AFRAgent : An Adaptive Feature Renormalization Based High Resolution Aware GUI agent

Neeraj Anand, Rishabh Jain, Sohan Patnaik, Balaji Krishnamurthy, Mausoom Sarkar

🏛 Institutions: Adobe
📅 Date: November 30, 2025
📑 Publisher: WACV 2026
💻 Env: Mobile
🔑 Keywords: model GUI grounding smartphone automation adaptive feature renormalization instruct-BLIP AFRAgent

TLDR

AFRAgent targets the loss of spatial detail that hurts mobile GUI automation models built on low-resolution vision features. It adds an adaptive feature renormalization module to enrich instruct-BLIP image embeddings with high-resolution information, and reports state-of-the-art results on Meta-GUI and AITW with a much smaller model than prior baselines.

Open paper arXiv Edit on GitHub Report issue