Research map / Continual learning / Continual fine-tuning of LLMs
PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning
Romain Cosentino
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 PLATE instead of LoRA for reasoning SFT | Qwen2.5-7B · IFEval after AM-DeepSeek-R1 SFT | accuracy drop: LoRA drops about 16 points; PLATE essentially eliminates it | Improved |
| Use PLATE instead of LoRA for reasoning SFT | Qwen2.5-7B · AIME, GSM8K, MATH-500 | accuracy gain: +13 points for both | No clear effect |
| Increase PLATE trainable parameters (ablation) | OLMo-2-7B on Tulu-3 · IFEval and MATH-500 forgetting | accuracy / forgetting: IFEval rises roughly linearly, forgetting nearly flat | Improved |
| Use PLATE instead of LoRA at low parameter budget | OLMo-2-7B on Tulu-3 · IFEval | accuracy: LoRA outperforms PLATE on learnability | Hurt |
| Use PLATE instead of O-LoRA | Qwen 2.5-3B · WikiText-2 retention | perplexity forgetting: O-LoRA shows 10-100x larger forgetting | Improved |
| Use PLATE instead of full fine-tuning | 3-layer MLP · MNIST 0-4 -> 5-9 | task-1 forgetting (%): 26 -> 1.85 | Improved |
| Use PLATE instead of LoRA | 3-layer MLP · MNIST 0-4 -> 5-9 | task-1 forgetting (%): 7-9 -> 1.85 (more than 4x better) | Improved |
| Use PLATE instead of O-LoRA | 3-layer MLP · MNIST 0-4 -> 5-9 | task-1 retention (%): O-LoRA retention ~88 at r=8 down to ~76 at r=128; PLATE 97.45 | Improved |
| Use PLATE instead of LoRA (rank sweep) | DistilBERT-base · AG News -> IMDB | task-1 forgetting (%): LoRA about 2-3 at higher rank; PLATE below 0.5 | Improved |
| Replace weight-derived Q with random orthonormal or high-energy Q (ablation) | PLATE · MNIST 0-4 -> 5-9 | predictive KL on task-1 inputs: 0.009 -> 0.076-0.113 when Q removed | Improved |
| Replace weight-derived Q with random orthonormal or high-energy Q (ablation) | PLATE · MNIST 0-4 -> 5-9 | task-1 forgetting (%): 0.20 -> 1.5-2.4 when Q removed | Improved |
| Use redundant-channel B instead of random B (ablation) | PLATE · MNIST 0-4 -> 5-9 | task-1 forgetting / KL: KL 0.006-0.007, forgetting about 0.19% | No clear effect |
Closest papers by meaning
- Rethinking the Stability-Plasticity Trade-off in Continual Learning from an Architectural Perspective (ICML 2025 · similarity 0.89)
- Plasticity Activation via Polar Operator: A Plug-in Method for Balancing Stability and Plasticity (ICML 2026 · similarity 0.89)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.88)
- Preserving Plasticity in Continual Learning via Dynamical Isometry (ICML 2026 · similarity 0.88)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.87)
- Mitigating Plasticity Loss in Continual Reinforcement Learning by Reducing Churn (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.
