Research map / Continual learning / Continual fine-tuning of LLMs
SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model
Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, Yunchao Wei
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 a small learning rate (0.0001) for the representation layer and 0.01 for the classifier (Slow Learner) (ablation) | Seq FT · Split CIFAR-100 (IN21K-Sup) | Last-Acc (%): 41.77 -> 88.86 | Improved |
| use the slow learner on a hard out-of-distribution benchmark (ablation) | Seq FT · Split ImageNet-R (IN21K-Sup) | Last-Acc (%): 26.95 -> 71.80 | Improved |
| update the representation layer with slow learner instead of freezing it (ablation) | Seq FT · Split CIFAR-100 (IN21K-Sup) | Last-Acc (%): 63.75 -> 88.86 | Improved |
| add classifier alignment with logit normalization on top of slow learner (ablation) | SLCA · Split CUB-200 (IN21K-Sup) | Last-Acc (%): 68.07 -> 84.71 | Improved |
| add classifier alignment with logit normalization on top of slow learner (ablation) | SLCA · Split Cars-196 (IN21K-Sup) | Last-Acc (%): 49.74 -> 67.73 | Improved |
| add logit normalization to classifier alignment (ablation) | SLCA · Split ImageNet-R (IN21K-Sup) | Last-Acc (%): 74.41 -> 77.00 | Improved |
| add logit normalization to classifier alignment (ablation) | SLCA · Split Cars-196 (IN21K-Sup) | Last-Acc (%): 67.90 -> 67.73 | No clear effect |
| align classifier during representation learning instead of post-hoc (ablation) | SLCA · Split CIFAR-100 | Last-Acc (%): worsen by 27.89% | Hurt |
| add classifier alignment to a fixed representation layer (ablation) | Fixed representation layer · Split CUB-200 (IN21K-Sup) | Last-Acc (%): 60.44 -> 82.71 | Improved |
| lower the representation learning rate too far (0.001 vs 0.0001) (ablation) | Seq FT · Split Cars-196 (IN21K-Sup) | Last-Acc (%): 49.74 -> 15.74 at lr 0.001 | Hurt |
| add classifier alignment to BiC with slow learner (ablation) | BiC · Split ImageNet-R (IN21K-Sup) | Last-Acc (%): 64.89 -> 74.49 | Improved |
| add classifier alignment to EWC with slow learner (ablation) | EWC · Split CIFAR-100 (IN21K-Sup) | Last-Acc (%): 89.30 -> 90.61 | Improved |
Later papers that compared with this method
22 later accepted papers report SLCA as a baseline in their results tables; 21 of them report a better number on most of the results they share with it.
- Random Amalgamation of Adapters for Flatter Loss Landscapes: Towards Class-Incremental Learning with Better Stability (AAAI 2026 · better on 5 of 5 shared results)
- Exemplar-Free Class Incremental Learning via Preserving Class-Discriminative Structure (CVPR 2026 · better on 2 of 3 shared results)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · better on 2 of 2 shared results)
- Continual Out-of-Distribution Detection with Analytic Neural Collapse (AAAI 2026 · better on 2 of 2 shared results)
- Sparse Tuning Enhances Plasticity in PTM-based Continual Learning (AAAI 2026 · better on 2 of 2 shared results)
- GR-LoRA: Gradient-Recycling Low-Rank Adaptation for Class-Incremental Learning (ICML 2026 · better on 1 of 1 shared result)
- LCA: Local Classifier Alignment for Continual Learning (ICLR 2026 · better on 1 of 1 shared result)
- DGS: Dual Gradient and Semantic-Shift Guided Low-Rank Adaptation for Class Incremental Learning (CVPR 2026 · better on 1 of 1 shared result)
14 more not listed.
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.
- DualPrompt — DualPrompt: Complementary Prompting for Rehearsal-Free Continual Learning (ECCV 2022) · 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
- SLCA++: Unleash the Power of Sequential Fine-Tuning for Continual Learning with Pre-training (IJCV 2026 · similarity 0.93)
- LCA: Local Classifier Alignment for Continual Learning (ICLR 2026 · similarity 0.89)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 2025 · similarity 0.88)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.88)
- An Effective Dynamic Gradient Calibration Method for Continual Learning (ICML 2024 · similarity 0.86)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · 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.
