DeCert framework improves generative reward model justification quality

The DeCert framework evaluates the validity and consistency of natural language rationales in generative reward models (GenRMs). It enables preference supervision without requiring costly reference explanations by checking if rationales actually support the verdict.