Search papers, labs, and topics across Lattice.
4
0
7
Claw-like agents are vulnerable to severe security breaches, with malicious plugins achieving a 100% success rate in attacks.
LLMs can substantially improve their ability to follow complex instructions and constraints by explicitly auditing their own context adherence before answering.
LLMs struggle with GIS tasks due to parameter misalignment and runtime anomalies, but a "Plan-and-React" architecture that mimics expert cognitive workflows significantly improves performance.
Meituan's new tree-based reranking method significantly boosts recommendation quality by generating lists in a coarse-to-fine manner, effectively balancing global context and local user interests.
