Energy-Conditioned Thinking: A Three-State Framework for Adaptive Depth and Halting

Current work on reasoning in large language models often relies on explicit chain-of-thought (CoT) as a linear token-level trace, implicitly assuming uniform compute per token and offering limited control over when to think deeply and when to stop. We propose the Structural Energy Framework (SEF), a metabolic view of reasoning that models implicit thinking as adaptive budget allocation across three recurrent states: Diffuse (D) for low-activation standby, Aggregation (A) for targeted activation and consolidation, and Goal-Directed Drive (G) for high-budget deep reasoning. SEF reframes implicit reasoning as latent state scheduling, providing a unifying lens over adaptive compute, halting, looped architectures, and thinking-token methods. We introduce measurable state diagnostics—state occupancy and transition patterns—that evaluate when models enter or exit deep thinking. Here, "energy" denotes a compute budget proxy, enabling measurable state telemetry. SEF offers a compact foundation for designing and auditing latent thinking systems beyond CoT.