Research map / Continual learning / Continual fine-tuning of LLMs
Learning to Modulate pre-trained Models in RL
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 a task-index oracle to L2M (ablation) | L2M · CW10 continual | success rate: 65% -> 76% | Helped |
| add a task-index oracle to L2M (ablation) | L2M · DMC6 continual | normalized score: 56% -> 70% | Helped |
| L2M versus prior Continual-World results | L2M · Continual-World v2 | success rate: roughly 40% -> 65% | Helped |
| use L2M instead of L2P-style prompt pool methods | L2M · CW10/DMC6 continual | performance | Helped |
| add EWC regularization for continual RL | Fine-tuned MDDT · CW10/DMC6 continual | forgetting | No clear effect |
| full fine-tuning versus L2M/L2P on pre-training tasks | MDDT · MT40/DMC10 after fine-tuning | performance retention | Hurt |
| scale model from 40M to 200M parameters (ablation) | MDDT · DMC10 pre-training | normalized score: 0.84 -> 1.33 | Helped |
| scale model from 40M to 200M parameters (ablation) | MDDT · MT40 pre-training | success rate: 0.6 -> 0.6 | No clear effect |
| scale model from 40M to 200M parameters (ablation) | MDDT · pre-training | mean reward: 983.92 -> 1030.1 | Mixed |
| use LoRA instead of Adapters | PEFT on MDDT · CW10 single-task | trainable parameter fraction: 2.1% -> 1% | Helped |
| use (IA)^3 instead of Prompt-tuning at similar parameter count | MDDT · CW10 single-task | success rate: 0.21% vs 0.21% params | Helped |
| use state embeddings as prompt-pool query instead of RTG or action (ablation) | L2M · CW10 | success rate | Helped |
Closest papers by meaning
- Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problem (ICML 2024 · similarity 0.90)
- Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data (ICLR 2025 · similarity 0.89)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.88)
- Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models (ICLR 2026 · similarity 0.88)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.88)
- Behavior Injection: Preparing Language Models for Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
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.