AI Paper Atlas

Research map / Continual learning / Continual fine-tuning of LLMs

PLATE: Plasticity-Tunable Efficient Adapters for Geometry-Aware Continual Learning

Romain Cosentino

ICML 2026 · Continual learning › Continual fine-tuning of LLMs · approach: Low-rank and model-merging adaptation · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
Use PLATE instead of LoRA for reasoning SFTQwen2.5-7B · IFEval after AM-DeepSeek-R1 SFTaccuracy drop: LoRA drops about 16 points; PLATE essentially eliminates itImproved
Use PLATE instead of LoRA for reasoning SFTQwen2.5-7B · AIME, GSM8K, MATH-500accuracy gain: +13 points for bothNo clear effect
Increase PLATE trainable parameters (ablation)OLMo-2-7B on Tulu-3 · IFEval and MATH-500 forgettingaccuracy / forgetting: IFEval rises roughly linearly, forgetting nearly flatImproved
Use PLATE instead of LoRA at low parameter budgetOLMo-2-7B on Tulu-3 · IFEvalaccuracy: LoRA outperforms PLATE on learnabilityHurt
Use PLATE instead of O-LoRAQwen 2.5-3B · WikiText-2 retentionperplexity forgetting: O-LoRA shows 10-100x larger forgettingImproved
Use PLATE instead of full fine-tuning3-layer MLP · MNIST 0-4 -> 5-9task-1 forgetting (%): 26 -> 1.85Improved
Use PLATE instead of LoRA3-layer MLP · MNIST 0-4 -> 5-9task-1 forgetting (%): 7-9 -> 1.85 (more than 4x better)Improved
Use PLATE instead of O-LoRA3-layer MLP · MNIST 0-4 -> 5-9task-1 retention (%): O-LoRA retention ~88 at r=8 down to ~76 at r=128; PLATE 97.45Improved
Use PLATE instead of LoRA (rank sweep)DistilBERT-base · AG News -> IMDBtask-1 forgetting (%): LoRA about 2-3 at higher rank; PLATE below 0.5Improved
Replace weight-derived Q with random orthonormal or high-energy Q (ablation)PLATE · MNIST 0-4 -> 5-9predictive KL on task-1 inputs: 0.009 -> 0.076-0.113 when Q removedImproved
Replace weight-derived Q with random orthonormal or high-energy Q (ablation)PLATE · MNIST 0-4 -> 5-9task-1 forgetting (%): 0.20 -> 1.5-2.4 when Q removedImproved
Use redundant-channel B instead of random B (ablation)PLATE · MNIST 0-4 -> 5-9task-1 forgetting / KL: KL 0.006-0.007, forgetting about 0.19%No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy