOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces PIRA (Per-Input Router Alignment), an inference-time safety intervention for Mixture-of-Experts language models. PIRA uses prompt-side risk and safety representations to identify, for each query, router experts that contribute to unsafe behavior and applies a query-dependent negative bias to suppress them. The intervention is activated only for sufficiently risky prompts.

Critical strengths
  1. PIRA directly exploits the input-conditioned sparsity of MoE routing. The per-query expert identification and risk-dependent modulation are conceptually well matched to the proposed safety problem.

  2. The method is tested across three MoE backbones, multiple attack families, helpfulness benchmarks, held-out attacks, and several ablations. The finding that suppression works while symmetric amplification causes routing collapse is particularly informative.

Critical weaknesses

The main evaluation focuses on selected single-turn jailbreak benchmarks. It does not demonstrate comparable robustness to adaptive, multi-turn, multilingual, or distributionally broader attacks. The paper’s claims should therefore be scoped more narrowly.

Questions for authors

Why were the six reported attack families selected?

Reviewer_Z393OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PIRA, a training-free inference-time safety intervention for MoE LLMs. It identifies Risk-Amplifying Experts for each query using gradient-based Per-Query Identification and adjusts their suppression strength through Query-Adaptive Bias Magnitude. Experiments on multiple MoE models show improved safety performance over existing baselines while largely preserving benign response capability.

Critical strengths

1. Training-free intervention: PIRA directly modifies router biases without updating model weights, making it practical to apply to existing MoE models.

2. Query-adaptive intervention: PIRA dynamically identifies and suppresses different experts based on each input, showing promising safety improvements while maintaining relatively high XSTest answer rates.

Critical weaknesses

1. Limited Novelty: Primarily a Combination of Existing Techniques: PIRA uses an input-risk direction and an output-safety direction in the offline stage, and then performs per-query identification and modulation (query-adaptive bias magnitude) in the online stage to dynamically intervene on experts. These designs, however, are not entirely new. Using hidden-state directions to identify a model’s refusal tendency, performing attribution via gradients, and dynamically adjusting steering strength based on the input have all been studied in related work (Refusal in Language Models Is Mediated by a Single Direction, NeurIPS 2024; GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs, ACL 2026; AdaSteer, EMNLP 2025). The main contribution of this paper is more about combining these existing ideas and applying them at the MoE router level.

2. Potentially unfair comparison: Although the authors reimplement SteerMoE and compare it with PIRA under the same experimental settings, the overall evaluation scope is much narrower than that of the original SteerMoE paper: (1) Benchmarks reduced from 4 to 1: the original SteerMoE paper uses four safety benchmarks, including TDC2023, MaliciousInstruct, AdvBench, and StrongREJECT, and also tests cross-lingual generalization with MultiJail; in contrast, the main safety experiments in this paper focus on a single benchmark, PandaGuard JBB, and only select six out of 25 attack variants. Although the authors add SORRY-Bench, it is only tested on OLMoE. (2) Models reduced from 6 to 4: the original SteerMoE paper evaluates six MoE models, while the main experiments here include only three models, with one additional Qwen3 model added in the appendix. As a result, the current results mainly show that PIRA outperforms SteerMoE under the experimental settings chosen by the authors, but they are not sufficient to show that this advantage generalizes to more safety benchmarks and different MoE architectures.

Questions for authors

1. Insufficient baseline comparison: The paper compares against only one baseline, SteerMoE. The authors claim that SteerMoE is the only published training-free method that intervenes at the router. But given how few baselines there are, the authors could do two things. First, add a comparison with unpublished work on the same technical route, such as MASCing: Configurable Mixture-of-Experts Behavior via Activation Steering Masks. Second, move the methods from other technical routes that are currently in the appendix into the main experiments and compare them thoroughly, so that the paper can show exactly where its method has advantages over those alternative routes.

2. Insufficient Ablation of the Modulation Components: The ablation in Table 2 sets both ρ(λ(q))\rho(\lambda(q)) and wℓ,e(q)w_{\ell,e}(q) to 1, which removes both the risk-related bias modulation and the expert-relative attribution weight at the same time. This makes it impossible to tell which component provides the main benefit, or whether combining the two is actually more effective than using either one alone. I suggest the authors remove these two factors separately, so that the actual role of each part of the modulation design can be clarified more clearly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.