World Action Models (WAMs) fuse driving policy and visual world modeling into a single generative network whose objectives are fixed after training. Adapting to a new driving goal normally requires retraining. GuidedWAM introduces a model‑agnostic framework that injects driving objectives into frozen WAMs at sampling time. Each objective is expressed as a differentiable energy on the sampler’s intermediate clean‑action estimate, requiring only forward access to the frozen model. Stability is achieved through guidance scheduling during sampling, per‑energy normalization, and velocity‑bounded correction. On the NAVSIM v2 hard‑split benchmark, three publicly released frozen WAMs improve by 1.8–3.6 EPDMS across three sampler seeds, with less than a 10% increase in per‑step sampling time. These findings position energy guidance as a practical post‑training control channel for WAMs.
Review: GuidedWAM demonstrates that energy‑based guidance can repurpose existing autonomous‑driving models with negligible overhead, offering a scalable way to adapt policies without costly retraining.