Unverified50% confidenceBenchmarkExact time
On real benchmarks in four vision and language settings, ConSDL improves dictionary consistency by an average of 17.6% and coefficient consistency by 35.9% over six SAE baseline methods.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedConSDL replaces the standard linear encoder with per-sample sparse coding and achieves state-of-the-art cross-seed consistency on multiple vision and language tasks.62% similarUnverifiedOn three fixed language model checkpoints, centered logit SAEs reconstruct the output logits better in 17 settings.61% similarUnverifiedAider项目的实践数据表明，使用unified diff格式的编辑成功率明显高于纯字符串替换方案，尤其在处理有重复代码片段的文件时差异显著60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065921API
curl https://kongchang.com/api/v1/knowledge/claims/1065921MCP
get_claim(id=1065921)