Expired50% confidenceBenchmarkExact time
By artificially enhancing attention to images or text at the appropriate moment, accuracy on visual-spatial reasoning tasks can be improved from 21% to 50%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/16/2026
First Seen
Valid until: 10/14/2026(expired)
Sources
Related Claims
Unverified原始 Self-Attention 的计算量随序列长度呈平方级增长，Anthropic 通过优化位置编码（如 ALiBi）等技术手段拓展了可用上下文74% similarUnverifiedOn visual cognition tasks requiring perception + reasoning + image reconstruction, humans can achieve approximately 80%, while SOTA multimodal models achieve only 30% to 50%.68% similarVerifiedContext length and computational load grow in an approximately quadratic relationship, influenced by the attention mechanism.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/533794API
curl https://kongchang.com/api/v1/knowledge/claims/533794MCP
get_claim(id=533794)