Search papers, labs, and topics across Lattice.
School of Information Management, Wuhan University, Wuhan, China
2
0
6
LLM inference gets a 2x speed boost without training, thanks to a clever technique that merges retrieval with logit-based speculation.
Grokking isn't just about memorization then generalization; it's a dimensional phase transition in the gradient field, revealing a fundamental shift in how networks learn.
