Jianzong Wang

Unified multimodal models suffer from internal conflict, but this work shows how to turn that interference into a surprisingly effective source of performance gains.

Forget reinforcement learning: robots can now evolve manipulation skills simply by reflecting on their successes and failures.