AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Eigenvectors of Experts are Training-free Non-collapsing Routers

Giang Do, Hung Le, Truyen Tran

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: MoE LLM compression · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace learned router with SSMoE eigenvector routing, with 25% experts droppedGPT-OSS-20B · 10-shot 8-benchmark averageaverage accuracy (%): 48.8 -> 54.1 vs originalImproved
replace learned router with SSMoE eigenvector routing, with 25% experts droppedGPT-OSS-120B · 10-shot 8-benchmark averageaverage accuracy (%): 52.6 -> 54.9 vs originalImproved
use SSMoE to choose which experts to drop instead of random drop (ablation)GPT-OSS-20B · 10-shot 8-benchmark averageaverage accuracy (%): 44.9 -> 54.1Improved
keep all experts with SSMoE routing (SSMoE-FULL) (ablation)GPT-OSS-20B · 10-shot 8-benchmark averageaverage accuracy (%): 54.1 -> 55.2Improved
SSMoE routing on GSM8K with 25% experts droppedGPT-OSS-20B · 10-shot GSM8Kaccuracy (%): 36.8 -> 30.8Hurt
use eigenvector-only (EV) representation in place of learned router embeddings (ablation)OLMoE-7B · MTEB clean averageaverage score: 29.9 -> 40.9Improved
add router mixing to the EV representation (SSMoE vs EV) (ablation)OLMoE-7B · MTEB clean averageaverage score: 40.9 -> 39.7No clear effect
use SSMoE instead of MoEE embeddings on Qwen-MoE-7BMoEE · MTEB clean averageaverage score: 42.6 -> 45.8Improved
use SSMoE under corrupted inputsSMoE / MoEE · MTEB corrupt averageaverage score: 21.2 -> 24.8 (OLMoE vs MoEE)Improved
apply SSMoE routing to CLIP-MoE for zero-shot retrievalCLIP-MoE · COCO I2T Recall@1 cleanRecall@1: 65.0 -> 65.7Improved
apply SSMoE routing to CLIP-MoE with noisy imagesCLIP-MoE · COCO I2T Recall@1 corruptRecall@1: 36.2 -> 37.2Improved
use averaged eigenvector descriptors instead of router-guided selection (ablation)SSMoE · GPT-OSS-20B 10-shot averageaverage accuracy (%): 54.1 -> 52.7Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy