On March 3rd, we’re delving into a new paper, "Reasoning Models Generate Societies of Thought," currently under review for Nature, and presented by University of Chicago Professor James Evans.
, DeepSeek-R1 and QwQ-32B) improve not just by generating longer chains of thought, but by implicitly simulating multi-agent-like internal dialogue—a “society of thought” where different perspectives (personality traits and domain expertise) diversify and debate possible solutions.
Using quantitative analysis and mechanistic interpretability on reasoning traces, the authors find these models show much higher perspective diversity and more conflict/reconciliation dynamics than standard instruction-tuned models.
They also report controlled RL experiments suggesting that rewarding accuracy can induce more conversational/internal-dialogue behaviors, and that adding conversational scaffolding can accelerate reasoning gains. 10825