Research map / Continual learning / Continual fine-tuning of LLMs
SABER: Continual Learning with Representation Conflict Management
Xuandi Luo, Huaidong Zhang, Yi Xie, Shengfeng He
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 orthogonal initialization of A_t (ablation) | LoRA baseline · ImageNet-R (T=10) | A_last: 67.65 -> 72.39 | Improved |
| add subspace alignment (SA) (ablation) | LoRA baseline · ImageNet-R (T=10) | A_last: 67.65 -> 73.25 | Improved |
| add knowledge recomposition (KR) (ablation) | LoRA baseline · ImageNet-R (T=10) | A_last: 67.65 -> 74.08 | Improved |
| add knowledge balancing (KB) to KR (ablation) | KR · ImageNet-R (T=10) | A_last: 74.08 -> 75.11 | Improved |
| add SA to KR (ablation) | KR · ImageNet-R (T=10) | A_last: 74.08 -> 78.53 | Improved |
| add KB to SA+KR (full SABER) (ablation) | SA+KR · ImageNet-R (T=10) | A_last: 78.53 -> 80.30 | Improved |
| add SA+KR+KB at long sequence (ablation) | LoRA baseline · ImageNet-R (T=20) | A_last: 62.98 -> 77.77 | Improved |
| use dense mutually orthogonal A_t instead of random init (ablation) | SABER · ImageNet-R (T=10) | A_last: 2.76 -> 80.30 | Improved |
| use dense mutually orthogonal A_t instead of standard basis (ablation) | SABER · ImageNet-R (T=10) | A_last: 76.29 -> 80.30 | Improved |
| use dense mutually orthogonal A_t instead of intra-ortho init (ablation) | SABER · ImageNet-R (T=10) | A_avg: 76.10 -> 85.98 | Improved |
| vary common-knowledge ratio alpha within 0.5-0.8 (ablation) | SABER · ImageNet-R (T=10) | A_avg: accuracy drops by at most 0.63% | No clear effect |
| set alpha=1.0 (no task-specific knowledge, SA only) (ablation) | SABER · ImageNet-R (T=10) | A_avg | Hurt |
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.
- CL-LoRA — CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025) · better on 10 of 10 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
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.88)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.88)
- When the Merge Coefficient Stops Mattering: Proximity Regularized Merging for Continual LoRA Adaptation (NeurIPS 2026 · similarity 0.87)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.87)
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 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.
