Xufang Luo

SHAPER enables embodied agents to self-evolve their skills and context without retraining, unlocking new possibilities for adaptation in fixed-interface scenarios.

Off-policy degree in RLVR updates can drastically change which tokens drive learning, leading to a new adaptive method that outperforms traditional baselines.

LVLMs aren't all that they seem: some lean heavily on language priors while others genuinely fuse vision and language, revealing surprisingly diverse strategies across model families.

Forget monolithic models: pMoE shows that ensembling diverse expert prompts within a single model framework yields surprisingly large gains in visual adaptation across a wide range of tasks.

LLM agents can learn to explore novel states and generalize to new tasks with a hybrid on- and off-policy RL framework that leverages memory.