Unverified50% confidenceFactExact time
CondenseVLA uses a set of learnable queries and multi-head cross-attention to distill the observation history into a fixed number of slot tokens.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLlama 3系列采用了GQA(Grouped Query Attention)设计以压缩KV Cache内存占用61% similarUnverified持久记忆系统本质上是RAG范式在对话历史领域的专项深化，检索源从静态知识库变为动态累积的用户交互记录60% similarUnverifiedQuery auditing records and analyzes each account's query history to identify accounts whose query distribution is abnormally concentrated, covers a specific task space, and whose input construction is systematic.60% similarUnverifiedMQAR is the Multi-Query Associative Recall task, a commonly used synthetic benchmark for evaluating the long-context information retrieval capabilities of efficient attention/state space models.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1073788API
curl https://kongchang.com/api/v1/knowledge/claims/1073788MCP
get_claim(id=1073788)