Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
1
0
1
4
Learned soft prefixes can shift correct logical judgments by up to 99 percentage points, exposing critical vulnerabilities in model reasoning stability.