TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents

TRACE. Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents. By Yuhao Wang and five others, from the Dalian University of Technology, the OPPO Research Institute, and one other institution. Published on the ninth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

GUI agents pile up high-resolution screenshots as a trajectory goes on, and inference latency and memory usage grow with them. Visual token pruning that needs no training can cut that cost, but cache reuse imposes a hard constraint. Once tokens are discarded, the visual evidence they carried cannot be recovered without encoding the screen again. Pruning therefore becomes an irreversible admission decision. What is kept has to stay useful for future targets that are not yet known, and has to keep covering the regions of the screen that can be operated, all under a tight budget.

TRACE, for Trajectory-robust Admission and Coverage-aware Evidence ordering, is a training-free framework built around that decision. It ranks visual evidence by combining three signals. A prior on likely interaction, derived from the layout and independent of any query. Relevance to the instruction. And novelty of features. Together they order evidence by both its potential future utility and its diversity. The layout prior matters because of the irreversibility. The current instruction says what is relevant now, but a later step may need a different part of the screen, and a prior that does not depend on the query is a way of keeping evidence that the future might want.

Part of the budget is then reserved for native visual tokens spread across the screen, which repairs gaps in spatial coverage without breaking the ordering. The result is a nested token order. Retained evidence can shrink monotonically as the budget tightens, and whatever is retained stays reusable throughout the trajectory. Nesting means a smaller budget keeps a prefix of what a larger budget would keep, so tightening the budget never asks for a token that has already been thrown away. Finally, a monotone KV contraction step incrementally contracts retired frames into a compact session state, so nothing has to be visually encoded or pruned a second time.

Extensive experiments across six GUI benchmarks and a range of models support its effectiveness under tight budgets. The source code will be released.

With a reused cache, pruning is a one-way door, and TRACE is designed around that fact rather than against it. Each piece answers one consequence of it: an ordering that holds at every budget, a reserved share of tokens for spatial coverage, and a way to retire old frames without paying to see them twice.