Research map / Architectures and efficiency / Pruning and efficient tuning
SHERL: Synthesizing High Accuracy and Efficient Memory for Resource-Limited Transfer Learning
Haiwen Diao, Bo Wan, Xu Jia, Yunzhi Zhuge, Ying Zhang, Huchuan Lu, Long Chen
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 |
|---|---|---|---|
| use SHERL instead of LST on T5-base GLUE | T5-base · GLUE (T5-base) | training memory (GB): 5.5 -> 2.9 | Improved |
| use SHERL instead of LST on T5-large | T5-large · GLUE (T5-large) | average score: 87.1 -> 87.5 | Improved |
| use SHERL instead of LST on T5-3B | T5-3B · GLUE (T5-3B) | average score: 88.1 -> 88.2 | No clear effect |
| use SHERL instead of UniPT on MSVD video-text retrieval | CLIP4Clip · MSVD | Rsum: 432.1 -> 429.7 | Hurt |
| use SHERL instead of UniPT on Flickr30K retrieval | VSEinf · Flickr30K | Rsum: 537.4 -> 542.3 | Improved |
| replace redundancy-aware aggregation with max pooling, average pooling, MHSA or plain linear attention (ablation) | SHERL early route · VL datasets (average accuracy) | average accuracy | Improved |
| add late regulation by the last pre-trained layer (ablation) | SHERL · GLUE (T5-base) | average accuracy | Improved |
| add Prompt tuning in late layers on top of SHERL (ablation) | SHERL · GLUE (T5-base) | average accuracy | No clear effect |
| add LoRA or Adapter in late frozen layers on top of SHERL (ablation) | SHERL · GLUE (T5-base) | average accuracy | Improved |
| insert blended features into multiple late decoder layers instead of a single earlier layer (ablation) | SHERL · GLUE (T5-base) | average accuracy | Improved |
| use SHERL instead of Fully-FT | T5-base · GLUE (T5-base) | average score: 85.2 -> 84.3 | Hurt |
Later papers that compared with this method
2 later accepted papers report SHERL as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Memory-Efficient Transfer Learning with Fading Side Networks via Masked Dual Path Distillation (CVPR 2026 · better on 7 of 7 shared results)
- MP-ISMoE: Mixed-Precision Interactive Side Mixture-of-Experts for Efficient Transfer Learning (AAAI 2026 · better on 5 of 5 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- UniPT — UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 2024) · better on 5 of 6 shared results
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
- UniPT: Universal Parallel Tuning for Transfer Learning with Efficient Parameter and Memory (CVPR 2024 · similarity 0.86)
- A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models (CVPR 2024 · similarity 0.85)
- Provable Sample-Efficient Transfer Learning Conditional Diffusion Models via Representation Learning (NeurIPS 2025 · similarity 0.85)
- DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts (NeurIPS 2026 · similarity 0.85)
- DTL: Disentangled Transfer Learning for Visual Recognition (AAAI 2024 · similarity 0.85)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.84)
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.
