Verifier-Guided Self-Consistency: Test-Time Reasoning with Imperfect Generators and Verifiers
Fadeeva et al.
A stochastic decoder steers generation with step-level and outcome signals while keeping several trajectories for a final vote, and can be read as approximate sampling from a verifier-tilted distribution.
Source published · In the Daily Brief of 6 November 2026
Filed under Reasoning and inference, Large language models
Test-time reasoning has to balance two imperfect signals: a generator that supplies diversity and a verifier that points towards correctness. Verifier-weighted voting cannot steer generation, because it only reweights trajectories already sampled. Beam-style search can steer, but it may concentrate on a narrow set of trajectories and amplify the verifier’s errors.
The method proposed here repeatedly draws candidate next steps from the base model and uses a step-level correctness estimate and a future-outcome estimate in an acceptance and resampling procedure, while maintaining multiple trajectories for answer-level self-consistency. The authors give a probabilistic interpretation as approximate sampling from a verifier-tilted trajectory distribution, and report higher average accuracy than representative baselines across several benchmarks, generators and verifier families.