🔗 Source: arXiv

Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

🚀 Technical Novelty

  • Mechanism: Decomposes agentic decision-making into three distinct, learnable stages: a configurator (System III) that dynamically decides when to plan, a simulative planner (System II) that uses the LLM as an in-language world model to predict future states, and reactive execution (System I) for direct action.
  • Nuance: Unlike unconstrained chain-of-thought or rigid always-on planners, this architecture explicitly controls planning presence and horizon via a learned regulator, preventing inefficient token bloat while preserving structured, verifiable deliberation across diverse tasks without per-domain engineering.

💡 Yield

  • SR2AM-v1.0-30B achieves Pass@1 performance competitive with 685B–1T parameter agentic models across math, science, tabular, and web reasoning tasks.
  • Reduces reasoning token consumption by 25.8–95.3% compared to similarly scaled competitors.
  • RL training shifts the model’s strategy from planning more often (+2.0%) to planning further ahead (+22.8% horizon), optimizing efficiency without sacrificing accuracy.

⚠️ Limitations

  • Supervised data construction relies on reconstructing plans from pretrained LLM traces or multi-module prompting, which may not perfectly capture optimal planning trajectories.
  • RL training requires careful trajectory filtering to prevent format collapse and depends on accurate LLM-judge rewards for answer correctness.
  • Currently validated on language-based interactive reasoning and tool-use environments rather than physical embodiment or highly dynamic real-world settings.