简体中文 · English

RARE: Editing MoE representations without changing expert routing¶

Fidelity: core-mechanism reproduction. Runs router row-space SVD, null-space projection, routing consistency checks, and downstream correction without loading the paper's six large MoE checkpoints.

Paper information¶

Field	Details
Paper	arXiv 2608.21236
Institution	Huazhong University of Science and Technology
First public release	2026-08-21 (arXiv v1)
Author code	No official or author-released code found in the paper (checked 2026-08-24).
Local adapter or method key	`rare`
Local reproduction code	`src/auto_research/reproductions/rare/`

Original paper summary¶

Background and main changes¶

Applying dense-LLM activation steering directly to MoE changes router logits. Sending tokens to different experts can invalidate the estimated behavior direction. RARE projects an arbitrary steering direction into the router's null space, then removes router-visible components introduced during propagation at later protected layers. This changes behavior representations without changing the original expert paths.

flowchart LR
  C["Positive / negative contrastive activations"] --> U["MeanDiff / Probe / LDA or other direction"]
  R["MoE router matrix"] --> Q["Row-space SVD"]
  U --> P["Null-space projection"]
  Q --> P
  P --> E["Edit hidden state"]
  E --> K["Runtime correction at later layers"]
  K --> O["Change behavior while retaining expert routes"]

Key figure from the paper¶

Figure 3: The core architecture, its main modules, and their connections. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

Let \(Q_l\) be an orthonormal basis for the row space of router \(R_l\):

\[ \Pi_l^\perp=I-Q_lQ_l^\top,\qquad h'_l=h_l+\Pi_l^\perp u_l,\quad R_l(h'_l-h_l)=0. \]

Later protected layers correct propagation drift using clean activations:

\[ \widetilde h_l=h_l^0+\Pi_l^\perp(h'_l-h_l^0),\qquad R_l\widetilde h_l=R_lh_l^0. \]

Offline and online results¶

Across six open MoE models, harmfulness steering averages 53.3% ASR while retaining 67.8% on MMLU. TruthfulQA MC1 rises from 41.0% to 58.6%, and CounterFact efficacy from 16.8% to 96.3%. Foundation-model papers are not subject to an industrial online A/B requirement.

Local reproduction¶

Local comparison: The baseline is the same steering direction without projection. The experimental arm adds only RARE's null-space projection and analytical correction. Route agreement increases from 0.8237 to 1.0000 (+21.40% relative), while steering gain remains 1.75 in both arms.

See deterministic mini-suite metrics for 2048 states, 64 dimensions, and 8 experts.

auto-research reproduce --paper rare --dataset-dir data --seed 42

Scope¶

DeepSeek-V2-Lite, Mixtral, Phi, Qwen3, and GPT-OSS were not loaded; JailbreakBench, TruthfulQA, and CounterFact were not run. The local run verifies the exact routing invariance in equation (8), not the paper's downstream results.