Ziyang Gong

Model-generated skills can actually hurt agent performance, and bigger models don't necessarily make for better skill extractors or consumers.

SkillOpt transforms agent skill development into a reproducible optimization process, achieving state-of-the-art results by treating skills as trainable parameters.

Visual degradations can cripple the spatial reasoning abilities of even state-of-the-art MLLMs, but targeted finetuning can restore—and even surpass—human-level performance.