Test‑time compute has become a dominant strategy for boosting Large Language Model (LLM) capabilities. Existing test‑time reasoning approaches rely heavily on external control—either fixed reasoning programs or costly expansion within constrained search spaces—thereby limiting generalization and efficiency. We introduce State of Thought (SoT), a new paradigm that lets an LLM’s internal reasoning state steer the reasoning process. Concretely, SoT extracts a compact dynamics‑geometric state from the model’s internal information flow and employs a $582$‑parameter controller on frozen backbones to selectively activate historical reasoning support useful under the current state, framing reasoning as a state‑conditioned process over evidence rather than an externally prescribed token chain.
Across three LLMs and sixteen datasets, SoT consistently improves mean‑baseline accuracy while reducing generated tokens by $62.6\%$ and end‑to‑end latency by $44.6\%$. Quantitative, general, symbolic‑and‑code, and long‑context reasoning see gains of $1.34\times$, $1.62\times$, $1.76\times$, and $2.51\times$ respectively. In evaluations on two VLM scales and three reasoning tasks, SoT raises mean accuracy by $3.8$ points, cuts completion tokens by $74.9\%$, and lowers latency by $73.5\%$ compared with search‑based methods. Even under constrained access, SoT retains $38.2\%/36.5\%$ mean accuracy improvements in training‑free/embedding‑only settings, and trajectory‑only judging achieves $84.1\%$ agreement across three API models. Together, endogenous state‑driven reasoning offers a generalizable and efficient alternative.
Review