Search papers, labs, and topics across Lattice.
Fudan University
4
0
9
5
EasyOPD unifies on-policy distillation methods, enabling seamless integration and superior performance across diverse tasks in large language models.
Existing guardrails falter under shifting policies, but PolicyShiftGuard achieves a remarkable 76.9 Avg. F1 score by adapting dynamically to policy changes.
DCPM reveals that separating memory processes can boost LLM agents' ability to infer user intentions and beliefs across sessions, leading to a dramatic increase in personalization accuracy.
Current AI memory systems struggle with fine-grained relational discrimination, revealing critical gaps in their ability to manage complex memory interactions over time.
