Lingjie Chen

Auxiliary draft models are no longer necessary for speculative decoding: distilling lightweight diffusion heads directly into standard LLMs yields lossless 3× generation speedups even at peak batch sizes.

PILL achieves up to 6.0 BLEU-2 improvement on text infilling while running 1.82x faster than the best existing method, revolutionizing efficiency in diffusion language models.

Bridging the gap between inference and adaptation in VLMs could lead to significant performance boosts by ensuring robust pseudo-labels that accurately reflect sample-level relationships.

RECONTEXT boosts long-context reasoning in LLMs by effectively reusing evidence from the input, leading to superior performance without the need for retraining.

A unified model that seamlessly integrates text and graph learning outperforms traditional methods by up to 3.9 points on benchmark tasks.