Search papers, labs, and topics across Lattice.
1
0
3
Token ranking heuristics for LLM prefill are surprisingly unstable across layers, but simply aggregating attention scores across layers can dramatically improve performance.
