Trajectory-Aware Decoding Control for Diffusion VLMs
September 11, 2026
A new training-free controller for Diffusion Vision-Language Models mitigates reasoning-budget mismatch by routing examples to early commitment or reasoning-supportive decoding. It uses trajectory signals from answer closure and representation revision pressure to optimize generation length.
HOW THIS AFFECTS YOU
●
builderYou can improve inference efficiency and accuracy in VLMs by implementing trajectory-aware routing.
●
researcherThis addresses the fundamental mismatch between static generation lengths and varying reasoning demands in diffusion models.