Researchers propose State of Thought (SoT), a reasoning paradigm that replaces externally imposed control with endogenous state management. The system extracts a dynamics-geometric state from the model's internal information transfer to govern how reasoning unfolds.
Quantitative Improvements Across Benchmarks
Testing across three large language models and sixteen datasets yielded consistent improvements over baselines. SoT achieved a 1.34x improvement in quantitative reasoning, 1.62x in general reasoning, 1.76x in symbolic-and-code tasks, and 2.51x in long-context scenarios.
Efficiency Gains in Tokens and Latency
The approach reduced generated tokens by 62.6% and end-to-end latency by 44.6%. In visual-language model evaluations across two scales and three reasoning tasks, it improved mean accuracy by 3.8 points while using 74.9% fewer completion tokens and achieving 73.5% lower latency than search-based methods.
Performance Under Constrained Access
SoT maintains robust performance even with limited data access. It retains 38.2% or 36.5% mean accuracy gains in training-free or embedding-only settings, respectively. Trajectory-only judging reached 84.1% agreement across three API models.
Source: https://arxiv.org/abs/2609.16055
