Research map / Architectures and efficiency / Pruning and efficient tuning
Multi-Head Adapter Routing for Cross-Task Generalization
Lucas Caccia, Edoardo Ponti, Zhan Su, Matheus Pereira, Nicolas Le Roux, Alessandro Sordoni
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-head routing Poly with multi-head routing Poly-S | Poly · T0 few-shot (T5-XL) | avg test accuracy: 68.0 -> 69.1 | Improved |
| route only the multi-head routing function (Poly-S-z) vs Poly-z | Poly-z · T0 few-shot (T5-XL) | avg test accuracy: 66.4 -> 68.3 | Improved |
| replace learned routing with random routing (Random-mu) (ablation) | Poly · T0 (T5-XL) | test accuracy: 68.0 -> 66.0 | Hurt |
| use Poly-S instead of Poly at 11B backbone | Poly · T0-11B few-shot | avg test accuracy: 74.9 -> 74.7 | No clear effect |
| routing-only tuning on 11B: Poly-S-z vs Poly-z | Poly-z · T0-11B few-shot | avg test accuracy: 70.0 -> 72.9 | Improved |
| Poly-S vs LoRA on SuperNI | LoRA · SuperNI | Rouge-L: 67.6 -> 68.5 | Improved |
| Poly-S vs Poly on SuperNI | Poly · SuperNI | Rouge-L: 67.8 -> 68.5 | Improved |
| routing-only tuning on SuperNI: Poly-S-z vs Poly-z | Poly-z · SuperNI | Rouge-L: 64.6 -> 68.0 | Improved |
| larger LoRA rank (LoRA-big) instead of LoRA | LoRA · SuperNI | Rouge-L: 67.6 -> 67.2 | No clear effect |
| average pre-trained modules (Poly-mu) instead of learned routing at adaptation (ablation) | Poly · T0 (T5-XL) | test accuracy: 68.0 -> 67.8 | No clear effect |
| Poly-mu averaging instead of shared LoRA adapter | LoRA · T0 (T5-XL) | test accuracy: 66.0 -> 67.8 | Improved |
| zero-shot: fine-tune averaged MHR-mu adapters for k=1000 steps vs k=0 (ablation) | MHR-mu · T0 zero-shot (T5-XL-LM) | avg accuracy: 48.0 -> 58.0 | Improved |
Closest papers by meaning
- PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization (NeurIPS 2024 · similarity 0.88)
- Selective Task Group Updates for Multi-Task Optimization (ICLR 2025 · similarity 0.86)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 2025 · similarity 0.86)
- Towards Modular LLMs by Building and Reusing a Library of LoRAs (ICML 2024 · similarity 0.86)
- Towards Few-Shot Adaptation of Foundation Models via Multitask Finetuning (ICLR 2024 · similarity 0.86)
- Parameter-Efficient Multi-Task Model Fusion with Partial Linearization (ICLR 2024 · 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.
