Reproduce “Efficient Agentic Reasoning Through Self-Regulated Simulative Planning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingkai Deng, Jinyu Hou, Lara Sá Neves, Varad Pimpalkhute, Taylor W. Killian, Zhengzhong Liu, Eric P. Xing
Researchers from the Institute of Foundation Models (IFM) and Carnegie Mellon University developed SR²AM, an AI agent architecture that decomposes decision-making into self-regulation, simulative planning, and reactive execution. This approach enabled SR²AM-v1.0-30B to achieve an overall Pass@1 score of 71.3, competitive with much larger models, while consuming 25.8–95.3% fewer reasoning tokens than other 30-32B agentic LLMs.
Key insights
- Decomposing agentic reasoning into distinct, self-regulated planning, simulative reasoning, and reactive execution systems provides superior accuracy-efficiency tradeoffs.
- A learned 'configurator' (System III) can dynamically decide when and how deeply to engage in planning, avoiding the inefficiency of constant deliberation.
- Grounded deliberation in future-state prediction via a world model offers a unified and explicit planning structure applicable across diverse tasks without domain-specific engineering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.