A research framework called CRAFT reduces collisions in autonomous vehicle simulators by 31% and traffic violations by 33% — without touching the underlying model.
The problem it solves is subtle but serious. Most AV simulators are trained on ego-centric driving logs, where the car recording the data only partially sees surrounding vehicles due to occlusions and sensor limits. That partial view gets baked into the model's learned behaviors. When you then run that simulator in a closed-loop environment — where it has full, global visibility of every agent — it starts acting strangely: phantom stops, unsafe merges, rule violations. The model learned from incomplete information and never got corrected.
CRAFT attacks this without retraining. It generates "what-if" rollouts from existing logged scenarios to surface failure cases, scores them against human-aligned driving priors, and trains a lightweight plug-in module called a Contextual Preference Evaluator. At inference time, that module reweights the base model's action candidates toward globally coherent choices.
The numbers are meaningful — 31.2% fewer collisions and 33.2% fewer violations are not marginal gains. More importantly, the plug-in approach matters for the industry: AV simulation stacks are expensive to retrain, and a drop-in corrective module is far easier to ship and audit. Whether CRAFT's gains hold across diverse edge cases beyond the paper's test set is the obvious open question — and one the authors don't fully answer.