PANDO: Efficient Multimodal AI Agents via Online Skill Distillation

Researchers introduced PANDO, a multimodal web-agent framework that learns from a single rollout to build a structured skill library.

PANDO uses progress reflection, skill routing, visual compression, and cache-aware prompting to reduce redundant actions and token use.

On 910 VisualWebArena tasks, it reached 58.3% success, outperforming SGV and a WALT reproduction while using far fewer tokens.

The work adds new efficiency metrics and shows web agents can get faster and leaner as they gain experience.