Unverified50% confidenceBenchmarkExact time
Experiments with MASA across multiple sparse attention methods, multiple benchmarks, and different LLM backbone models show consistent accuracy improvements.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/9/2026
First Seen
Valid until: 1/7/2027
Sources
Related Entities
Related Claims
UnverifiedAblation experiments show that wider convolutional receptive fields improve the accuracy of sentence scoring.68% similarUnverifiedMASA (Matrix Approximation Sparse Attention) uses a closed-form solution score to measure the extent to which each sparse unit can reduce the approximation error of the matrix product.67% similarUnverified模型在大海捞针测试中对中间位置信息的检索能力明显弱于首尾位置，这被称为注意力沉降现象67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990375API
curl https://kongchang.com/api/v1/knowledge/claims/990375MCP
get_claim(id=990375)