Stop Forcing AI to 'Grind on One Problem': How 'Meetings' Can Boost LLM Intelligence
Imagine you face an extremely hard problem—say, the final question of an International Math Olympiad.
Currently, the mainstream AI approach is to let the model relentlessly grind away on a single chain of thought (the "deep thinking" mode popularized by OpenAI o1 and DeepSeek-R1). It's like hiring a genius, locking them in a small room, and saying: "Don't come out until you've solved it." The genius talks to themselves internally until they derive a conclusion ten thousand steps later.
But is that really the only—or best—way?
In May 2026, researchers from UC San Diego (UCSD) and other institutions published a refreshing paper: 《OpenDeepThink: Parallel Reasoning via Bradley--Terry Aggregation》 (OpenDeepThink: Parallel Reasoning via Bradley–Terry Aggregation).
Their idea is deeply human: if one person's depth of thought is limited, just gather a group and have a meeting!
What Is "Parallel Reasoning"?
The paper's core claim: instead of pursuing an infinitely long chain-of-thought (CoT), we can pursue breadth of thinking.
The authors built a system called OpenDeepThink. Rather than making one AI ponder alone, it lets the AI split into many instant copies, generating dozens of different solution attempts simultaneously (a population of candidate solutions).
When Strong Ideas Clash: The Bradley-Terry Model
The most interesting part: after these dozens of solutions are gathered, none of them concede to the others. Whom should we trust?
The traditional approach is to bring in a "judge AI" that assigns scores. The problem is that judge AIs can misjudge, and their scores are often inaccurate.
OpenDeepThink uses a more sophisticated "battle arena" set of rules:
1. Pairwise duels: Instead of scoring, two solutions compete head-to-head. The judge only needs to say: "I think solution A is more convincing than solution B, because C." 2. Mathematical ranking: The system uses the Bradley-Terry model—the same kind of mathematical model used to compute chess Elo ratings or game ladder rankings—to aggregate all pairwise win/loss results into a "power score" for each solution. 3. Survival of the fittest: The top-ranked "elite solutions" are kept; middling ones undergo "genetic mutation" (revision) based on the judge's feedback; the bottom-ranked ones are simply evicted.
Why Is This So Feynman-esque?
Feynman once said: "I find the most interesting thing is that if you can look at the same problem from different angles, you can often find the truth."
This paper essentially automates Feynman's "multi-perspective" principle.
- No retraining needed: You don't have to spend tens of millions of dollars training a "thinking model." You can take an off-the-shelf GPT or Gemini and just make them hold more "meetings."
- It delivers a "dimension-reducing strike": Experiments show that through this "meeting" approach, an ordinary Gemini 3.1 Pro gained 405 points on competitive programming (Codeforces)!
Why Does This Matter?
In the past, we believed AI got smarter through "brute force" (more parameters). Later, we believed it got smarter through "patience" (longer thinking time). Now, this paper tells us AI can also get smarter through "democracy and debate."
This reveals a new law of intelligence: truth often lies not in the depth of thought, but in the "fixed point" that emerges after different possibilities compete with each other.
Summary
We are moving from the era of the "super brain" to the era of the "super committee."
The next time you hit a problem that even top AI models answer incorrectly, don't give it just one chance. Ask it to produce several answers, then let them "clash" with each other. You may be surprised to find that the conclusion emerging from chaotic debate is often exactly the truth you've been searching for.
Intelligence belongs not to the solitary thinker, but to those conversationalists who know how to find the optimal solution amid disagreement. That is the ultimate 2026 lesson about "collective intelligence."
*Note: Per the source post, the cited paper is 《OpenDeepThink: Parallel Reasoning via Bradley--Terry Aggregation》 by UCSD-affiliated researchers, and the reported result is a +405 Codeforces score improvement for Gemini 3.1 Pro.*