Research map / Architectures and efficiency / Pruning and efficient tuning
TaskIT: Memory-Efficient Fine-Tuning of Multi-LoRA LLMs via Cross-Task Importance Transfer
Cheng Fang, Zimu Zhou, Ke Ma, Bin Guo
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 task similarity assessment instead of uniform task similarities (ablation) | TaskIT · image tasks | accuracy: 75.3 -> 80.4 | Improved |
| add module importance predictor instead of uniform importances (ablation) | TaskIT · image tasks | accuracy: 74.1 -> 80.4 | Improved |
| add block-based memory predictor instead of uniform memory model (ablation) | TaskIT · image tasks | peak memory (GB): 9.9 -> 7.8 | Improved |
| use DP scheduler instead of greedy (ablation) | TaskIT · image tasks | accuracy: 77.5 -> 80.4 | Improved |
| fine-tune with TaskIT instead of best Zero-FT (AdapterFusion) | multi-LoRA Janus-Pro · cross-modal tasks | avg accuracy: +10.4% | Improved |
| TaskIT instead of best non-LoRA baseline UniPT | multi-LoRA Janus-Pro · cross-modal tasks | memory and accuracy: -13.8% memory, +1.5% accuracy | Improved |
| TaskIT instead of best LoRA baseline AutoLoRA | multi-LoRA Janus-Pro · cross-modal tasks | memory and accuracy: -45.5% memory, -0.3% accuracy | Mixed |
| TaskIT versus UniPT on uni-modal image tasks | multi-LoRA Janus-Pro · VTAB | avg accuracy: +2.1% accuracy, -11.4% memory | Improved |
| smaller memory budget from 12 GB to 8 GB in continual learning (ablation) | TaskIT · continual task sequence | accuracy gain vs LoRA | Hurt |
| fall back to uniform importance when no task has similarity above 0.35 | TaskIT · audio captioning (t9) | importance transfer | No clear effect |
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 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
- Sparse High Rank Adapters (NeurIPS 2024 · similarity 0.86)
- RandLoRA: Full rank parameter-efficient fine-tuning of large models (ICLR 2025 · similarity 0.86)
- SeedLoRA: A Fusion Approach to Efficient LLM Fine-Tuning (ICML 2025 · similarity 0.86)
- MeteoRA: Multiple-tasks Embedded LoRA for Large Language Models (ICLR 2025 · similarity 0.86)
- Transformer-Squared: Self-adaptive LLMs (ICLR 2025 · similarity 0.85)
- Efficient Multi-task LLM Quantization and Serving for Multiple LoRA Adapters (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.
