Selective layer looping improves masked diffusion model training efficiency and reasoning performance via depth scaling without added parameters and flexible inference compute scaling.
EHR-ReasonCon introduces a reasoning-intensive benchmark for clinical note-table consistency verification, and EHR-Inspector achieves state-of-the-art results via LLM-based verification with table exploration.
Condition-dependent source distributions for flow matching improve text-to-image generation via variance regularization and directional alignment, accelerating convergence up to 3x in FID.
SELFCI uses complementary self-distillation to decouple information suppression from task resolution, improving contextual integrity without degrading utility.
AMUSE integrates Muon's rapid bulk progress with Schedule-Free averaging via time-varying interpolation to suppress oscillations, requiring no learning rate schedules and improving training efficiency across vision and LLM tasks.
TRQAM introduces trust-region Q-adjoint matching with adaptive path-space KL control via projected dual descent, enabling stable off-policy flow-policy fine-tuning and achieving 68% success on OGBench.
Kernel Discovery uses an LLM-driven evolutionary framework to search broad kernel spaces for high-dimensional Bayesian optimization, achieving average rank 1.2 out of 17.
STRATA predicts lipid nanoparticle transfection by aligning molecular structure and composition ratio representations to model component interactions. It improves accuracy and generalizes to unseen molecules and ratios.
MotionGrounder enables multi-object motion transfer via a diffusion transformer with flow-based motion signals, object-caption alignment loss, and a new object grounding score. It outperforms baselines in multi-object controllable video generation.