Key Features

• System I + II + III decomposition: a configurator (System III) decides per-turn whether to plan, continue an existing plan, or act directly; a simulative planner (System II) constructs plans grounded in predicted future states; reactive execution (System I) handles fine-grained reasoning and tool use.
• SFT + RL training: supervised learning on data encoding the self-regulated planning structure, followed by reinforcement learning (GRPO) for task success.
• Agentic tool use: web search (SerpAPI), web browsing with LLM summarization, and stateless Python code execution (SandboxFusion).
• Reasoning efficiency: 5,518 reasoning tokens per trajectory on average — 51% fewer than MiroThinker-v1.5-30B and 95% fewer than ASearcher-Web-QWQ-v2 for comparable or better accuracy.
• RL learns to plan further, not more often: RL increases average planning horizon by 22.8% while planning frequency grows only 2.0 percentage points.

Quick Start

See the GitHub repository for setup and inference instructions.

Main Results

SR²AM-v1.0-30B sits above the size-vs-accuracy trendline in (a) and on the Pareto frontier of reasoning-token efficiency vs. accuracy among 30/32B models in (b). The full benchmark breakdown is in the paper.

Citation

bibtex
Copy code
@article{deng2026sr2am,
  title={Efficient Agentic Reasoning Through Self-Regulated Simulative Planning},
  author={Deng, Mingkai and Hou, Jinyu and Neves, Lara Sá and
          Pimpalkhute, Varad and Killian, Taylor W. and
          Liu, Zhengzhong and Xing, Eric P.},
  journal={arXiv preprint arXiv:2605.22138},
  year={2026}
}

License

Released under the Apache License 2.0.