Jointly decides when waiting agent sessions should run on GPUs and which node should serve them. Scheduling accounts for tool-return readiness, time-decaying node cache value, and token demand.
Decides when a diffusion vision-language model needs to process an image by checking whether text-only predictions agree with its latest image-conditioned predictions. Retains 99.6% of baseline scores across ten benchmarks at 35.3% of decoding FLOPs.
Tracks reward variation and accuracy to detect when training tasks stop providing useful learning signal, then refreshes the multimodal task pool without external examples or human-written answer labels.
Improves few-step text-to-image distillation by matching feature means and per-coordinate standard deviations. It retains DMD’s full score correction and uses the generator’s EMA copy for features, avoiding an external encoder.