Reasoning Models Generate Societies of Thought
2026/01/15 by Junsol Kim, Shiyang Lai, Nino Scherrer +2 · 17 voices
#cs.CL #cs.CY #cs.LG
paper · pdf
Abstract
Large language models have achieved remarkable capabilities across domains, yet mechanisms underlying sophisticated reasoning remain elusive. Recent reasoning models outperform comparable instruction-tuned models on complex cognitive tasks, attributed to extended computation through longer chains of thought. Here we show that enhanced reasoning emerges not from extended computation alone, but from simulating multi-agent-like interactions -- a society of thought -- which enables diversification and debate among internal cognitive perspectives characterized by distinct personality traits and domain expertise. Through quantitative analysis and mechanistic interpretability methods applied to reasoning traces, we find that reasoning models like DeepSeek-R1 and QwQ-32B exhibit much greater perspective diversity than instruction-tuned models, activating broader conflict between heterogeneous personality- and expertise-related features during reasoning. This multi-agent structure manifests in conversational behaviors, including question-answering, perspective shifts, and the reconciliation of conflicting views, and in socio-emotional roles that characterize sharp back-and-forth conversations, together accounting for the accuracy advantage in reasoning tasks. Controlled reinforcement learning experiments reveal that base models increase conversational behaviors when rewarded solely for reasoning accuracy, and fine-tuning models with conversational scaffolding accelerates reasoning improvement over base models. These findings indicate that the social organization of thought enables effective exploration of solution spaces. We suggest that reasoning models establish a computational parallel to collective intelligence in human groups, where diversity enables superior problem-solving when systematically structured, which suggests new opportunities for agent organization to harness the wisdom of crowds.
Cited by
Discussions
- Very interesting paper! I've been thinking about similar things for a bit now, wondering if there's some limit to how large a single 'mind' can get, requiring horizontal scaling, either externally (li [bsky, 16 points, 0 comments]
- this specific paper completely vindicated [bsky, 7 points, 0 comments]
- There's some research indicating that reasoning models learn to do this in thought traces, and that's part of what makes them effective! arxiv.org/abs/2601.108... [bsky, 6 points, 1 comments]
- Reasoning Models Generate Societies of Thought [hn, 3 points, 0 comments]
- Explore our full paper on how reasoning models like OpenAI’s o-series, DeepSeek-R1, and QwQ reason via internal argument: arxiv.org/abs/2601.10825 [bsky, 3 points, 0 comments]
- This new paper from researchers at Google, Chicago, Santa Fe, "Reasoning Models Generate Societies of Thought" is truly remarkable and wild. Models are getting really good at reasoning but not because [bsky, 3 points, 1 comments]
- Despite the anthropomorphizing language this is fascinating. Paper: Reasoning Models Generate Societies of Thought ( www.arxiv.org/abs/2601.10825 ) And can someone neologize a better word for anthropo [bsky, 2 points, 0 comments]
- Resonates with Mercier & Sperber's social origins of reason and @santafe.edu complexity research on collective intelligence. Proud to pursue this with Google's Paradigms of Intelligence, @knowledgelab [bsky, 2 points, 1 comments]
- They do cite Minsky in their fuller paper: arxiv.org/abs/2601.10825 [bsky, 1 points, 2 comments]
- A study reveals that reasoning models can simulate a 'society of thought' to enhance cognitive diversity, leading to improved problem-solving. These findings suggest that organized diversity in AI rea [bsky, 1 points, 0 comments]
- #AI #LLM #Reasoning New arXiv shows top reasoning models behave like internal debate teams. Steering a “conversational surprise” feature (SAE 30939) doubled Countdown accuracy from 27.1% to 54.8%, boo [bsky, 1 points, 0 comments]
- we doing Societies of Thought now www.arxiv.org/abs/2601.10825 [bsky, 0 points, 0 comments]
- 4. Die Psychologie der LLM: Amodei betrachtet die innere Struktur der LLM als Sammlungen von Persönlichkeiten mit eigener Psychologie und das korrespondiert mit dem DeepMind-Paper "Reasoning Models Ge [bsky, 0 points, 1 comments]
- The Republic of Letters anno 2026: Reasoning LLM Models Generate Societies of Thought arxiv.org/pdf/2601.10825 [bsky, 0 points, 0 comments]
- Example 4: Reasoning models generate societies of thought. A cool paper that argues that reasoning models roleplay different personas in their chain of thought, which helps them solve more difficult p [bsky, 0 points, 1 comments]
- Reasoning Models Generate Societies of Thought arxiv.org/abs/2601.10825 [bsky, 0 points, 0 comments]
- "Enhanced reasoning emerges not from extended computation alone, but from multi-agent [LLM] interactions - a society of thought - which enables diversification and debate among internal cognitive pers [bsky, 0 points, 0 comments]
Related