Search papers, labs, and topics across Lattice.
University of Michigan
5
0
7
Deductive stereotyping in LLMs leads to biased reasoning, but targeted injection phrases can dramatically enhance fairness across diverse tasks.
Misfired alignment in LLMs can lead to a 18.9% failure rate in reasoning about stereotypes, revealing a critical flaw in current safety-oriented training methods.
A single biased example can completely undermine the alignment of large language models, revealing a critical flaw in their post-training safety mechanisms.
LLMs can be tricked into unsafe responses to children, but even subtle cues about the user's age can dramatically improve safety.
Even state-of-the-art models like Gemini and Claude can completely miss critical user information when it's buried in semantically unrelated past interactions, tanking personalization performance.
