Research map / Reinforcement learning / Visual and representation RL
Meta-RL Induces Exploration in Language Agents
Yulun Jiang, Liangze Jiang, Damien Teney, Michael Moor, Maria Brbic
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 |
|---|---|---|---|
| train with cross-episode Meta-RL (LaMer) instead of single-episode GiGPO | Qwen3-4B · Sokoban | pass@3: 44.1 -> 55.9 | Improved |
| train with Meta-RL instead of GiGPO | Qwen3-4B · MineSweeper | pass@3: 55.1 -> 74.4 | Improved |
| train with Meta-RL instead of GiGPO | Qwen3-4B · Webshop | pass@3: 75.2 -> 89.1 | Improved |
| train with Meta-RL instead of GiGPO | Qwen3-4B · MineSweeper | pass@1: 52.0 -> 44.1 | Hurt |
| train with Meta-RL instead of GiGPO | Qwen3-4B · ALFWorld Cool (OOD) | success rate: 58.1 -> 81.0 | Improved |
| train with Meta-RL instead of GiGPO | Llama3.1-8B-Instruct · Sokoban | pass@3: 6.3 -> 20.3 | Improved |
| keep reflection only in inter-episode memory instead of both trajectories and reflections (ablation) | LaMer · Sokoban | pass@3: 55.9 -> 56.4 | Improved |
| add self-reflection to memory (vs trajectory-only) (ablation) | LaMer · Sokoban | pass@3: 34.8 -> 56.4 | Improved |
| give GiGPO baseline inter-episode memory (ablation) | GiGPO · Webshop | pass@3: 75.2 -> 74.0 | Hurt |
| give GiGPO baseline inter-episode memory (ablation) | GiGPO · MineSweeper | pass@3: 55.1 -> 60.4 | Improved |
| increase cross-episode discount gamma_traj (ablation) | LaMer · Sokoban/MineSweeper/Webshop | pass@3 | Mixed |
| use Meta-RL instead of RL on harder tasks | LaMer · Sokoban (more boxes) | success rate: 10% gap on hardest | Improved |
Closest papers by meaning
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.90)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.90)
- Expanding LLM Agent Boundaries with Strategy-Guided Exploration (NeurIPS 2026 · similarity 0.89)
- LLM-Explorer: A Plug-in Reinforcement Learning Policy Exploration Enhancement Driven by Large Language Models (NeurIPS 2025 · similarity 0.89)
- First-Explore, then Exploit: Meta-Learning to Solve Hard Exploration-Exploitation Trade-Offs (NeurIPS 2024 · similarity 0.89)
- ReMA: Learning to Meta-Think for LLMs with Multi-agent Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
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.
