Search papers, labs, and topics across Lattice.
3
0
8
Memory overload in autoregressive video generation can be tackled by absorbing historical context into model weights, achieving up to 50% cache reduction with minimal quality loss.
Pre-load auditing of Agent Skills can achieve >97% accuracy in detecting malicious intent, even against semantics-preserving rewrites, by combining role-aware evidence extraction with semantic verification.
Decoupling fact injection from text generation lets you edit LLMs with greater precision, improving fine-grained question answering without sacrificing overall editing performance.
