Search papers, labs, and topics across Lattice.
King鈥檚 College London
2
0
5
ASRD leverages trusted Anchor Tokens to dramatically improve decoding quality and speed in diffusion LLMs, achieving a 6.4% accuracy boost while accelerating inference by 7.2x.
EDIT reveals that targeted interventions based on internal model diagnostics can dramatically enhance LLM grading accuracy, outperforming conventional methods.
