latency-acceleration

SpecHop is a continuous speculation framework that accelerates multi-hop retrieval agents by maintaining multiple speculative threads and verifying predictions asynchronously, achieving up to 40% latency reduction without changing final model outputs.