Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering

Researchers introduced CorVer, a lightweight corpus-grounded reward for reinforcement learning in factual question answering.

It uses Wikipedia co-occurrence statistics to provide sentence-level feedback and token-level advantages with little computation.

Across multiple models and benchmarks, CorVer outperformed the raw baseline in every test setting.

It also beat neural-verifier methods in most comparable cases while training much faster.