Reasoning models may be doing more than extending their chains of thought. A study described by researchers from Google, the University of Chicago, and the Santa Fe Institute found that models including Deepseek-R1 and QwQ-32B can produce a kind of internal exchange, where different simulated perspectives question, correct, and refine the answer.
The researchers call this pattern a "society of thought". The phrase matters because it frames AI reasoning as something closer to a structured argument than a single uninterrupted calculation.
Why internal disagreement matters
Standard language models often move through a problem in a direct sequence. The study suggests that stronger reasoning models behave differently on complex tasks: they pause, shift perspective, raise objections, and revise their own path.
That distinction is important for understanding AI reasoning. If a model can catch its own mistake during generation, the output may improve not because the model has more facts, but because its process includes checks against its own assumptions.
The researchers analyzed over 8,000 reasoning problems and compared reasoning models with standard instruction-tuned models. Deepseek-R1 showed more question-answer sequences and more frequent shifts in perspective than Deepseek-V3. QwQ-32B showed more explicit conflicts between viewpoints than Qwen-2.5-32B.
What the researchers observed
The team used an LLM-as-judge method, with Gemini 2.5 Pro classifying reasoning traces. The source says agreement with human raters was substantial, giving the researchers a way to compare these internal patterns across models and tasks.
One example came from a complex multi-stage Diels-Alder synthesis. Deepseek-R1 moved between perspectives and challenged its own reasoning. At one point, it wrote, "But here, it's cyclohexa-1,3-diene, not benzene," identifying a mistake while still working through the problem.
Deepseek-V3 handled the same kind of work differently. It followed what the study described as a "monologic sequence", did not second-guess itself in the same way, and reached the wrong answer.
That contrast captures the central finding: the stronger model did not merely produce more text. It produced a process with internal friction, where one part of the reasoning appeared to push back against another.
Diverse simulated voices
The study also examined the kinds of perspectives that appeared inside the reasoning process. Deepseek-R1 and QwQ-32B showed higher personality diversity than instruction-tuned models across the Big Five dimensions: Extraversion, Agreeableness, Conscientiousness, Neuroticism, and Openness.
There was one notable exception. Diversity was lower for conscientiousness, with the simulated voices appearing disciplined and diligent. The authors connect this to research on team dynamics: variation in socially oriented traits such as extraversion and neuroticism can help group performance, while variation in task-oriented traits such as conscientiousness can hurt it.
A creative writing task offered a clearer view of how these roles can appear. The LLM-as-judge identified seven perspectives in Deepseek-R1's chain of thought, including a "creative ideator" with high openness and a "semantic fidelity checker" with low agreeableness. That checker objected with the line: "But that adds 'deep-seated' which wasn't in the original."
In plain terms, the model appeared to stage both invention and critique. One perspective proposed changes; another guarded fidelity to the prompt. The value came from their tension.
Tests beyond observation
The researchers also tested whether the conversation-like pattern was connected to better performance. Using a technique from mechanistic interpretability, they examined which internal features a model activates. In Deepseek-R1-Llama-8B, they found a feature tied to conversational signals such as surprise, realization, or acknowledgment.
When they artificially boosted this feature during text generation, accuracy on a math task doubled from 27.1 to 54.8 percent. The models also checked intermediate results more often and caught their own mistakes more frequently.
Controlled reinforcement learning experiments pointed in a similar direction. Base models "spontaneously increase conversational behaviours" when rewarded for accuracy, even without explicit training on dialogue structures.
The effect was stronger in models previously trained with dialogue-like thought processes. In Qwen-2.5-3B, dialogue-trained models reached about 38 percent accuracy after 40 training steps, while monologue-trained models stalled at 28 percent.
The dialogue-like structure also transferred beyond math. Models trained on math problems with simulated multi-perspective discussions learned faster even when detecting harmful or toxic content.
What this suggests for future AI
The authors compare the findings with research on collective intelligence in human groups. Mercier and Sperber's "Enigma of Reason" theory argues that human thinking evolved primarily as a social process. Bakhtin's "dialogical self" describes thought as an internalized conversation among perspectives.
The study does not claim that reasoning traces are literally discussions among simulated human groups. It also does not claim they are definitely a single mind imitating multi-agent interaction. The point is narrower: structured diversity inside a model's reasoning may help problem-solving.
That conclusion sits beside a more cautious view of reasoning models. In the summer of 2025, Apple researchers raised doubts about the "thinking" capabilities of reasoning models. Their study said models like Deepseek-R1 break down as problem complexity increases and reason less, which they called a "fundamental scaling limit."
Other studies have reached similar conclusions, though the finding remains controversial. Taken together, the work suggests that the next question is not only whether AI models can reason, but what kinds of internal structure make their reasoning more reliable.