Seoyoung Lee

Chaining low-rank KV corrections directly across sequential agent workflows eliminates redundant prefills, slashing peak GPU memory by 3.7x and cutting time-to-first-token in half without sacrificing model accuracy.

Conditional branching in navigation tasks reveals hidden failures in agent decision-making that standard metrics overlook.

A4D redefines robot interaction by enabling planning based on what objects can do, not just how they look, achieving unprecedented accuracy and speed in novel scenarios.

Achieve reliable UGV deployment in hidden spaces using only RGB data by enforcing consistency between predicted camera motion and onboard platform egomotion to recover metric scale.

Achieve 75% task success in computer-use agents by explicitly modeling and reusing user intent, outperforming standard RL and trajectory-based approaches.