Free Energy Heuristics: Fast-And-Frugal Cognition as Active Inference Under Uncertain Precision

The paper presents a theoretical framework linking fast-and-frugal heuristics and active inference to meta-uncertainty in large language models (LLMs). It demonstrates that excessive reasoning in chain-of-thought (CoT) can degrade model performance in high-meta-uncertainty scenarios, particularly in mid-to-large models, while showing no negative impact in cases with definite answers. The study introduces the FEH-79 benchmark to evaluate these effects across seven models and finds significant accuracy drops in high-uncertainty contexts, emphasizing the importance of understanding meta-uncertainty for optimizing LLM performance.