Research map / Large language models / Agents, code and math
Beyond ReAct: A Planner-Centric Framework for Complex Tool-Augmented LLM Reasoning
Xiaolong Wei, Yuehu Dong, Xingliang Wang, Xingyu Zhang, Zhejun Zhao, Dongdong Shen, Long Xia, Dawei Yin
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 GRPO RL stage after SFT (ablation) | Qwen3-8B planner · ComplexTool-Plan Easy | DAG exact match: 0.781 -> 0.803 | Improved |
| add GRPO RL stage after SFT (ablation) | Qwen3-8B planner · ComplexTool-Plan Hard | DAG exact match: 0.295 -> 0.319 | Improved |
| add GRPO RL stage after SFT (ablation) | Qwen3-4B planner · ComplexTool-Plan Hard | DAG exact match: 0.241 -> 0.259 | Improved |
| add GRPO RL stage after SFT (ablation) | Qwen3-1.7B planner · ComplexTool-Plan Easy | DAG exact match: 0.723 -> 0.756 | Improved |
| add GRPO RL stage after SFT (ablation) | Qwen3-4B planner · ComplexTool-Plan Hard | edge F1: 0.627 -> 0.621 | Hurt |
| add GRPO RL stage after SFT (ablation) | Qwen3-8B planner · ComplexTool-Plan Hard | edge F1: 0.657 -> 0.659 | No clear effect |
| scale planner size from 1.7B to 8B (SFT+RL) (ablation) | Qwen3 planner · ComplexTool-Plan Hard | DAG exact match: 0.218 -> 0.259 -> 0.319 | Improved |
| use fine-tuned Qwen3-8B planner instead of prompted proprietary LLM | DAG planning · ComplexTool-Plan Hard | DAG exact match: 0.106 -> 0.319 | Improved |
| use plan-then-execute planner (Qwen3-8B RL) instead of ReAct with GPT-4 | StableToolBench · StableToolBench average | SoPR: 48.2 -> 59.8 | Improved |
| use Qwen3-8B planner instead of GPT-4 (DFSDT) | StableToolBench · StableToolBench average | SoPR: 70.3 -> 59.8 | Hurt |
| use global DAG planning instead of iterative DTA-Llama | StableToolBench · StableToolBench | inference steps: 2.48 -> 2.29 | Improved |
Closest papers by meaning
- Tool-Planner: Task Planning with Clusters across Multiple Tools (ICLR 2025 · similarity 0.90)
- ToolTree: Efficient LLM Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning (ICLR 2026 · similarity 0.90)
- Thought of Search: Planning with Language Models Through The Lens of Efficiency (NeurIPS 2024 · similarity 0.88)
- Hierarchical Planning for Complex Tasks with Knowledge Graph-RAG and Symbolic Verification (ICML 2025 · similarity 0.88)
- Graph-enhanced Large Language Models in Asynchronous Plan Reasoning (ICML 2024 · similarity 0.87)
- NaviAgent: Graph‑Driven Bilevel Planning for Scalable Tool Orchestration (ICML 2026 · 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.
