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—sampling multiple candidate solutions in parallel—rather than only extending a single reasoning trace. It addresses the selection bottleneck of choosing the best candidate without a ground-truth verifier, where pointwise LLM judging tends to be noisy and biased. Each generation, the LLM judge evaluates random pairs of candidates and aggregates pairwise votes via a Bradley-Terry model into a global ranking. Top-ranked candidates are preserved; the top three quarters are mutated using natural-language critiques generated during the pairwise comparison, while the bottom quarter is discarded. Over eight sequential LLM call rounds (about 27 minutes wall-clock), OpenDeepThink improves Gemini 3.1 Pro's effective Codeforces Elo by +405 points. The pipeline transfers to both weaker and stronger models without retuning, and on the multi-domain HLE benchmark, gains concentrate in objectively verifiable domains while shrinking on subjective ones. The authors also release CF-73, a curated set of 73 expert-rated Codeforces problems with Grandmaster annotations and 99% local-evaluation agreement with official verdicts. Paper: arXiv 2505.08635.

Paper Overview

Field: Machine Learning 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. Each generation, the LLM judges random pairs of candidates and aggregates votes via Bradley-Terry into a global ranking; top-ranked candidates are preserved and the top three quarters are mutated using the natural-language critiques produced during comparison; the bottom quarter is discarded.

Key Results

Codeforces: +405 points of effective Elo for Gemini 3.1 Pro over eight sequential LLM call rounds (~27 minutes wall-clock time).
Transferability: The pipeline moves to weaker and stronger models without retuning.
HLE benchmark: Gains concentrate in objectively verifiable domains and diminish on subjective domains.
CF-73 dataset released: 73 expert-rated Codeforces problems with International Grandmaster annotations; 99% agreement between local evaluation and official verdicts.

Full Abstract (source)

> 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, we introduce OpenDeepThink, a population-based test-time compute framework that selects via pairwise Bradley-Terry comparison. Each generation, the LLM judges random pairs of candidates and aggregates votes via Bradley-Terry into a global ranking; top-ranked candidates are preserved and the top three quarters are mutated using the natural-language critiques produced during comparison; the bottom quar... (truncated in source)

Paper link: <https://arxiv.org/abs/2505.08635>

--- *Auto-collected on 2026-05-16*