Search papers, labs, and topics across Lattice.
3
0
5
0
Autoregressive video models can now perform explicit, goal-directed reasoning by jointly predicting the final outcome and intermediate steps, achieving state-of-the-art performance with 75% fewer training steps.
Clamping harmful features in models may only mask misbehavior, allowing for a surprising 95.8% recovery of suppressed actions even under intervention.
Visually indistinguishable adversarial images can trigger catastrophic failures in visual world models, exposing critical vulnerabilities in their deployment.
