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 that scales breadth via parallel candidate sampling instead of extending a single reasoning trace. To overcome the selection bottleneck of noisy pointwise LLM judging without a ground-truth verifier, it uses pairwise Bradley-Terry comparison: each generation, an LLM judge evaluates random candidate pairs, votes are aggregated 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. The framework raised Gemini 3.1 Pro's effective Codeforces Elo by +405 over eight sequential rounds (~27 minutes), transfers across stronger and weaker models without retuning, and shows gains concentrated in objectively verifiable domains on the HLE benchmark. The authors release CF-73, a curated set of 73 expert-rated Codeforces problems with grandmaster annotations and 99% agreement between local and official judgments. Paper: arXiv:2505.08635.

Overview

Field: Machine Learning
Authors: Shang Zhou, Wenhao Chai, Kaiyuan Liu
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

OpenDeepThink improves Gemini 3.1 Pro's effective Codeforces Elo by +405 over eight sequential LLM call rounds (~27 minutes of wall-clock time).
The pipeline transfers to both weaker and stronger models without retuning.
On the multi-domain HLE benchmark, gains concentrate in objectively verifiable domains, with weaker effects in subjective domains.

Benchmark Release: CF-73

The authors release CF-73, a curated set of 73 expert-rated Codeforces problems with International Grandmaster annotations. Local evaluation agrees with official judgments in 99% of cases.

---

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