Reinforcement learning trains the system to optimize computational allocation.
Eric Xing#1357
1/4 Frontier LLMs are converging on adaptive reasoning.
But controlling how much to think is not the same as controlling what kind of thinking to do.
SR²AM introduces self-regulated simulative reasoning: an agent that simulates possible futures through a world model and learns when that simulation is worth the cost.
2/4 The system decomposes deliberation into three processes: reactive execution (System I), future-state simulation via LLM-as-world-model (System II), and a learned configurator (System III) that decides when to simulate, how far ahead, and when to act directly.
RL trains the configurator to plan further ahead, not more often. Allocation, not compression.
1/4 Frontier LLMs are converging on adaptive reasoning.
But controlling how much to think is not the same as controlling what kind of thinking to do.
SR²AM introduces self-regulated simulative reasoning: an agent that simulates possible futures through a world model and learns when that simulation is worth the cost.
Reinforcement learning trains the system to optimize computational allocation.
Eric Xing#13571/4 Frontier LLMs are converging on adaptive reasoning.
But controlling how much to think is not the same as controlling what kind of thinking to do.
SR²AM introduces self-regulated simulative reasoning: an agent that simulates possible futures through a world model and learns when that simulation is worth the cost.