Paper Overview
Field: NLP Authors: Lecheng Kong, Like Hui, Haitao Mao Published: 2026-08-12 arXiv: 2508.05137
Abstract
Test-time scaling often uses an external verifier, such as compilers and test cases in coding or trained value functions in robotics applications, to obtain high-quality rollouts. Verifier-free test-time scaling (VF-TTS) is gaining extensive attention as a mechanism to enhance Large Language Model (LLM) reasoning, primarily because we do not have access to such high-quality verifiers in many real-world applications. Among existing VF-TTS methods, confidence-based VF-TTS methods, which compute and rank rollouts solely by confidence, are particularly promising. Such methods introduce near-zero overhead for sample evaluation and require minimal access to internal model states, making the methods highly flexible across models and tasks.
In this paper, the authors demonstrate a critical limitation of existing confidence-based VF-TTS methods: they catastrophically collapse on complex tasks. A notable observation is that uniformly high confidence typically indicates failed exploration, tending toward confidently wrong answers.
Key Idea
The core insight is that robust epistemic search requires a specific confidence trajectory pattern: methods should perform exploratory branching at the start (manifesting as low initial confidence) and converge to high final-confidence solutions.
Method: Consilience
To realize this insight, the authors introduce consilience, a novel selection framework that explicitly evaluates the temporal asymmetry of confidence during reasoning. It is implemented through a composite metric that:
- Actively penalizes high initial confidence
- Strictly requires final certainty
Results
Extensive experiments on graduate-level math problems and free-form code generation show that consilience effectively outperforms existing baselines, validating the novel perspective on completion confidence.
---
*Automatically collected on 2026-08-12*