Research map / Reinforcement learning / RL from feedback
Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning
Xin Cheng, Shuo He, Lang Feng, Haiyang Xu, Ming Yan, Lei Feng, Bo An
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 |
|---|---|---|---|
| add episode-level advantage A^E to step-level advantage (ablation) | GraphGPO · ALFWorld | success rate (%): 91.67 -> 92.71 | Improved |
| add episode-level advantage A^E to step-level advantage (ablation) | GraphGPO · Sokoban | success rate (%): 83.07 -> 86.98 | Improved |
| add episode-level advantage A^E to step-level advantage (ablation) | GiGPO · Sokoban | success rate (%): 62.5 -> 76.92 | Improved |
| add episode-level advantage A^E to step-level advantage (ablation) | GiGPO · ALFWorld | success rate (%): 91.41 -> 90.88 | No clear effect |
| add dynamic sampling | GraphGPO · ALFWorld | success rate (%): 92.71 -> 98.43 | Improved |
| add dynamic sampling | GraphGPO · WebShop | success rate (%): 78.65 -> 85.68 | Improved |
| add dynamic sampling | GiGPO · ALFWorld | success rate (%): 90.88 -> 96.35 | Improved |
| use graph-based advantage instead of anchor-state GiGPO advantage | GiGPO · ALFWorld Qwen2.5-1.5B | success rate (%): 90.88 -> 92.71 | Improved |
| use graph-based advantage instead of GRPO trajectory-level advantage | GRPO · ALFWorld Qwen2.5-7B | success rate (%): 83.33 -> 95.31 | Improved |
| use graph-based advantage instead of GiGPO | GiGPO · Sokoban | success rate (%): 76.92 -> 86.98 | Improved |
| vary distance discount omega from 0.1 to 0.6 (ablation) | GraphGPO · ALFWorld (1.5B) | success rate (%): 92.71 -> 90.36 | No clear effect |
| vary distance discount omega from 0.2 to 0.8 (ablation) | GraphGPO · WebShop (1.5B) | success rate (%): 78.65 -> 75.00 | No clear effect |
Closest papers by meaning
- Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning (NeurIPS 2026 · similarity 0.94)
- Group-in-Group Policy Optimization for LLM Agent Training (NeurIPS 2025 · similarity 0.90)
- Segment Policy Optimization: Effective Segment-Level Credit Assignment in RL for Large Language Models (NeurIPS 2025 · similarity 0.88)
- StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction (NeurIPS 2026 · similarity 0.88)
- Agentic Reinforced Policy Optimization (ICLR 2026 · similarity 0.86)
- Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · similarity 0.86)
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.
