LOVER: Unsupervised Verifier Boosts LLM Reasoning via Logical Rules
Researchers propose LOVER, an unsupervised verifier that uses logical rules to enhance reasoning in large language models without costly supervised data. The method treats the verifier as a binary latent variable, applying negation, intra-group, and inter-group consistency constraints across multiple reasoning paths. Experiments on 10 datasets show LOVER matches 95% of supervised verifier performance. Code is open-source.
Key facts
- LOVER is an unsupervised verifier regularized by logical rules.
- It treats the verifier as a binary latent variable.
- Three logical constraints: negation consistency, intra-group consistency, inter-group consistency.
- Groups reasoning paths by final answer.
- Outperforms unsupervised baselines on 10 datasets.
- Achieves 95% of supervised verifier performance on average.
- Compatible with any off-the-shelf LLM.
- Source code at https://github.com/wangx
Entities
—