Tag
Speculation on how far ahead Anthropic's internal frontier models might be from the publicly released Fable 5 model, given months of feedback and AI-assisted development cycles.
NVIDIA researchers developed a technique to speed up bounding box detection by 10x by eliminating the autoregressive token-by-token prediction step used in VLM grounding models.
Draft-OPD introduces on-policy distillation with target-assisted rollouts and error replay to overcome the offline-to-inference mismatch in training draft models for speculative decoding, achieving over 5x lossless acceleration and improving upon EAGLE-3 and DFlash by 23% and 13% respectively.