Research map / Architectures and efficiency / Pruning and efficient tuning
RUTaL: Residual Upcycling with Task Ladder for Efficient Multi-Task Learning
Haiming Yao, Wei Luo, Qiyu Chen, Jianxing Liao, Wei You
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 TSLA task-specific ladder branch (ablation) | RUTaL (TGRU) · PASCAL-Context (Swin-Tiny) | Delta m (%): +2.42% -> +7.79% | Improved |
| add TGRU residual upcycling (ablation) | RUTaL (TSLA) · PASCAL-Context (Swin-Tiny) | Delta m (%): +6.75% -> +7.79% | Improved |
| use low-rank residual experts instead of fully fine-tuned MoE experts (ablation) | TGRU · PASCAL-Context (Swin-Tiny) | Delta m (%): +4.72% -> +7.60% | Improved |
| reparameterized aggregation with shared up-projection instead of per-expert residuals (ablation) | TGRU · PASCAL-Context (Swin-Tiny) | Delta m (%): +7.60% -> +7.79% | Improved |
| use both projector with cross-attention and task modulation in TAS/LAE (ablation) | TSLA · PASCAL-Context (Swin-Tiny) | Delta m (%): +6.71% / +3.01% -> +7.79% | Improved |
| increase experts from 1 to 4 (ablation) | RUTaL · PASCAL-Context (Swin-Tiny) | Delta m (%): +7.61% -> +7.79% | Improved |
| increase experts from 4 to 64 (ablation) | RUTaL · PASCAL-Context (Swin-Tiny) | Delta m (%): +7.79% -> +8.07% | No clear effect |
| increase rank r from 8 to 16 (ablation) | RUTaL · PASCAL-Context (Swin-Tiny) | Delta m (%): +7.79% -> +7.71% | No clear effect |
| increase projection dim from D/4 to D (ablation) | TSLA · PASCAL-Context (Swin-Tiny) | Delta m (%): +7.79% -> +7.82% | No clear effect |
| scale Swin backbone from Tiny to Large (ablation) | RUTaL · PASCAL-Context (Swin-Tiny) | Delta m (%): +7.79% -> +12.35% | Improved |
| use RUTaL instead of TADFormer on Swin-Tiny | multi-task PEFT · PASCAL-Context (Swin-Tiny) | Delta m (%): +4.24% -> +7.79% | Improved |
| use RUTaL instead of TADFormer on NYUDv2 | multi-task PEFT · NYUDv2 (Swin-Tiny) | SemSeg mIoU: 37.30 -> 42.76 | Improved |
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.
- VMT-Adapter — VMT-Adapter: Parameter-Efficient Transfer Learning for Multi-Task Dense Scene Understanding (AAAI 2024) · better on 3 of 3 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
- Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learner (NeurIPS 2024 · similarity 0.90)
- TADFormer: Task-Adaptive Dynamic TransFormer for Efficient Multi-Task Learning (CVPR 2025 · similarity 0.88)
- DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations (CVPR 2025 · similarity 0.85)
- MTLoRA: Low-Rank Adaptation Approach for Efficient Multi-Task Learning (CVPR 2024 · similarity 0.85)
- PELA: Learning Parameter-Efficient Models with Low-Rank Approximation (CVPR 2024 · similarity 0.85)
- Low-Rank Rescaled Vision Transformer Fine-Tuning: A Residual Design Approach (CVPR 2024 · 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.
