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, presented by researchers including Shang Zhou and Jingbo Shang (arXiv:2605.15177). Instead of only extending a single reasoning trajectory, it scales breadth by sampling multiple candidate solutions in parallel and addresses the selection bottleneck using pairwise Bradley-Terry comparisons rather than noisy pointwise LLM judging. In each generation round, an LLM judge evaluates random candidate pairs, votes are aggregated into a global Bradley-Terry ranking, top candidates are retained, 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 time), OpenDeepThink raised Gemini 3.1 Pro's effective Codeforces Elo by +405 points. The pipeline transfers across weaker and stronger models without retuning, and gains on the multi-domain HLE benchmark concentrate in objectively verifiable domains, reversing in subjective ones. The authors also release CF-73, a benchmark of 73 expert-graded Codeforces problems with Grandmaster annotations and 99% agreement with official verdicts.

Paper Overview

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

Abstract (translated from the post)

Test-time compute scaling is a major axis for improving LLM reasoning. Existing methods mostly scale 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 a ground-truth verifier, 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 round, an LLM judge evaluates random pairs of candidates, and the votes are aggregated through Bradley-Terry into a global ranking. The top-ranked candidates are retained; the top three-quarters are mutated using the natural-language critiques produced during the comparisons; the bottom quarter is discarded.

Key results

Over eight sequential rounds of LLM calls (~27 minutes wall-clock time), OpenDeepThink improves Gemini 3.1 Pro's effective Codeforces Elo by +405 points.
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.
The authors release CF-73, a set of 73 expert-graded Codeforces problems with Grandmaster annotations and 99% agreement with official verdicts in local evaluation.

---

*Auto-collected on 2026-05-17 from zhichai.net.*