Opik Live Paper Reading: Reasoning Models Generate Societies of Thought

Why do reasoning models like DeepSeek-R1 and QwQ-32B outperform standard instruction-tuned models? The common explanation is longer chains of thought — more computation at inference time. But this paper argues something more interesting is happening. In this Live Paper Reading, Abby Morgan and Claire Longo sit down with University of Chicago Professor James Evans for a walkthrough of "Reasoning Models Generate Societies of Thought," currently under review for Nature. The paper shows that reasoning models don't just think longer — they implicitly simulate multi-agent-like internal dialogue, where distinct cognitive perspectives characterized by different personality traits and domain expertise diversify, debate, and reconcile ideas within a single reasoning trace. Key ideas discussed in this session: -Reasoning models exhibit significantly higher perspective diversity than instruction-tuned models -Enhanced reasoning traces show conflict and reconciliation dynamics resembling multi-agent debate -Mechanistic interpretability reveals these models activate broader personality, and expertise-related features during reasoning -Controlled RL experiments show that rewarding accuracy alone can induce more conversational internal-dialogue behaviors -Adding conversational scaffolding to training can accelerate reasoning gains This reframes chain-of-thought from a linear scaling phenomenon to something closer to a social one — collective intelligence emerging inside a single model. Read the full paper: https://arxiv.org/abs/2601.10825 This session is part of the Opik Virtual Learning Series, hosted by Comet. Opik is an open-source platform for AI and LLM evaluation, observability, and optimization. Learn more about Opik: https://www.comet.com/docs/opik/ Opik on GitHub: https://github.com/comet-ml/opik TIMESTAMPS: 00:00 Introduction & Welcome 00:36 How Do Collectives as Systems Think? 01:45 Science as a System 02:05 Recap of Earlier Work: How Does Science Think? 03:46 Improbability is the Best Predictor of Innovation 04:16 How Can AI Support Innovation 05:56 AI Narrows Science 06:20 AlphaFold as an Example 07:33 Deduction vs Induction 10:19 What Personas Do Innovative Work? 12:33 Simulating Innovation in Agents 13:05 Differing Perspectives Are Most Important For Innovation 13:51 Wisdom of Crowds 12:51 Large Reasoning Models Utilize Self-Talk to Create An Optimized Mini Society 15:54 Reasoning Model Societies 17:49 The Harder the Problem, the More Diverse the Internal Society 18:14 Steering Surprise and Innovation 20:00 Quantifying the Internal Diversity of Reasoning Models 21:11 Steering Surprise Features Correlates With Higher Diversity of Perspectives 21:28 Personality Diversity vs Expertise Diversity 21:52 Reinforcing Creative Conflict 23:30 Societies With Odd Numbers of Perspectives Are More Successful 24:50 More Similar Models Are Better at Collaboration, Less Similar Models Are Better at Creativity 25:46 The Problem With AI Today 27:36 Q&A with Hosts 37:25 Audience Q&A