Research map / Architectures and efficiency / Pruning and efficient tuning
PYRA: Parallel Yielding Re-Activation for Training-Inference Efficient Task Adaptation
Yizhe Xiong, Hui Chen, Tianxiang Hao, Zijia Lin, Jungong Han, Yuesong Zhang, Guoxin Wang, Yongjun Bao, Guiguang Ding
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 |
|---|---|---|---|
| add PYRA token modulation to ToMe + LoRA | ToMe + LoRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 70.43 -> 72.06 | Improved |
| add PYRA token modulation to ToMe + LoRA | ToMe + LoRA · VTAB-1k high compression, ViT-L/16 | average accuracy: 74.10 -> 75.66 | Improved |
| add PYRA token modulation at low compression | ToMe + LoRA · VTAB-1k low compression, ViT-B/16 | average accuracy: 74.10 -> 74.69 | Improved |
| add the W_r generator (tokens) without activation (ablation) | PYRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 70.43 -> 70.54 | Improved |
| add both W_r and W_D generators without activation (ablation) | PYRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 70.43 -> 71.32 | Improved |
| add sigmoid re-activation on top of both generators (ablation) | PYRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 71.32 -> 72.06 | Improved |
| use parallel yielding generators instead of a directly learned W_{D x r} (ablation) | PYRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 71.49 -> 72.06 | Improved |
| use PYRA generators instead of a gated MLP generator (ablation) | PYRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 71.06 -> 72.06 | Improved |
| increase LoRA hidden dimension to 4x parameters instead of adding PYRA (ablation) | ToMe + LoRA baseline · VTAB-1k high compression, ViT-B/16 | average accuracy: 70.43 -> 71.00 (PYRA: 72.06) | Improved |
| increase modulation generator rank above 1 (ablation) | PYRA · VTAB-1k high compression, ViT-B/16 | average accuracy | No clear effect |
| one-stage training with token merging instead of two-stage (ablation) | ToMe + LoRA · VTAB-1k high compression, ViT-B/16 | average accuracy: 69.25 -> 70.43 | Improved |
| parameter-efficient distillation instead of direct PEFT on smaller backbone | ViT-S/16 with LoRA · VTAB-1k high compression | average accuracy: 71.85 -> 70.99 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DiffRate — DiffRate : Differentiable Compression Rate for Efficient Vision Transformers (ICCV 2023) · better on 2 of 2 shared results
- SPViT — SPViT: Enabling Faster Vision Transformers via Latency-Aware Soft Token Pruning (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · similarity 0.86)
- APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference (ICML 2024 · similarity 0.86)
- RECAST: Reparameterized, Compact weight Adaptation for Sequential Tasks (ICLR 2025 · similarity 0.86)
- SeedLoRA: A Fusion Approach to Efficient LLM Fine-Tuning (ICML 2025 · similarity 0.86)
- RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models (NeurIPS 2025 · similarity 0.86)
- CorDA: Context-Oriented Decomposition Adaptation of Large Language Models for Task-Aware Parameter-Efficient Fine-tuning (NeurIPS 2024 · similarity 0.85)
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.
