Search papers, labs, and topics across Lattice.
2
0
4
4
LLMs morph into riskier conversationalists when playing directive support roles like "Coach" or "Inform" for caregivers, revealing a troubling quality-safety trade-off.
Open-source LLMs, trained with Reddit community feedback, can match the performance of proprietary models in generating supportive mental health responses, offering a privacy-preserving alternative.
