Dimitris N. Metaxas

Forget opaque embeddings: Cross-Layer Transcoders reveal how ViT layers contribute to the final representation, pinpointing the critical few that drive performance.

Halve the training cost of your diffusion transformer without sacrificing generative performance by using multi-patch hierarchies.

Few-step diffusion language models get a boost from trajectory self-distillation with direct discriminative optimization, narrowing the quality gap with slower, full-step decoding.