Andrei Bursuc

Synthetic scenes with automatically grounded object identities unlock substantial real-world perceptual gains in MLLMs via on-policy self-distillation.

Leading MLLMs still struggle to articulate and localize physical changes across revisited scenes, exposing a fundamental blind spot in multi-image spatial grounding that targeted synthetic data can overcome.

Training self-driving policies directly from perspective images can rival traditional methods that rely on privileged data, achieving impressive zero-shot transfer performance.

Post-training on synthesized safety-critical scenarios can dramatically enhance the reliability of autonomous driving systems, reducing failures in rare but critical events.

MLLMs can get a surprising visual reasoning boost from a simple trick: adding just a dash of visually grounded self-supervision to instruction tuning.