Aspect-wise Pairwise Tournaments for Test-Time Selection
Summary
- Researchers introduce Aspect-wise Pairwise Tournaments (APT), a new method for test-time selection in large language models that uses single-elimination tournaments based on pairwise aspect comparisons.
- APT uses multiple off-the-shelf aspect verifiers specified in natural language, enabling users to steer candidate solution selection toward specific properties without retraining.
- Evaluated across five domains and three generator LLMs, APT outperforms binary multi-aspect approval, self-consistency, and reward model verification, achieving the largest gains on open-ended tasks.
Why it matters
APT improves test-time compute scaling for LLMs by enabling more precise candidate selection through natural language-specified aspect comparisons without requiring model retraining.