AI Paper Atlas

Research map / Continual learning / Continual fine-tuning of LLMs

Multi-Head Attention as a Source of Catastrophic Forgetting in MoE Transformers

Anrui Chen, Ruijun Huang, Xin Zhang, Fang Dong, Hengjie Cao, Zhendong Huang, Yifeng Yang, Mengyi Chen, Jixian Zhou, Mingzhi Dong, Yujiang Wang, Jinlong Hou and 6 more

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation · 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 single global routing with head-wise routing (MH-MoE)LoRAMoE · TRACE, Qwen3-0.6BOP: 37.8 -> 46.7Improved
replace single global routing with head-wise routing (MH-MoE)LoRAMoE · TRACE, Qwen3-0.6BBWT: -11.2 -> -4.5Improved
replace single global routing with head-wise routing (MH-MoE)LoRAMoE · TRACE, Qwen3-8BOP: 55.1 -> 56.9Improved
replace single global routing with head-wise routing (MH-MoE)LoRAMoE · TRACE, Qwen3-8BBWT: -5.5 -> -5.1Improved
use head-wise routing at matched route-space size and parameters (ablation)LoRAMoE (K=26, top-5) · TRACE, Qwen3-0.6BBWT: -9.3 -> -4.5Improved
use head-wise routing at matched route-space size and parameters (ablation)LoRAMoE (K=26, top-5) · TRACE, Qwen3-0.6BOP: 43.1 -> 46.7Improved
use head-wise routing under task order 1 (ablation)LoRAMoE · TRACE order 1BWT: -21.4 -> -8.0Improved
use head-wise routing under task order 2 (ablation)LoRAMoE · TRACE order 2BWT: -18.3 -> -9.6Improved
use head-wise routing under task order 3 (ablation)LoRAMoE · TRACE order 3BWT: -17.0 -> -10.9Improved
increase number of routing heads from 2 to 4 (ablation)MH-MoE · TRACEOP: 53.1 -> 52.6Hurt
increase number of routing heads from 4 to 16 (ablation)MH-MoE · TRACEOP: 52.6 -> 56.9Improved
use head-wise routing instead of single routingLoRAMoE · Qwen3-8B training speedtokens/s: 3215.4 -> 3184.7Hurt

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