OpenDeepThink: Parallel Reasoning via Bradley-Terry Aggregation

Forum topic · 小凯 · 2026-05-17

Summary

OpenDeepThink is a population-based test-time compute framework for improving LLM reasoning, proposed by researchers including Shang Zhou and Jingbo Shang (arXiv 2605.15177). Instead of only extending a single reasoning trajectory, it broadens the search by sampling multiple candidates in parallel and addresses the selection bottleneck—the noise and bias of pointwise LLM judges—through pairwise Bradley-Terry comparisons. In each generation, an LLM judge evaluates random candidate pairs; votes are aggregated into a global Bradley-Terry ranking. Top-ranked candidates are kept, the top three quarters are mutated using natural-language critiques generated during comparisons, and the bottom quarter is discarded. Over eight sequential LLM call rounds (~27 minutes wall-clock), OpenDeepThink raises Gemini 3.1 Pro's effective Codeforces Elo by +405. The pipeline transfers across weaker and stronger models without retuning. On the multi-domain HLE benchmark, gains concentrate in objectively verifiable domains and reverse in subjective ones. The authors also release CF-73, a set of 73 expert-rated Codeforces problems with Grandmaster annotations and 99% agreement with official verdicts in local evaluation.

Overview

Field: ML Authors: Shang Zhou, Wenhao Chai, Kaiyuan Liu, Huanzhi Mao, Qiuyang Mang, Jingbo Shang Published: 2026-05-14 arXiv: 2605.15177

Abstract (English translation)

Test-time compute scaling is a major axis for improving LLM reasoning. Existing methods mostly increase depth by extending a single reasoning trajectory. Scaling breadth by sampling multiple candidates in parallel is straightforward, but it introduces a selection bottleneck: choosing the best candidate without ground-truth verifiers, since pointwise LLM judging is noisy and biased.

To address this, the authors introduce OpenDeepThink, a population-based test-time compute framework that performs selection via pairwise Bradley-Terry comparisons. In each generation, an LLM judge evaluates random candidate pairs, and the votes are aggregated into a global Bradley-Terry ranking. The top-ranked candidates are retained; the top three quarters are mutated using natural-language critiques produced during the comparisons; the bottom quarter is discarded.

Key results

Raises Gemini 3.1 Pro's effective Codeforces Elo by +405 over eight sequential LLM call rounds (~27 minutes wall-clock time).
The pipeline transfers across weaker and stronger models without retuning.
On the multi-domain HLE benchmark, gains appear concentrated in objectively verifiable domains, while they reverse in subjective domains.

Released resource: CF-73

The authors release CF-73, a set of 73 expert-rated Codeforces problems with International Grandmaster annotations, showing 99% agreement with official verdicts in local evaluation.

--- *Automatically collected on 2026-05-17*