Unverified50% confidenceFactExact time
BCSA performs exact windowed attention on the most recent 4,096 tokens, pools more distant context into blocks at a 4:1 ratio, and then uses a learnable indexer to read the top 512 blocks.
1
Sources
50%
Confidence
Long-term
Relevance
10/5/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn large model API call scenarios, cache hit rate refers to the reuse ratio of the KV Cache (key-value cache), where the model can directly read the attention key-value pairs of already-processed tokens from the cache.63% similarUnverifiedSHADOW writes attention states to disk at 1-bit precision, consuming 288 bytes per token, with the index occupying only 22 bytes per token, record lookup taking approximately 1 microsecond, and attention states being reinjected into the model in approximately 0.03 ms, without using a vector database or embedding model.62% similarUnverified一个200K上下文窗口的请求，其KV Cache可能占据数十GB显存61% similarVerifiedThe context window is the maximum amount of information a large model can process in a single session, measured in tokens; modern top-tier models' context windows have expanded from the early 4K tokens to 200K tokens or even higher.60% similarUnverified当前长上下文窗口（100K-1M tokens）的普及使Agent能够维持复杂任务的长期记忆60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/977112API
curl https://kongchang.com/api/v1/knowledge/claims/977112MCP
get_claim(id=977112)