VaSE proposes training-free KV-cache eviction for reasoning models

VaSE’s introduction describes it as a training-free approach to KV-cache eviction for reasoning models. Its creator says the method keeps large-magnitude value states and evicts stochastically to reduce the capability drop that can come from capping cache memory. A researcher said they planned to present the work at COLM 2026’s Efficient Reasoning Workshop on October 9.