KV-PRM

KV-PRM, a novel Process Reward Model that leverages the KV cache instead of text re-encoding, achieves up to a 5,000x reduction in scoring FLOPs, 37x lower latency, and 34x less memory per sequence in…

Researchers introduce KV-PRM, a process reward model that reduces scoring cost from O(L²) to O(L) by reading the KV cache from LLM generation instead of re-encoding trajectory text, achieving up to 5,…