Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Canva Research
5
0
8
Speculative decoding drafters can achieve over 2x speedup on 256K context by treating distant tokens as shared far memory with sharpened attention.
Enforcing uniform load balancing across video tokens fundamentally breaks visual coherence, but bifurcating experts into semantic and generic tiers resolves patch fragmentation while driving emergent coarse-to-fine denoising.
Results show that resume-screening evaluations should assess not only whether a system identifies stronger candidates, but also whether those decisions remain stable when the same competence evidence is presented differently, and whether those decisions remain stable when the same competence evidence is presented differently.
No existing controllable video generation model excels in both visual fidelity and inherent reactivity, revealing a fundamental limitation in current benchmarks.