Classification improves stepwise reasoning checks in language models
Stepwise Reasoning Verification: Don't Generate, Classify!
Artificial Intelligence
Summary
Checking the reasoning of large language models is important when they solve math and science problems. The authors find that instead of generating next steps, it works better to classify whether each new reasoning step logically follows from the previous steps. They test this idea on several datasets and show that smaller models using classification can outperform much larger models that generate answers. This means verifying reasoning can be done more accurately and efficiently without huge models.
What this means in practice
- •For ai system developers: Create verification tools that check each reasoning step of language models efficiently using smaller classification models.
- •For software quality teams: Improve error detection in automated reasoning by classifying reasoning step correctness rather than generating solutions.
Authors
Yifei Shi, Michalis Korakakis, Rami Aly, Eric Chamoun, Lukasz Szpruch, Andreas Vlachos
Abstract
Verifying the reasoning of large language models is increasingly important as they are applied to complex mathematical and scientific problems. Existing work focuses on large decoder models and reports mixed results for generative and discriminative verifiers, leaving unclear which paradigm is more effective and whether large models are necessary. In this work, we cast stepwise verification as a natural language inference task: the problem and partial solution prefix form the premise, while the next reasoning step serves as the hypothesis to be classified as entailed or non-entailed. This formulation enables error localisation at individual reasoning steps and direct comparison of encoder classification, decoder classification, and decoder verdict generation across model scales and annotation budgets. Across five reasoning datasets, a 0.15B encoder classifier outperforms most generative verifiers with up to 4B parameters. On identical decoder backbones, classification outperforms generation from 0.5B to 4B parameters and matches it on the strongest 7B backbone---reaching comparable performance with roughly one tenth as many annotated problems. Our findings establish classification as a parameter- and data-efficient approach to stepwise verification and show that strong verification does not require large generative decoders.