LLMs don't know how to think (tictacguy.github.io)

Meta has unveiled a groundbreaking approach called Meta-Reasoning, which redefines how Large Language Models (LLMs) handle cognitive processes by asserting that reasoning should be managed externally rather than within the model itself. This system introduces a Cognitive Controller that governs the cognitive operations of the LLM, effectively treating it as a stateless execution substrate that produces structured outputs without making decisions or generating language autonomously. The innovative framework emphasizes exploration of cognitive pathways rather than rote pattern repetition, shaking up conventional expectations around LLM capabilities. The significance of this development for the AI/ML community lies in its potential to enhance reasoning flexibility and robustness in LLMs. By imposing strict cognitive constraints and formal output protocols, Meta-Reasoning aims to prevent the stalling effects of dominant thinking patterns in LLMs. Furthermore, it establishes a new standard for cognitive observability, tracking cognitive trajectory and performance through metrics and formal traces of reasoning operations. This enables a systematic approach to understanding model behavior, offering insights into how LLMs can improve their reasoning dynamics and avoid pitfalls like hallucinations, thus paving the way for more reliable and adaptable AI systems.