KVShareArena: KV-Cache Reuse Across Contexts and Model Checkpoints
KVShareArena tests cache reuse where prefix-only benchmarks do not: mixed prompt sources and same-family checkpoint changes.
The paper says reused KV caches can fail when retrieved chunks or agent reports are inserted inside a new prompt, because positions and attention history no longer match. Its benchmark measures how much each repair method recovers versus full recomputation, while also accounting for compute, memory, latency, and cache build cost. Position correction works until questions require several sources at once; then partial re-encoding or trained methods recover more, while unrepaired caches can underperform no cache. Cross-checkpoint reuse barely hurts training-free methods, but a checkpoint-specific adapter loses quality. ArXiv · AI/CL/LG's note
The paper says reused KV caches can fail when retrieved chunks or agent reports are inserted inside a new prompt, because positions and attention history no longer match. Its benchmark measures how much each repair method recovers versus full recomputation, while also accounting for compute, memory, latency, and cache build cost. Position correction works until questions require several sources at once; then partial re-encoding or trained methods recover more, while unrepaired caches can underperform no cache. Cross-checkpoint reuse barely hurts training-free methods, but a checkpoint-specific adapter loses quality. ArXiv · AI/CL/LG's note
score 5