Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Guiding Pretraining in Reinforcement Learning with Large Language Models
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 LLM-suggested goals as an exploration reward (ELLM) instead of novelty-only rewards | RL pretraining agent · Crafter pretraining | unique achievements per episode: about 6 vs less than 3 | Helped |
| Remove novelty bias (allow repeated rewards for same goal) (ablation) | ELLM · Housekeep and Crafter pretraining | exploration diversity / success | Hurt |
| Add textual state captions to policy inputs (pixels + text) (ablation) | ELLM and baselines · Crafter pretraining | unique achievements per episode | Helped |
| Replace finetuning with guided exploration (pretrained policy used for epsilon-greedy actions in new agent) | ELLM pretrained policy on Crafter downstream tasks · Crafter downstream | success rate | Helped |
| Use a learned ClipCap captioner instead of ground-truth transition captioner (ablation) | ELLM · Crafter pretraining | achievement rewards / pretraining performance | No clear effect |
| Goal-conditioning the policy on suggested goals (ablation) | ELLM in Crafter pretraining · Crafter pretraining | unique achievements per episode | No clear effect |
| Goal-conditioned ELLM policy during Housekeep finetuning | ELLM in Housekeep · Housekeep finetuning | success rate | Hurt |
| Add LLM common-sense priors via ELLM rewards in Housekeep pretraining | ELLM · Housekeep pretraining and downstream | rearrangement success rate | Helped |
| Add NovelD baseline comparison | Crafter pretraining · Crafter pretraining | achievements per episode | No clear effect |
Closest papers by meaning
- Expanding LLM Agent Boundaries with Strategy-Guided Exploration (NeurIPS 2026 · similarity 0.91)
- Representation-Based Exploration for Language Models: From Test-Time to Post-Training (ICLR 2026 · similarity 0.90)
- Can large language models explore in-context? (NeurIPS 2024 · similarity 0.90)
- On the Modeling Capabilities of Large Language Models for Sequential Decision Making (ICLR 2025 · similarity 0.89)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.89)
- SENSEI: Semantic Exploration Guided by Foundation Models to Learn Versatile World Models (ICML 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.