Research map / Continual learning / Continual fine-tuning of LLMs
MoCL: Metabolic Optimization for Curvature-Aware Continual Learning
Jiajun Lai, Qi Liu, Shijie Li, Huaiguang Jiang
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 |
|---|---|---|---|
| use soft metabolic gating instead of binary hard projection (ablation) | MoCL · ImageNet-R (N=20) | A_last: 77.68 -> 79.22 | Improved |
| use soft metabolic gating instead of binary hard projection (ablation) | MoCL · ImageNet-R (N=20) | A_avg: 84.11 -> 85.03 | Improved |
| use Tsallis entropy gate instead of Shannon entropy (ablation) | MoCL · ImageNet-R (N=20) | A_last: 78.43 -> 79.22 | Improved |
| use Tsallis entropy gate instead of Shannon entropy (ablation) | MoCL · ImageNet-R (N=20) | A_avg: 84.57 -> 85.03 | Improved |
| use Tsallis gate instead of sigmoid gating (ablation) | MoCL · ImageNet-R (N=20) | A_last: 74.88 -> 79.22 | Improved |
| use Tsallis gate instead of sigmoid gating (ablation) | MoCL · ImageNet-R (N=20) | A_avg: 82.54 -> 85.03 | Improved |
| use K-FAC activation and gradient subspaces instead of gradient subspace only (ablation) | MoCL · ImageNet-R (N=20) | A_last: 76.85 -> 79.22 | Improved |
| use K-FAC activation and gradient subspaces instead of activation subspace only (ablation) | MoCL · ImageNet-R (N=20) | A_last: 75.90 -> 79.22 | Improved |
| use K-FAC activation and gradient subspaces instead of gradient subspace only (ablation) | MoCL · ImageNet-R (N=20) | A_avg: 83.74 -> 85.03 | Improved |
| use K-FAC activation and gradient subspaces instead of activation subspace only (ablation) | MoCL · ImageNet-R (N=20) | A_avg: 82.56 -> 85.03 | Improved |
| replace CoSO with MoCL as the method (more tasks) | MoCL · ImageNet-R (N=5 to N=20) | A_last: 82.29 vs 81.53 (N=5); 79.22 vs 77.57 (N=20) | Improved |
| replace CoSO with MoCL | MoCL · ImageNet-A (N=10) | A_avg: 68.36 vs 71.27 | 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.
- SD-LoRA — SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025) · better on 7 of 7 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
- From Generic to Dedicated: A Novel Optimizer for Online Continual Learning (NeurIPS 2026 · similarity 0.89)
- Online Curvature-Aware Replay: Leveraging $\mathbf{2^{nd}}$ Order Information for Online Continual Learning (ICML 2025 · similarity 0.89)
- Energy-Structured Low-Rank Adaptation for Continual Learning (ICML 2026 · similarity 0.89)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.88)
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · similarity 0.87)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · 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.
