实体 Diff-SAE

Diff-SAE

PulseAugur coverage of Diff-SAE — every cluster mentioning Diff-SAE across labs, papers, and developer communities, ranked by signal.

总计 · 30天

1

90 天内 1

发布 · 30天

0

90 天内 0

论文 · 30天

1

90 天内 1

层级分布 · 90 天

时间线

2026-05-08 research_milestone Researchers published a paper detailing a new method for detecting backdoors in language models. 来源

情绪 · 30 天

1 天有情绪数据

最近 · 第 1/1 页 · 共 1 条

TOOL · CL_25606 · May 8 · 06:30

New Diff-SAE method excels at detecting language model backdoors

Researchers have developed a new method using Sparse Autoencoders (SAEs) to detect backdoor attacks in language models. Their Differential SAE (Diff-SAE) architecture proved significantly more effective than Crosscoders…