itstargetnews
A new Google study suggests that advanced reasoning models can achieve high performance by simulating multi-agent-like debates involving different perspectives, personality traits, and domain expertise.
Their experiments show that this internal debate, which they call “society of thought“Significantly improves model performance in complex reasoning and planning tasks. Researchers found that leading reasoning models such as DeepSeek-R1 and QwQ-32B, trained through reinforcement learning (RL), innately developing this ability to engage socially in mental conversations without explicit instruction.
These findings offer a road map for how developers can build more robust LLM applications and how businesses can train better models using their own internal data.
What is the society of thought?
A core premise of cognitive society is that reasoning models can learn to simulate social, multi-agent dialogues to refine their logic. This assumption draws on cognitive science, especially the idea that human reason is primarily a social process to solve problems through argument and engagement with different perspectives.
The researchers wrote that "Diversity of thinking, which comes from diversity of skills and personality traits, improves problem solving, especially when accompanied by genuine opposition." Consequently, they suggest that combining different perspectives allows LLMs to develop strong reasoning strategies. By simulating conversations between different internal personas, models can perform important checks (such as verification and backtracking) that help avoid common pitfalls such as unwanted biases and sycophancy.
In models like the DeepSeek-R1, it is "COMMUNITY" directly reflects the content of the chain of thought. The researchers note that you don’t need separate models or prompts to force this interaction; the debate emerges autonomously within the reasoning process of a model example.
Examples of cognitive society
The study provides tangible examples of how this internal friction leads to better results. In an experiment involving a complex organic chemistry synthesis problem, DeepSeek-R1 simulates a debate among many different internal perspectives, including a "Planner" and a "Critical Verifier."
Planner initially suggests a standard reaction path. However, the Critical Verifier (described as having high conscientiousness and low agreeableness) is distracted from challenging the hypothesis and providing a counterargument with new facts. Through this adversarial check, the model discovers the error, reconciles conflicting views, and corrects the synthesis path.
A similar dynamic appears in creative works. When asked to rewrite the sentence, "I cast my hatred into the burning fire," the model simulates a negotiation between a "Creative Ideator" and a "Semantic Fidelity Checker." After the ideator proposed a version using the word "deep seated," the examiner replied, "But that adds ‘deep-seated,’ which was not in the original. We must avoid adding new ideas." The model eventually settled on a compromise that kept the original meaning while improving the style.
Perhaps the most remarkable evolution occurred in the "counting game," a math puzzle where the model must use certain numbers to reach a target value. Early in training, the model attempts to solve the problem using a monologue approach. As it is known through RL, it is strongly divided into two distinct personas: a "Methodical Problem-Solution" make calculations and a "Inquisitive Thinker" monitor the progress, who will interrupt the failed paths with comments like "Again no luck… Maybe we can try using negative numbers," prompting the Methodical Solver to switch strategies.
These findings challenge the assumption that longer thought chains automatically result in higher accuracy. Instead, different behaviors such as looking at answers through different lenses, confirming early assumptions, backtracking, and exploring alternatives, drive improvements in reasoning. The researchers reinforced this by artificially driving the model’s activation space to induce conversational surprise; this intervention activates a wider range of personality- and skill-related areas, doubling the accuracy of complex tasks.
The implication is that social reasoning emerges autonomously through RL as a function of the driving model to produce correct responses, rather than through explicit human supervision. In fact, training models of monologues do not generate raw RL that naturally develops multi-agent conversations. On the contrary, performing closely monitored-tuning (SFT) in multi-party conversations, and the debate far surpasses SFT in standard thought chains.
Implications for business AI
For developers and business decision makers, these insights offer practical guidelines for building more powerful AI applications.
Slow engineering for ‘conflict’
Developers can improve the reasoning of general-purpose models by explicitly prompting them to adopt a social structure of thought. However, it is not enough to simply ask the model to chat with herself.
"It is not enough to ‘have a debate’ but to have different views and dispositions that make debate inevitable and allow that debate to examine and discriminate between alternatives," James Evans, co-author of the paper, told VentureBeat.
Instead of generic roles, developers should design prompts that assign conflicting dispositions (eg, a risk-averse compliance officer versus a growth-focused product manager) to force the model to bias alternatives. Even the simple hints that guide the expression model "surprise" can trigger these higher reasoning paths.
Design for social scaling
As developers scale the test-time compute to allow models "THINK" longer, they must construct this time as a social process. Applications should facilitate a "CORPORATELY" process where the model uses pronouns such as "we," asks his own questions, and clearly debates the alternatives before coming together on an answer.
This approach can also be extended to multi-agent systems, where different personalities assigned to different agents participate in critical debate to reach better decisions.
Stop cleaning your training data
Perhaps the most important implications lie in how companies train or refine their own models. Traditionally, data teams scrubbed their datasets to create "Golden Answers" which provides perfect, linear paths to a solution. The study suggests that this may be a mistake.
Models well adapted to conversational data (eg, transcripts of multi-agent debate and resolution) improve reasoning much faster than those trained on pure monologues. There is even value in debates that do not lead to a correct answer.
"We trained the conversational scaffolding that led to the wrong answer, then strengthened the model and found that it performed as well as reinforced the correct answer, suggesting that conversational behaviors in exploring solutions are most important for new problems," Evans said.
This means that businesses should stop dumping "troubled" engineering logs or Slack threads where problems are solved iteratively. the "DISPUTE" where the model learns the exploratory behavior.
Exposing the ‘black box’ for trust and auditing
For high-stakes business use cases, simply getting an answer is not enough. Evans argued that users should see internal resistance to trust output, suggesting a shift in user interface design.
"We need a new interface that systematically reveals the internal debates to us so that we can ‘participate’ in calibrating the right answer," Evans said. "We are better at debate; AIs are better at debate; and we do better when exposed to the AI debate."
The strategic case for open weights
These findings provide a new argument to "build versus buy" debate about open-weight models versus proprietary APIs. Many proprietary reasoning models hide their chain-of-thought, treating internal debate as a trade secret or a safety liability.
But Evans argued that "no one has given a reason to expose this society of thought before," but that the value of auditing these internal conflicts has become undeniable. As long as proprietary providers offer full transparency, businesses in high-compliance sectors may find that open-weight models offer a different advantage: the ability to see the objection, not just the decision.
"I believe that the big, proprietary models will start serving (and licensing) the information once they realize there is value in it," Evans said.
Research suggests that the job of an AI architect is shifting from pure model training to something closer to organizational psychology.
"I believe this opens up a whole new frontier of small group and organizational design within and between models that will likely enable new types of performance," Evans said. "My team is working on it, and I hope others are too."
