Research map / Continual learning / Continual fine-tuning of LLMs
GR-LoRA: Gradient-Recycling Low-Rank Adaptation for Class-Incremental Learning
Yipeng Lin, Fengqiang Wan, Yang Yang
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 Local Mismatch Suppression (ablation) | GR-LoRA with CA · ImageNet-R (20 tasks) | A_Last (%): 78.22 -> 79.12 | Improved |
| add Global Mismatch Suppression (ablation) | GR-LoRA with CA · ImageNet-R (20 tasks) | A_Last (%): 78.22 -> 79.65 | Improved |
| add both LMS and GMS (ablation) | GR-LoRA with CA · ImageNet-R (20 tasks) | A_Last (%): 78.22 -> 80.23 | Improved |
| add LMS and GMS (ablation) | GR-LoRA with CA · ImageNet-R (20 tasks) | A_Avg (%): 83.77 -> 85.05 | Improved |
| recycle orthogonal-discarded gradients into task-specific modules (GR-LoRA with CA vs LoRA-DRS) (ablation) | LoRA-DRS · ImageNet-R (20 tasks) | A_Last (%): 77.90 -> 78.22 | Improved |
| share an orthogonal module across tasks instead of single independent LoRA per task (ablation) | GR-LoRA · ImageNet-R (20 tasks) | A_Last (%): 70.50 -> 80.23 | Improved |
| use full variant instead of efficient variant (E-GR-LoRA) (ablation) | GR-LoRA · ImageNet-R (20 tasks) | A_Last (%): 79.21 -> 80.23 | Improved |
| reduce LoRA rank from 10 to 5 (ablation) | GR-LoRA · ImageNet-R (20 tasks) | A_Last (%): 80.23 -> 80.22 | No clear effect |
| increase LoRA rank from 10 to 20 (ablation) | GR-LoRA · ImageNet-R (20 tasks) | A_Last (%): 80.23 -> 80.53 | No clear effect |
| use entropy-based multi-branch task selection (GR-LoRA) instead of LoRA-DRS | GR-LoRA · ImageNet-R (50 tasks) | inference time per image (s): 1.78 -> 2.58 | Hurt |
| use GR-LoRA instead of MACIL | GR-LoRA · CUB-200 (10 tasks) | A_Last (%): 90.23 -> 89.91 | Hurt |
| use GR-LoRA instead of MACIL | GR-LoRA · ImageNet-R (10 tasks) | A_Last (%): 81.82 -> 82.09 | 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.
- SLCA — SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model (ICCV 2023) · 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
- SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025 · similarity 0.92)
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.91)
- SDS-LoRA: Overcoming Anisotropic Gradient Scaling in Low-Rank Adaptation (NeurIPS 2026 · similarity 0.90)
- Flat-LoRA: Low-Rank Adaptation over a Flat Loss Landscape (ICML 2025 · similarity 0.89)
- GraLoRA: Granular Low-Rank Adaptation for Parameter-Efficient Fine-Tuning (NeurIPS 2025 · similarity 0.89)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.89)
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.
