[HUGGINGFACE]score: 0.48
PANDO Cuts Web Agent Inference Cost via Online Skill Distillation on 910 Tasks
May 25, 2026
PANDO is a single-rollout web agent framework that builds a structured Skill Library online, using confidence-based skill demotion, hierarchical routing, visual compression, and cache-aware prompting to reduce inference cost as experience accumulates. Evaluated on all 910 VisualWebArena tasks, it targets the inefficiencies of repeat-action loops and low prompt-cache reuse common in current multimodal agents.
paper
HOW THIS AFFECTS YOU
●
builderIf you're building multimodal web agents, PANDO's cache-aware prompting and skill routing could reduce per-task inference cost without requiring multiple model passes.
●
researcherThe online distillation approach with confidence-based demotion is a concrete alternative to offline skill discovery — VisualWebArena at 910 tasks is a meaningful evaluation scale.