Research map / Large language models / Agents, code and math
Translate Policy to Language: Flow Matching Generated Rewards for LLM Explanations
Xinyi Yang, Liang Zeng, Heng Dong, Chao Yu, Xiaoran Wu, Huazhong Yang, Yu Wang, Milind Tambe, Tonghan Wang
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 rectified flow reward model instead of using Proxy LLM rewards directly (ablation) | Explanation LLM (Llama-3.1-8B-Instruct) · SMAC | accuracy: 0.640 -> 0.764 | Improved |
| add rectified flow reward model instead of using Proxy LLM rewards directly (ablation) | Explanation LLM (Llama-3.1-8B-Instruct) · MMLU | accuracy: 0.703 -> 0.772 | Improved |
| add rectified flow reward model instead of using Proxy LLM rewards directly (ablation) | Explanation LLM (Llama-3.1-8B-Instruct) · MathQA | accuracy: 0.694 -> 0.804 | Improved |
| use sentence-level rewards instead of sparse last-token reward (ablation) | Ours · SMAC | accuracy: 0.752 -> 0.764 | Improved |
| use sentence-level rewards instead of sparse last-token reward (ablation) | Ours · MMLU | accuracy: 0.757 -> 0.772 | Improved |
| use sentence-level rewards instead of sparse last-token reward (ablation) | Ours · MathQA | accuracy: 0.773 -> 0.804 | Improved |
| add cross-attention in the rectified flow reward model (ablation) | Ours · SMAC | accuracy: 0.662 -> 0.764 | Improved |
| add cross-attention in the rectified flow reward model (ablation) | Ours · MMLU | accuracy: 0.721 -> 0.772 | Improved |
| add cross-attention in the rectified flow reward model (ablation) | Ours · MathQA | accuracy: 0.728 -> 0.804 | Improved |
| use rectified flow rewards instead of PPO with sentence-level reward (ablation) | PPO · SMAC | accuracy: 0.688 -> 0.764 | Improved |
| use Ours instead of SFT on Qwen2.5-7B-Instruct backbone | Explanation LLM · SMAC | accuracy: 0.594 -> 0.783 | Improved |
| apply rejection sampling to PPO baseline (ablation) | PPO · SMAC | accuracy: 0.653 -> 0.642 | Hurt |
Closest papers by meaning
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.89)
- Learning Agentic Policy from Action Guidance (NeurIPS 2026 · similarity 0.88)
- Reinforce LLM Reasoning through Multi-Agent Reflection (ICML 2025 · similarity 0.88)
- The Synergy of LLMs & RL Unlocks Offline Learning of Generalizable Language-Conditioned Policies with Low-fidelity Data (ICML 2025 · similarity 0.88)
- RLZero: Direct Policy Inference from Language Without In-Domain Supervision (NeurIPS 2025 · similarity 0.88)
- Reinforcing LLM Agents via Policy Optimization with Action Decomposition (NeurIPS 2024 · 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.
