Research map / Reinforcement learning / RL from feedback
Grounded Reinforcement Learning: Learning to Win the Game under Human Commands
Shusheng Xu, Huaijie Wang, Yi Wu
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 |
|---|---|---|---|
| Replace unconditioned RL training (NA commands) with random-command training (ablation) | RED · MiniRTS | win rate | Hurt |
| Replace NA training commands with a learned human-proxy command model (ablation) | RED · MiniRTS | win rate | Hurt |
| Add RED self-distillation over BC-only Switch policy under Oracle commands | RED vs Switch · MiniRTS Oracle commands | win rate (%): 78.7 -> 92.6 | Improved |
| Add RED over Joint soft-constraint objective under NA commands | RED vs Joint · MiniRTS NA commands | win rate (%): 51.8 -> 57.8 | Improved |
| Add RED over Joint under Oracle commands | RED vs Joint · MiniRTS Oracle commands | win rate (%): 90.2 -> 92.6 | Improved |
| Add RED over Joint under Random commands | RED vs Joint · MiniRTS Random commands | win rate (%): 32.3 -> 29.8 | No clear effect |
| Add RED over Switch under Random commands | RED vs Switch · MiniRTS Random commands | win rate (%): 12.3 -> 29.8 | Improved |
| Add IRL learned reward instead of BC-based constraint | IRL · MiniRTS Oracle commands | win rate (%): 89.3 -> 57.3 | Hurt |
| Add IRL learned reward on unit building tasks (Dragon) | IRL vs RED · MiniRTS Dragon build success | success rate (%): 88.8 -> 1.8 | Hurt |
| Raise ratio ||Dk||:||D|| from 1:1 to 2:1 (more RL samples) (ablation) | RED · MiniRTS Oracle commands | win rate (%): 92.6 -> 87.7 | Hurt |
| Lower ratio ||Dk||:||D|| from 1:1 to 1:0.25 (ablation) | RED · MiniRTS Oracle commands | win rate (%): 92.6 -> 89.6 | Hurt |
| Raise ratio ||Dk||:||D|| from 1:0.25 to 2:1 (ablation) | RED · MiniRTS validation NLL | NLL: 2.92 -> 3.15 | Hurt |
Closest papers by meaning
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.89)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.89)
- DUO: Diverse, Uncertain, On-Policy Query Generation and Selection for Reinforcement Learning from Human Feedback (AAAI 2025 · similarity 0.88)
- Uni-RLHF: Universal Platform and Benchmark Suite for Reinforcement Learning with Diverse Human Feedback (ICLR 2024 · similarity 0.88)
- Zeroth-Order Policy Gradient for Reinforcement Learning from Human Feedback without Reward Inference (ICLR 2025 · similarity 0.88)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · similarity 0.87)
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.
