ACCT: Asymmetric Cross-LLM Context Transfer

Researchers propose ACCT, an asymmetric cross-LLM context transfer method designed to reduce computational costs by reusing KV caches between models of different scales within the same family. ACCT uses three components—the Cartographer, Allocator, and Compiler—to transform source KV caches into sparse target states under a shared budget. Evaluations across multiple benchmarks show that the method matches or exceeds target inference performance while achieving significant speedups over standard prefill. Open article →