TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents
TRACE treats screenshot pruning as an irreversible choice, then orders what to keep so GUI agents can reuse it under smaller budgets.
The paper targets GUI agents whose growing screenshot histories raise latency and memory costs. Its training-free method ranks visual evidence using layout-based interaction signals, instruction relevance, and feature novelty. It also reserves tokens across the screen to preserve spatial coverage of operable regions. The authors say tests across six GUI benchmarks and multiple models show the approach works under tight budgets. HF Daily Papers' note
The paper targets GUI agents whose growing screenshot histories raise latency and memory costs. Its training-free method ranks visual evidence using layout-based interaction signals, instruction relevance, and feature novelty. It also reserves tokens across the screen to preserve spatial coverage of operable regions. The authors say tests across six GUI benchmarks and multiple models show the approach works under tight budgets. HF Daily Papers' note
score 4