AI's secret to getting smarter is actually having a "group chat" inside its brain?!
Google's latest research shows that top-tier reasoning models like DeepSeek-R1 spontaneously "split" into different virtual personalities when solving problems, such as outgoing, rigorous, and skeptical ones...
The reasoning process of large models is a wonderful social gathering and debate among these personalities; it's like the left and right brain competing:
"Is this line of thinking correct? Let's try verifying it this way...""No, the previous assumption overlooked the xx condition"...
Interestingly, AI gets smarter the more it argues.
Research found that when facing high-difficulty tasks like GPQA graduate-level science questions or complex mathematical derivations, this internal conflict of opinions becomes more intense.
Compared to that, for simple tasks like Boolean expressions or basic logical reasoning, the model's internal dialogue significantly decreases.
The model's reasoning process is "left and right brain competing"
The team analyzed the thought trajectories of models like DeepSeek-R1 and QwQ-32B and found that their reasoning processes are full of conversational sense.
The virtual roles split internally not only have very different personalities but also cover more problem-solving angles.
Creative roles are good at proposing novel ideas, critical roles focus on finding and fixing errors, and executive roles are responsible for implementing and verifying...
Through the exchange of these personalities, the collision of different viewpoints allows the model to examine solutions more comprehensively.
Even netizens say that when they think, they also experience "left and right brain competing".
However, this multi-role interaction is not deliberately designed by developers but is spontaneously formed by the model in the process of pursuing reasoning accuracy.
So how did the experiment prove this?
The team used Sparse Autoencoders (SAE) to deeply decode the black box of AI's reasoning, successfully "eavesdropping" on the AI's internal group chat.
First, researchers had the AI perform complex mathematical or logical reasoning tasks. While the model produced its chain of thought, the team simultaneously extracted the activation values of its hidden layer neurons.
However, this data at the time was a complex nonlinear signal composed of hundreds of millions of parameters and could not directly correspond to any semantics.
Inputting this activation data into the SAE, through the SAE's sparse constraint mechanism, the chaotic activations could be decomposed into independent dialogue semantic features such as "self-questioning and self-answering" and "switching perspectives";
By analyzing the activation frequencies of these features and their collaborative relationships in the time series, the team successfully identified different internal logical entities.
Then, labeling the above features with virtual role tags such as "planner" and "verifier" successfully decoded the multi-role dialogue behavior inside the AI.
"Oh!" makes reasoning more accurate
By comparing the reasoning trajectories of DeepSeek-R1 with those of general instruction models like DeepSeek-V3 and Qwen-2.5-32B-IT, it was found that conversational behaviors appear significantly more frequently in reasoning models.
Here's another interesting discovery—
"Oh!" makes reasoning more accurate.
When the team used activation addition to strengthen the model's conversational features and amplify discourse markers like "Oh!" that express surprise and turns, the model's accuracy in the Countdown arithmetic reasoning task directly doubled from 27.1% to 54.8%.
A more critical experimental evidence came from reinforcement learning training.
Researchers did not provide any training signals for conversational structure, only rewarding the model for answering questions correctly, and found that the model would spontaneously learn to use conversational thinking;
And by first fine-tuning the model with multi-agent conversational data, then performing reasoning training, the improvement speed was much faster than directly training reasoning or fine-tuning with monologue reasoning data.
In the Qwen-2.5-3B and Llama-3.2-3B model systems, the accuracy of the conversational fine-tuned model in the early training stage was more than 10% higher than that of the monologue fine-tuned model, and the gap even widened to 22% by the later stages of training for Llama-3.2-3B.
This discovery precisely echoes the famous theory in human evolutionary biology, the Social Brain Hypothesis.
The hypothesis suggests that the evolution of the human brain is primarily to cope with complex social relationships and group interaction needs.
Now it seems that AI is the same; to get smarter, it must first learn to socialize with different "personalities"!
https://arxiv.org/abs/2601.10825https://x.com/sebkrier/status/2013331596863041731
— End —