Research map / Continual learning / Continual fine-tuning of LLMs
LCA: Local Classifier Alignment for Continual Learning
Tung Tran, Danilo Vasconcellos Vargas, Khoat Than
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 Classifier Alignment (LCA) after incremental merging (ablation) | IM (ViT-B/16-IN1K) · Overall (7 datasets) | average accuracy: 80.9 -> 85.6 | Improved |
| add LCA after incremental merging (ablation) | IM · ImageNet-A | average accuracy: 66.5 -> 75.0 | Improved |
| add LCA after incremental merging (ablation) | IM · VTAB | average accuracy: 84.6 -> 95.2 | Improved |
| add LCA after incremental merging (ablation) | IM · OmniBenchmark | average accuracy: 81.1 -> 81.4 | No clear effect |
| use LCA robustness term instead of plain classifier alignment (CA) (ablation) | IM Max · ImageNet-A | average accuracy: 70.86 -> 73.73 | Improved |
| use LCA instead of CA (ablation) | IM MaxAbs · VTAB | average accuracy: 91.71 -> 95.2 | Improved |
| add classifier alignment (CA) after merging (ablation) | IM Max · ImageNet-A | average accuracy: 66.26 -> 70.86 | Improved |
| choose Min, Max or MaxAbs merge operator (ablation) | IM+LCA · ImageNet-A / VTAB / CARS | average accuracy: 73.00-75.0 (IN-A) | No clear effect |
| add LCA to MOS and SLCA as final alignment step (ablation) | MOS, SLCA · seven CIL datasets | average accuracy | Improved |
| add LCA to IM (ablation) | IM · CIFAR100-C and CIFAR100-P | robustness accuracy: +2% (C), +2.5% (P) | Improved |
| make robustness penalty lambda too strong (ablation) | LCA · CIFAR100 and another dataset | accuracy | Hurt |
| add LCA to VPT-based PEFT merging (ablation) | IM with VPT · VTAB | accuracy | 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++: Unleash the Power of Sequential Fine-Tuning for Continual Learning with Pre-training (IJCV 2026) · 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
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.87)
- Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification (ICLR 2025 · similarity 0.87)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.87)
- Continual learning with the neural tangent ensemble (NeurIPS 2024 · similarity 0.86)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.86)
- An Effective Dynamic Gradient Calibration Method for Continual Learning (ICML 2024 · similarity 0.86)
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.
