Soheil Kolouri

Shrinking an LLM's residual stream based on activation variance alone discards the wrong directions—weighting pruned subspaces by downstream output sensitivity dramatically improves compressed model performance without sacrificing closed-form computational efficiency.

KG-FairDiff achieves substantial reductions in demographic disparities in text-to-image generation while preserving the original intent of prompts, making it a game-changer for equitable AI deployment.

Dataset distillation's dirty secret: soft labels bloat storage, but a VQ-VAE slashes their size by 30-40x without sacrificing performance.