Search papers, labs, and topics across Lattice.
3
0
8
Background leakage in image editing can be effectively mitigated by SAM-Flow's localized approach, ensuring high fidelity in semantic modifications.
Forget slow matrix multiplies: this new quantization method lets you run LLaMA-2-7B in 3-bit with only bit shifts and additions, beating heavier methods like AWQ.
LLM code generation benchmarks are likely overestimating model capabilities: adversarial test suite scaling reveals substantial weaknesses in even state-of-the-art models.
