DeepMind's GenRM improves LLM accuracy by having models verify their own outputs

DeepMind's GenRM trains LLMs to verify responses based on next-token prediction and chain-of-thought (CoT) reasoning.