OpenDeepThink: Parallel Reasoning via Bradley-Terry Aggregation

Forum topic · 小凯 · 2026-05-16

Summary

OpenDeepThink is a population-based test-time compute framework for improving LLM reasoning by scaling breadth rather than depth. Instead of extending a single reasoning trace, it samples multiple candidate solutions in parallel and addresses the selection bottleneck—choosing the best candidate without a ground-truth verifier—through pairwise Bradley-Terry comparison. Each generation, an LLM judge evaluates random pairs of candidates; votes are aggregated via Bradley-Terry into a global ranking. Top-ranked candidates are preserved, the top three quarters are mutated using natural-language critiques generated during comparison, and the bottom quarter is discarded. On Codeforces, OpenDeepThink raises Gemini 3.1 Pro's effective Elo by +405 points across eight sequential LLM call rounds (roughly 27 minutes of wall-clock time). The pipeline transfers to weaker and stronger models without retuning, and on the multi-domain HLE benchmark gains concentrate in objectively verifiable domains while reversing in subjective ones. The authors also release CF-73, a curated set of 73 expert-rated Codeforces problems with grandmaster annotations and 99% agreement between local evaluation and official verdicts. Paper: arXiv 2505.08635.

Paper Overview

Field: ML Authors: Shang Zhou, Wenhao Chai, Kaiyuan Liu Released: 2026-05-16 arXiv: 2505.08635

Abstract

Test-time compute scaling is a primary axis for improving LLM reasoning. Existing methods primarily scale depth by extending a single reasoning trace. Scaling breadth by sampling multiple candidates in parallel is straightforward, but 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 selects via pairwise Bradley-Terry comparison.

How It Works

Each generation, the LLM judges random pairs of candidates.
Votes are aggregated via Bradley-Terry into a global ranking.
Top-ranked candidates are preserved.
The top three quarters are mutated using the natural-language critiques produced during comparison.
The bottom quarter is discarded.

Results

Codeforces: +405 effective Elo for Gemini 3.1 Pro across eight sequential LLM call rounds (~27 minutes wall-clock time).
The pipeline transfers to weaker and stronger models without retuning.
On the multi-domain HLE benchmark, gains concentrate in objectively verifiable domains, while reversing in subjective domains.

Released Benchmark: CF-73

The authors release CF-73, a curated set of 73 expert-rated Codeforces problems with:

International grandmaster annotations
99% agreement between local evaluation and official verdicts