Research map / Reinforcement learning / RL from feedback
ResT: Reshaping Token-Level Policy Gradients for Tool-Use Large Language Models
Zihan Lin, Xiaohan Wang, Jie Cao, Jiajun Chai, Guojun Yin, Wei Lin, Ran He
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 |
|---|---|---|---|
| use ResT token reweighting instead of GRPO | Qwen3-8B · BFCL multi-turn overall | accuracy: 36.00 -> 40.13 | Improved |
| use ResT instead of GRPO | Qwen3-14B · BFCL multi-turn overall | accuracy: 38.88 -> 44.25 | Improved |
| use ResT instead of GRPO | Qwen3-4B-2507 · BFCL multi-turn overall | accuracy: 41.62 -> 50.38 | Improved |
| apply SFT or GRPO to Qwen3-4B-2507 | Qwen3-4B-2507 · BFCL multi-turn overall | accuracy: 48.00 -> 43.00 (SFT); 41.62 (GRPO) | Hurt |
| use ResT instead of GRPO | Qwen3-8B · API-Bank overall | accuracy: 66.50 -> 70.69 | Improved |
| use ResT instead of SFT+GRPO | Qwen3-4B-2507 · API-Bank overall | accuracy: 68.51 -> 68.68 | No clear effect |
| add dynamic scaling reward (ablation) | ResT · API-Bank Qwen3-8B | accuracy: 64.15 -> 70.69 | Improved |
| add gradients for chain-of-thought tokens (ablation) | ResT · API-Bank Qwen3-8B | accuracy: 66.33 -> 70.69 | Improved |
| add curriculum learning weights (ablation) | ResT · API-Bank Qwen3-8B | accuracy: 65.83 -> 70.69 | Improved |
| add curriculum learning weights (ablation) | ResT · API-Bank Qwen3-4B-2507 | accuracy: 67.84 -> 68.68 | Improved |
| use ResT instead of SFT on Llama-3.2-3B | Llama-3.2-3B-Instruct · BFCL live | accuracy: 63.04 -> 61.53 | Hurt |
| use ResT on Llama-3.2-3B-Instruct vs GRPO | Llama-3.2-3B-Instruct · BFCL non-live AST | accuracy: 74.38 -> 83.19 | Improved |
Closest papers by meaning
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.91)
- Nemotron-Research-Tool-N1: Exploring Tool-Using Language Models with Reinforced Reasoning (ICLR 2026 · similarity 0.91)
- ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- ToolRL: Reward is All Tool Learning Needs (NeurIPS 2025 · similarity 0.90)
- Reinforcement Mid-Training (ICLR 2026 · similarity 0.90)
- STRIDE: Learnable Stepwise Language Feedback for LLM Reasoning (NeurIPS 2026 · 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.
