💬 NLP

RARE: Decoupling Representation Steering from Expert Routing in Mixture-of-Experts Language Models

The paper introduces RARE, a router-agnostic framework that decouples representation steering from expert routing in Mixture-of-Experts models by projecting behavioral perturbations onto the router's null space, thereby significantly improving steering effectiveness for harmfulness, truthfulness, and factual editing while preserving model utility.

Original authors: Zhibo Zhang, Zhen Ouyang, Ling Shi, Kailong Wang

Published 2026-08-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhibo Zhang, Zhen Ouyang, Ling Shi, Kailong Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Large language models are the engines behind modern artificial intelligence, capable of writing stories, solving problems, and answering questions. To make these massive systems efficient, many are now built using a design called a mixture of experts. Imagine a large office where a manager receives every incoming request and decides which specific specialist should handle it. In these models, the "manager" is a routing mechanism that looks at each word of a sentence and sends it to a small, specialized group of internal processors, or "experts," rather than using the entire brain of the computer for every single task. This allows the model to be incredibly large and smart while remaining fast and economical to run.

Researchers have long sought ways to guide or "steer" these models toward specific behaviors, such as making them more truthful or less likely to generate harmful content. A popular technique involves nudging the internal mathematical states of the model during the thinking process, a method that works well for older, simpler models where every part of the brain is always active. However, when scientists tried to apply this same nudging technique to the newer mixture-of-experts models, it often failed. The problem was that the nudge intended to change the behavior also accidentally confused the manager, causing it to send the request to the wrong specialists. This mismatch disrupted the model's natural flow, leading to poor results.

A team of researchers has now solved this puzzle by developing a new framework called RARE. Their work reveals a crucial insight: the manager in these models is primarily concerned with the topic of the request, not the specific behavior the model is expected to show. Whether a user asks a question about coding or asks the model to refuse a harmful request, the manager tends to send the request to the same set of specialists if the topic remains the same. The researchers found that the old methods failed because they tried to change the behavior by altering the path the data took, which confused the manager. Instead, their new approach changes the behavior without changing the path.

The RARE framework works by carefully filtering the "nudge" before it is applied. It removes any part of the instruction that would make the manager change its mind about which specialists to use. By doing this, the model continues to send the request to the correct experts, but those experts then process the information in a way that produces the desired behavior. The researchers tested this method on six different large models and across three distinct tasks: making the models less harmful, making them more truthful, and updating specific facts they know.

The results were striking. When trying to stop the models from following harmful instructions, the new method succeeded in 53.3% of cases, a significant improvement over previous attempts, while still keeping the models 67.8% accurate on general knowledge tests. In tests of truthfulness, the method improved the models' ability to give correct answers from 41.0% to 58.6%. Perhaps most dramatically, when asked to update a specific fact, the success rate jumped from 16.8% to 96.3%. These numbers suggest that by respecting the model's internal routing system, researchers can effectively guide its behavior without breaking its underlying intelligence.

The study also compared five different ways of calculating the necessary nudge to find which one worked best. They found that a method based on analyzing the shape and spread of the data, rather than just its average direction, provided the most consistent and powerful results across all the different models. This finding suggests that the geometry of the internal data matters just as much as the direction of the change.

Ultimately, this research demonstrates that controlling the behavior of modern, complex AI models requires a delicate balance. You cannot simply force a change; you must work within the model's existing architecture. By preserving the natural path that the model chooses for a given topic, it is possible to steer its output toward safety, truth, or accuracy. This approach offers a reliable way to adapt these powerful tools for specific needs without compromising their general capabilities, providing a clearer path forward for making artificial intelligence more aligned with human values.

Technical Summary: RARE – Decoupling Representation Steering from Expert Routing in MoE Models

Problem Statement

Large Language Models (LLMs) are increasingly adopting Mixture-of-Experts (MoE) architectures to scale capacity while maintaining inference efficiency. However, applying Representation Engineering—a technique that controls model behavior by modifying intermediate hidden states—to MoE models introduces a critical structural mismatch.

In dense models, steering vectors are added to hidden states to shift behavior. In MoE models, the router determines which experts process a token based on these hidden states. Directly applying steering perturbations alters the router logits, causing tokens to be routed to different expert sets than they would be in a clean inference pass. This "routing drift" disrupts the model's native conditional computation path, leading to a significant degradation in steering effectiveness. While recent MoE-specific methods attempt to control behavior by directly manipulating router logits or expert activation, these approaches override the model's input-conditioned computation path, potentially degrading response relevance and general capabilities.

