SR²AM: Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

Source: HuggingFace daily papers, 2026-05-23. arxiv: 2605.22138

TL;DR

Reactive agents trained end-to-end with adaptive chain-of-thought tend to over-think: reasoning length explodes without reliable accuracy gains because the model has no control over whether, when, or how deeply to plan. SR²AM (Self-Regulated Simulative Reasoning Agentic LLM) decomposes agent decision-making into three systems: System I (reactive execution of fine-grained actions), System II (simulative reasoning grounded in future-state prediction via a world model), and System III (self-regulation, a learned configurator that decides when and how deeply to plan). Across math, science, tabular analysis, and web search, v0.1-8B and v1.0-30B match Pass@1 of 120-355B and 685B-1T parameter systems respectively, while v1.0-30B uses 25.8-95.3% fewer reasoning tokens than comparable agentic LLMs. RL increases average planning horizon by 22.8% while planning frequency only grows 2.0%: the model learns to plan further ahead rather than more often.

What this paper actually proposes

The three-system decomposition is a clean operationalization of the dual-process / control-hierarchy idea that has been informally circulating since GPT-4. Most prior agentic work conflates planning and execution into a single chain-of-thought, then tries to control reasoning depth via heuristics (max-tokens cap, "think harder" prompts, reward shaping). SR²AM separates them: a configurator decides how much planning to allocate to each step, the planner simulates with an LLM world model, the executor acts.

The 25.8-95.3% reduction in reasoning tokens at matched accuracy is the practical claim. The structural claim: that learned self-regulation extends beyond planning to "how agents govern their own learning and adaptation": is a research-program statement worth holding the field to.

Connections to prior wiki state

This is the most direct reply to a tension that has been building in agent research since the May Worth Watching prediction that token economics will become the binding constraint on agent deployment (Microsoft canceling internal Claude Code licenses this week is the strongest evidence yet that the constraint is now real). SR²AM gives a concrete answer: don't reduce model size, reduce planning frequency. A 30B model that uses 50% fewer tokens than its competitors at matched accuracy has the same effect on unit economics as a 2x model compression: but with no quality loss.

The "RL learns to plan further ahead, not more often" finding is also notable in light of Karpathy's Autoresearch (last-week roundup from Ken Huang's compound-engineering taxonomy) and the broader observation that autonomous research / agentic systems benefit more from longer-horizon single plans than from many short plans. The empirical signal here lines up.

Gaps

The world model is the same LLM as the executor. The paper does not explore whether a smaller, faster world model would suffice (separating the simulator from the executor is the obvious extension). The configurator's decisions: when to plan, how deep: are learned end-to-end but not interpretable. We don't see an analysis of what triggers the configurator to plan deeper.

Also: math, science, tabular, web search are all task families where a structured plan helps. Whether SR²AM helps on tasks with high ambient noise (open-ended dialogue, creative writing) is unaddressed.

Research angle

The strongest open question is whether the configurator generalizes across task families. If the same configurator learned on math + web search makes good planning-allocation decisions on software engineering or biomedical literature search, then SR²AM is the start of a general-purpose self-regulation layer. If not, every domain needs its own configurator and the architecture is closer to a sophisticated multi-skill MoE.

For routing: SR²AM is a temporal router (when and how much to think) where Maestro (today's paper) is a resource router (which expert to call). The two systems are orthogonal and likely compose. A combined SR²AM + Maestro system would route both when and to whom.

Raw source

raw/huggingface/2026-05-23-efficient-agentic-reasoning-through-self-regulated-simulativ.md