🛰️ Daily AI Frontier
‹ back to 2026-07-30

Why Are GUI Agents Correct but Late? Decode on the Decision-Time Critical Path, Tested with Pre-Compiled Policy Trees

arXiv cs.LG LLM Agents Zihan Dong, Rui Qian, Qishi Zhan, Dongshen Peng, Kaixin Li, Yu Li 2026-07-30

TL;DR - AAPT pre-computes bounded GUI action trees during idle periods, moving costly multimodal-model decoding off the decision-time critical path. This raises success on transient events from 0.50 to 0.79 without incorrect actions.

  • A lightweight observer routes screen changes to pre-authorized actions for immediate execution.
  • Baselines that decoded during execution had zero success within the contested decision window.
  • Ablations identify observer speed, valid planning, and especially accurate branch routing as critical.
  • Results replicated on another multimodal model across 126 paired trials; reactive methods remain better when actions cannot be enumerated beforehand.

view merged work →