LLMs' Inner Dialogue: How Models Simulate a Collective Mind

Modern large language models (LLMs) show remarkable reasoning abilities, but how exactly do they get there? The conventional view was that Chain-of-Thought, in which a model builds a logical chain step by step, is the key to success. But a recent study by scientists at Google Research and the University of Chicago found something far more interesting: inside an LLM, reasoning is not merely sequential. It is a complex process that simulates a multi-sided dialogue, a kind of "council of minds".

Instead of a monologue, the model generates different perspectives that clash, debate and eventually reach reconciliation. The phenomenon, dubbed "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".

The researchers identified four key patterns in this "conversational dynamic": asking questions, shifting perspectives, conflict and reconciliation. They also found that, while reasoning, the models reproduce 12 socio-emotional roles described in Bales' IPA (Interaction Process Analysis) system, which points to a high degree of complexity in these internal interactions.

The key factor affecting reasoning accuracy is diversity of perspectives. The more different viewpoints a model generates, the higher the chance it finds the right solution. Experiments with activation steering, RL training and transfer effects confirmed this hypothesis, showing that encouraging diversity inside the model improves its performance.

Imitating internal dialogue opens new directions for LLM development. Rather than simply growing model size and training data, developers can focus on building better mechanisms for generating and managing different perspectives. This may lead to models that can solve harder problems and make more balanced decisions. The finding has far-reaching implications for the whole AI industry, because it suggests that the key to truly intelligent machines may lie not only in more computing power, but also in understanding and reproducing the complex cognitive processes of human thought. Ultimately, the study highlights the importance of an interdisciplinary approach to AI development, one that combines computer science, psychology and philosophy.