Yihan Zeng

MLLMs can achieve 10% gains on multimodal reasoning benchmarks by using ground-truth anchored data curation and scaffold-stripping to avoid cognitive drift during self-evolution.

End-to-end joint optimization of planning and execution in an image restoration agent unlocks significantly improved performance compared to independently trained tools and all-in-one models.

RL's inherent resilience to catastrophic forgetting can be harnessed to improve continual learning in GUI agents, outperforming SFT alone.