Search papers, labs, and topics across Lattice.
SambaNova Systems, Inc
2
0
5
LongAttnComp not only matches full-context accuracy but also significantly boosts efficiency in long-context reasoning tasks, outperforming existing methods across diverse model families.
Forget same-family constraints: you can compress prompts for LLaMA with a Qwen draft model and still get 90-100% of the original performance.
