Jiashun Liu

Current multimodal agents can build sophisticated perception-control pipelines but fail to reliably execute physical tasks, losing object states and mistaking unfinished goals for completion.

Tracking both task advantage and actual reward gains can drastically improve the efficiency of multi-task reinforcement learning for LLMs.

Merging existing expert capabilities can yield significant performance boosts, but choosing the right fusion method can make all the difference in multi-domain reinforcement learning.

Plasticity loss in RL isn't just about forgetting; it's about vanishing gradients, and a simple sample re-weighting can bring back the learning.