Internal Dialogue of LLMs: How Models Simulate Collective Reasoning

Modern large language models (LLM) demonstrate striking reasoning capabilities, but how exactly do they achieve such results? It was traditionally believed that the Chain-of-Thought method, in which a model constructs a logical chain step-by-step, was the key to success. However, a recent study conducted by scientists from Google Research and University of Chicago revealed something far more interesting: within LLMs, there occurs not simply sequential reasoning, but a complex process that mimics multi-sided dialogue, a kind of "meeting of minds."

Instead of a monologue, the model generates different perspectives that come into conflict, debate, and ultimately reach reconciliation. This phenomenon, called "Society of Thought," suggests that LLMs have spontaneously learned to imitate what philosophers and psychologists have long described as the nature of thinking – an internal dialogue between different "voices."

Researchers identified four key patterns in this "conversational dynamics": asking questions, shifting perspectives, conflict, and reconciliation. Moreover, they discovered that in the reasoning process, models reproduce 12 socio-emotional roles described in the Bales' IPA (Interaction Process Analysis) system, which testifies to the high complexity of internal interactions.

A key factor influencing the accuracy of reasoning is the diversity of perspectives. The more different points of view a model generates, the higher the probability of finding the correct solution. Experiments with activation steering, RL-training, and transfer effects confirmed this hypothesis, showing that stimulating diversity within the model leads to improved performance.

The imitation of internal dialogue opens new horizons in LLM development. Instead of simply increasing model size and training data volume, one can focus on creating more effective mechanisms for generating and managing different perspectives. This could lead to the creation of models capable of solving more complex tasks and making more balanced decisions. This discovery has far-reaching consequences for the entire artificial intelligence industry, as it shows that the key to creating truly intelligent machines may lie not only in increasing computational power, but also in understanding and reproducing complex cognitive processes inherent to human thinking. Ultimately, this research emphasizes the importance of an interdisciplinary approach to AI development, combining knowledge from computer science, psychology, and philosophy.