Google's most recent research uncovers a fascinating phenomenon: top-tier reasoning models, exemplified by DeepSeek - R1, naturally develop multiple virtual personas with unique traits during the process of problem - solving. Their reasoning mechanism is akin to a lively social debate. The more intense this "internal debate" grows, the more intelligent the model's performance becomes.

When confronted with highly challenging tasks, the conflicts of viewpoints within the model escalate significantly. In contrast, simpler tasks result in relatively fewer such conflicts. These virtual personas boast a wide array of characteristics, encompassing multiple angles for problem - solving. Their interactions drive the model to scrutinize issues in a more comprehensive manner.

Remarkably, this multi - role interaction isn't a result of artificial design. Instead, it emerges spontaneously as the model strives for reasoning accuracy. The research team managed to decode the AI's "internal group chat" by employing Sparse Autoencoders (SAE). They discovered that the frequency of conversational behavior in reasoning models is substantially higher than that in ordinary instruction - following models.

Experiments have demonstrated that reinforcing discourse markers (like "oh") or adopting conversational thinking training can markedly enhance the model's reasoning accuracy. For example, in the Countdown arithmetic reasoning task, after enhancing conversational features, the model's accuracy soared from 27.1% to 54.8%.

Moreover, reinforcement learning training reveals that dialogue - finetuned models make far greater progress in reasoning tasks compared to monologue - finetuned models. This discovery is in line with the Social Brain Hypothesis, which posits that for AI to achieve greater intelligence, it must first learn to engage in social interactions with different "personalities."

For the complete paper, please visit: https://arxiv.org/abs/2601.10825.