Search papers, labs, and topics across Lattice.
Shanghai Jiao Tong University
2
7
5
15
LLMs' final layers might be holding your recommendation system back: representations from middle layers actually perform *better*, and a modular compression approach can unlock significant gains.
Current LLM evaluation benchmarks often conflate chatbots and true AI agents, leading to misaligned research efforts, but this survey provides a framework for targeted evaluation based on environmental complexity and agent capabilities.
