Duel-Evolve: Reward-Free Test-Time Scaling via LLM Self-Preferences

TL;DR

Duel-Evolve is a reward-free, test-time optimization method for LLM outputs that uses pairwise preferences from the same LLM to guide iterative candidate refinement, achieving significant accuracy improvements without external rewards.

Contribution

It introduces Duel-Evolve, a novel evolutionary algorithm that leverages LLM self-preferences and Bayesian modeling for test-time optimization without external rewards or labels.

Findings

20 percentage points higher accuracy on MathBench

Over 12 percentage points improvement on LiveCodeBench

No need for reward models or ground-truth labels

Abstract

Many applications seek to optimize LLM outputs at test time by iteratively proposing, scoring, and refining candidates over a discrete output space. Existing methods use a calibrated scalar evaluator for the target objective to guide search, but for many tasks such scores are unavailable, too sparse, or unreliable. Pairwise comparisons, by contrast, are often easier to elicit, still provide useful signal on improvement directions, and can be obtained from the LLM itself without external supervision. Building on this observation, we introduce Duel-Evolve, an evolutionary optimization algorithm that replaces external scalar rewards with pairwise preferences elicited from the same LLM used to generate candidates. Duel-Evolve aggregates these noisy candidate comparisons via a Bayesian Bradley-Terry model, yielding uncertainty-aware estimates of candidate quality. These quality estimates guide allocation of the comparison budget toward plausible optima using Double Thompson Sampling, as well as selection of high-quality parents to generate improved candidates. We evaluate Duel-Evolve on MathBench, where it achieves 20 percentage points higher accuracy over existing methods and baselines, and on LiveCodeBench, where it improves over comparable iterative methods by over 12 percentage points. Notably, the method requires no reward model, no ground-truth labels during search, and no hand-crafted scoring function. Results show that pairwise self-preferences provide strong optimization signal for test-time improvement over large, discrete output spaces.