The core challenge is how to coordinate representation-level control with the conditional routing mechanism without disrupting the native expert allocation.

Methodology: RARE

The authors introduce RARE (Router-Agnostic Representation Engineering), a framework designed to steer MoE models while preserving their native routing patterns. The methodology consists of three main components:

1. Empirical Motivation

Through controlled probes, the authors established that MoE routing is primarily sensitive to the semantic content of the input query rather than the specific response behavior (e.g., refusal vs. compliance) or safety intent.

Finding: Under a fixed query topic, changing the response behavior (e.g., from compliant to refusing) results in minimal changes to the router's expert selection.
Implication: Effective steering should modify the hidden representations processed by the naturally selected experts without altering the router's decision.

2. Router-Agnostic Projection

RARE decouples behavioral steering from expert routing by projecting perturbations onto the null space of the router matrix.

Null Space Projection: For a router weight matrix $R_\ell$ at layer $\ell$ , the router computes logits $z_\ell = R_\ell h_\ell$ . To ensure the router logits remain unchanged ( $R_\ell \delta h_\ell = 0$ ), the perturbation $\delta h_\ell$ must lie in the null space of $R_\ell$ .
Mechanism: Given a raw perturbation vector $u_\ell$ (derived from contrastive hidden states), RARE projects it using a projector $\Pi^\perp_\ell = I - Q_\ell Q_\ell^\top$ , where $Q_\ell$ is an orthonormal basis for the row space of $R_\ell$ . The resulting perturbation is invisible to the router, preserving the original expert selection.

3. Downstream Routing Correction

Because perturbations propagate through nonlinear transformer blocks, local invariance at the injection layer does not guarantee invariance at subsequent routers.

Runtime Correction: RARE applies a correction at selected downstream "protected" layers. It compares the current edited input $h'_\ell$ against a reference clean input $h^0_\ell$ and removes any router-visible deviation, ensuring the router logits at these layers remain consistent with the clean pass.

Perturbation Estimators

The framework is agnostic to how the raw perturbation is constructed. The authors evaluated five estimators:

MeanDiff: First-order translation between class centers.
Probe: Linear boundary normal.
LowRank: Projection onto a principal subspace.
LDA: Linear Discriminant Analysis accounting for within-class covariance.
AffineGaussian: Second-order alignment using whitening-recoloring principles.

Key Contributions

Empirical Insight: The paper provides evidence that MoE routing tracks query semantics more closely than response behaviors, motivating the preservation of native routing as a prerequisite for effective steering.
RARE Framework: A novel, router-agnostic steering method that projects behavioral perturbations onto router-invisible subspaces and corrects propagated routing drift, enabling representation editing without disrupting expert allocation.
Comprehensive Evaluation: A systematic study of five perturbation estimators across six heterogeneous open-weight MoE models (DeepSeek, Mixtral, Phi-3.5, Phi-mini, Qwen3, GPT-oss) and three behavioral control tasks.

Experimental Results

The evaluation covered three scenarios: Harmfulness Steering (jailbreaking), Truthfulness Steering, and Factual Editing.

Harmfulness Steering: RARE (using the AffineGaussian estimator) achieved an average Attack Success Rate (ASR) of 53.3% while retaining 67.8% MMLU accuracy. This represents a significantly stronger efficacy-utility trade-off compared to baselines like RepE (18.7% ASR) and MoE-specific methods like SteerMoE (46.3% ASR but lower utility retention).
Truthfulness Steering: RARE improved average TruthfulQA MC1 accuracy from a baseline of 41.0% to 58.6%.
Factual Editing: RARE increased CounterFact efficacy from 16.8% to 96.3%.
Estimator Performance: AffineGaussian consistently yielded the strongest steering performance across all models and tasks. LDA offered a better trade-off when considering locality (preserving neighboring knowledge).
Ablation: Removing the router-agnostic projection pipeline reduced the average ASR from 53.3% to 14.6%, confirming that preserving routing is critical for performance.

Significance and Claims

The paper claims that routing consistency is a fundamental architectural consideration for adapting representation engineering to MoE models. By decoupling behavioral control from expert routing, RARE demonstrates that it is possible to modify internal representations to alter model behavior while maintaining the model's native computation path and general capabilities.

The authors position RARE not as a method to override the model's safety or logic, but as a lightweight, efficient mechanism to steer behavior that respects the underlying MoE architecture. The results suggest that future representation engineering for MoE models must account for the router's sensitivity to semantic content to avoid the failure modes caused by routing drift.