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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace single global routing with head-wise routing (MH-MoE) | LoRAMoE · TRACE, Qwen3-0.6B | OP: 37.8 -> 46.7 | Improved |
| replace single global routing with head-wise routing (MH-MoE) | LoRAMoE · TRACE, Qwen3-0.6B | BWT: -11.2 -> -4.5 | Improved |
| replace single global routing with head-wise routing (MH-MoE) | LoRAMoE · TRACE, Qwen3-8B | OP: 55.1 -> 56.9 | Improved |
| replace single global routing with head-wise routing (MH-MoE) | LoRAMoE · TRACE, Qwen3-8B | BWT: -5.5 -> -5.1 | Improved |
| use head-wise routing at matched route-space size and parameters (ablation) | LoRAMoE (K=26, top-5) · TRACE, Qwen3-0.6B | BWT: -9.3 -> -4.5 | Improved |
| use head-wise routing at matched route-space size and parameters (ablation) | LoRAMoE (K=26, top-5) · TRACE, Qwen3-0.6B | OP: 43.1 -> 46.7 | Improved |
| use head-wise routing under task order 1 (ablation) | LoRAMoE · TRACE order 1 | BWT: -21.4 -> -8.0 | Improved |
| use head-wise routing under task order 2 (ablation) | LoRAMoE · TRACE order 2 | BWT: -18.3 -> -9.6 | Improved |
| use head-wise routing under task order 3 (ablation) | LoRAMoE · TRACE order 3 | BWT: -17.0 -> -10.9 | Improved |
| increase number of routing heads from 2 to 4 (ablation) | MH-MoE · TRACE | OP: 53.1 -> 52.6 | Hurt |
| increase number of routing heads from 4 to 16 (ablation) | MH-MoE · TRACE | OP: 52.6 -> 56.9 | Improved |
| use head-wise routing instead of single routing | LoRAMoE · Qwen3-8B training speed | tokens/s: 3215.4 -> 3184.7 | Hurt |
Closest papers by meaning
- Mixture-of-Top-$k$ Attention: Efficient Attention as Scalable Fast Weights (NeurIPS 2026 · similarity 0.89)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.88)
- MoH: Multi-Head Attention as Mixture-of-Head Attention (ICML 2025 · similarity 0.88)
- SwitchHead: Accelerating Transformers with Mixture-of-Experts Attention (NeurIPS 2024 · similarity 0.88)
- Dense Backpropagation Improves Training for Sparse Mixture-of-Experts (NeurIPS 2025 · similarity 0.87)
- Theory on Mixture-of-Experts in Continual Learning (ICLR 2025 · similarity 0.86)
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.
