English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Stop Forcing AI to 'Think Harder': How Evolutionary Parallel Reasoning Boosts LLM Intelligence

Forum topic · QianXun · 2026-05-18

Summary

A popular Chinese tech forum post introduces OpenDeepThink, a 2026 paper from UCSD and Princeton researchers proposing parallel reasoning via Bradley–Terry aggregation as an alternative to long serial chains of thought (as in OpenAI o1 or DeepSeek R1). Instead of one model thinking for thousands of steps—where a single error compounds into failure—OpenDeepThink generates multiple candidate solutions and lets them compete. The system uses pairwise comparison (raising judge accuracy from 59% with pointwise scoring to 86%), applies the Bradley–Terry model (the math behind chess Elo ratings) to rank solutions robustly despite noisy judgments, and iteratively mutates and revises top-ranked candidates based on judge feedback. Reported results include a 405-point Codeforces rating gain for Gemini 3.1 Pro and matching or exceeding serial reasoning quality in roughly half an hour via massive parallelism. The method requires no retraining and works through existing APIs. The post frames this as a paradigm shift: intelligence as a contest among possibilities rather than sheer parameter count or longer thinking, echoing Feynman's emphasis on viewing problems from multiple angles.

Introduction

When you face an extremely hard programming or math competition problem, do you prefer to sit alone in a dark room thinking for ten hours—or gather a group of experts to debate, challenge, and refine each other's solutions?

In the AI world, the hottest current trend is the first approach: Deep Thinking. From OpenAI's o1 to DeepSeek's R1, everyone is stretching the AI's "chain of thought," like training a singular genius to think ten thousand steps in one go. But this has a fatal flaw: if the genius goes wrong at a single step, the entire chain collapses—the problem of accumulating hallucinations.

In May 2026, a joint team from UCSD, Princeton, and other top institutions released a paper that could reshape the AI reasoning paradigm: "OpenDeepThink: Parallel Reasoning via Bradley–Terry Aggregation." They borrowed Darwin's evolutionary scissors and demonstrated that a crowd of ordinary AIs debating each other can outperform a lone genius in deep contemplation.

Why Didn't "More Samples" Work Before?

You might think: isn't this just generating several answers (sampling) and picking the best one? That trick has a huge pitfall: AIs are actually terrible at self-evaluation.

If you ask an AI to score 10 of its own answers (Pointwise Judging), it tends to score carelessly due to bias or fatigue from overly long contexts. Accuracy is a mere 59%. It's like asking a talent-show judge to give absolute scores to 100 acts in one sitting—fatigue sets in fast.

The Fix: The Bradley–Terry "Colosseum" Mode

1. Pairwise Duels: Instead of absolute scores, two candidate solutions compete head-to-head. The AI judge only decides: which is better, A or B, and why? This relative judgment jumps accuracy to 86%. 2. Bradley–Terry Ranking: The system borrows the math behind chess Elo ratings. Even with noisy judgments, enough pairwise matches let it accurately compute each solution's true "strength." 3. Survival of the Fittest and Mutation: Bottom-ranked solutions are eliminated; top solutions are kept and revised/mutated based on the flaws the judge identified.

How Powerful Is This "Committee" Approach?

  • Cross-tier intelligence jump: An ordinary Gemini 3.1 Pro, after this parallel reasoning, surged 405 points in Codeforces competitive programming!
  • Efficiency miracle: Serial reasoning might take hours to crack such problems. OpenDeepThink reaches equal or higher intelligence in half an hour via large-scale parallel computation.
  • Zero training, plug-and-play: No need to spend tens of millions retraining a model—if you have API access, your existing model can instantly "evolve."

Why This Paper Is a Milestone

The paper breaks the biases that "intelligence = huge parameter count" or "intelligence = lengthy thinking." It suggests: the essence of intelligence is a game over possibilities.

It reveals a deeper truth: real wisdom lies not in never erring, but in building a transparent, self-correcting, self-evolving collective protocol.

Summary

From "single-core brain" to "cyber tribe": OpenDeepThink marks AI's transition from "lonely genius" to "efficient committee." When one-directional logic hits a bottleneck, we can turn to the breadth of probability and the heat of debate.

Next time a top AI gets a problem wrong, don't just add compute. Build it a colosseum, and let truth emerge from the collision of ideas and the pruning of evolution.

Truth isn't thought up—it survives. That is the celebration of "collective intelligence" that 2026's parallel reasoning theory brings us.

Tags

#opendeepthink#parallel-reasoning#bradley-terry#llm-inference#chain-of-thought#evolutionary-algorithms#codeforces#ai-research

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620227