Research map / Continual learning / Continual fine-tuning of LLMs
PCLR: Progressively Compressed LoRA for Multimodal Continual Instruction Tuning
Weicheng Meng, Jingyang Qiao, Zhizhong Zhang, Shaohui Liu, Yuan Xie
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 LoRA Rank Pool architecture with compression and learning (ablation) | LoRA · CoIN LLaVA-1.5-7B | Avg.ACC / Forgetting: Avg.ACC +25.92, Forgetting -26 | Improved |
| add integration process (distillation after compression) (ablation) | PCLR · CoIN LLaVA-1.5-7B | Avg.ACC / Forgetting: 54.66 -> 60.78 Avg.ACC; 11.29 -> 5.09 Forgetting | Improved |
| add progressive learning process (ablation) | PCLR · CoIN LLaVA-1.5-7B | Avg.ACC / Forgetting: 60.78 -> 62.19 Avg.ACC; 5.09 -> 3.39 Forgetting | Improved |
| use progressive compression instead of aggressive compression (ablation) | PCLR · CoIN LLaVA-1.5-7B | Forgetting: 5.09 -> 3.39 | Improved |
| use progressive compression instead of conservative compression (ablation) | PCLR · CoIN LLaVA-1.5-7B | New.ACC: 63.43 -> 65.16 | Improved |
| use progressive compression instead of reverse progressive (ablation) | PCLR · CoIN LLaVA-1.5-7B | Avg.ACC: 61.26 -> 62.19 | Improved |
| use progressive compression instead of centralized compression (ablation) | PCLR · CoIN LLaVA-1.5-7B | Avg.ACC: 61.00 -> 62.19 | Improved |
| increase integration data usage from none to 5k samples (ablation) | PCLR · CoIN LLaVA-1.5-13B | Avg.ACC / Forgetting: 59.40 -> 64.50 Avg.ACC; 8.86 -> 3.11 Forgetting | Improved |
| use full dataset instead of 5k for integration (ablation) | PCLR · CoIN LLaVA-1.5-13B | Avg.ACC / Forgetting: 64.50 -> 65.51 Avg.ACC; 3.11 -> 2.08 Forgetting; 11 min -> 115 min cost | Improved |
| combine PCLR with LwF within task groups (ablation) | PCLR · Continual-NExT LLaVA-1.5-hf | Avg.ACC / Forgetting: 47.94 -> 52.67 Avg.ACC; 7.71 -> 4.58 Forgetting | Improved |
| change instruction templates (Diverse, 10Type) (ablation) | PCLR · CoIN LLaVA-1.5-7B | Avg.ACC: 62.19 -> 62.00 and 61.98 | No clear effect |
| change task order (Reverse, Alphabet) (ablation) | PCLR · CoIN LLaVA-1.5-7B | Avg.ACC: 62.19 -> 62.18 and 60.62 | 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.
- SEFE — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 2 of 2 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
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.88)
- Large Continual Instruction Assistant (ICML 2025 · similarity 0.88)
- SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025 · similarity 0.88)
- SAFE: Slow and Fast Parameter-Efficient Tuning for Continual Learning with Pre-Trained Models (NeurIPS 2024 · similarity 0.88)
- CoIN: A Benchmark of Continual Instruction Tuning for Multimodel Large Language Models (NeurIPS 2024 · similarity 0.87)
- Gated Integration of Low-Rank Adaptation for Continual Learning of Large Language Models (NeurIPS 2025 · similarity 0.87)
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.
