Hugging Face Daily Papers · September 14, 2026 · 3 min read

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

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

<a href=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/h1EsJEcxErOIS8_zmo54t.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/h1EsJEcxErOIS8_zmo54t.png\" alt=\"02-fig1-motivation\"></a><br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/UEnU7f5pcj44AXoA4mYF4.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/UEnU7f5pcj44AXoA4mYF4.png\" alt=\"03-fig2-overview\"></a><br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/yblLNGUkyAaSb_lOHbPeJ.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/yblLNGUkyAaSb_lOHbPeJ.png\" alt=\"04-fig4-keepmaps\"></a><br><a href=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/afxp7EiwxE4ozlgHRWzln.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/62dea2a3a8ccfacec7102dc3/afxp7EiwxE4ozlgHRWzln.png\" alt=\"05-fig5-efficiency\"></a></p>\n","updatedAt":"2026-09-14T14:22:55.046Z","author":{"_id":"62dea2a3a8ccfacec7102dc3","avatarUrl":"/avatars/2875ff8489a0b1d3378096a75af8285f.svg","fullname":"Yuhao Wang","name":"YHY-Test","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.47684600949287415},"editors":["YHY-Test"],"editorAvatarUrls":["/avatars/2875ff8489a0b1d3378096a75af8285f.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.10297","authors":[{"_id":"6aa802c15dd4cb9b4cc02201","name":"Yuhao Wang","hidden":false},{"_id":"6aa802c15dd4cb9b4cc02202","name":"Mu Qiao","hidden":false},{"_id":"6aa802c15dd4cb9b4cc02203","name":"Xindong Zhang","hidden":false},{"_id":"6aa802c15dd4cb9b4cc02204","name":"Yunzhi Zhuge","hidden":false},{"_id":"6aa802c15dd4cb9b4cc02205","name":"Lei Zhang","hidden":false},{"_id":"6aa802c15dd4cb9b4cc02206","name":"Huchuan Lu","hidden":false}],"publishedAt":"2026-09-09T00:00:00.000Z","submittedOnDailyAt":"2026-09-14T00:00:00.000Z","title":"TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents","submittedOnDailyBy":{"_id":"62dea2a3a8ccfacec7102dc3","avatarUrl":"/avatars/2875ff8489a0b1d3378096a75af8285f.svg","isPro":false,"fullname":"Yuhao Wang","user":"YHY-Test","type":"user","name":"YHY-Test"},"summary":"GUI agents accumulate high-resolution screenshots as the trajectory unfolds, increasing inference latency and memory usage. Training-free visual token pruning can reduce this cost, but cache reuse introduces a fundamental constraint. Once tokens are discarded, the corresponding visual evidence cannot be recovered without re-encoding. Pruning therefore becomes an irreversible admission decision that must remain useful for unknown future targets while preserving coverage of operable regions under tight budgets. To address these challenges, we propose \\method{}, a training-free framework for \\textbf{Trajectory-robust Admission and Coverage-aware Evidence ordering}. Specifically, we combine a query-independent layout-derived interaction prior with instruction relevance and feature novelty to rank visual evidence according to both potential future utility and diversity. Then, we reserve part of the budget for native visual tokens distributed across the screen, repairing missing spatial coverage without breaking the ordering. Together, these mechanisms produce a nested token order, allowing retained visual evidence to shrink monotonically across budgets while remaining reusable throughout the trajectory. Finally, our monotone KV contraction incrementally contracts retired frames into compact session state, avoiding repeated visual encoding or pruning. Extensive experiments across six GUI benchmarks and diverse models verify the effectiveness of our proposed under tight budgets. The source code will be released.","upvotes":0,"discussionId":"6aa802c15dd4cb9b4cc02207","githubRepo":"https://github.com/924973292/TRACE","githubRepoAddedBy":"user","ai_summary":"TRACE is a training-free framework that ranks visual evidence by future utility and diversity, reserves native tokens for spatial coverage, and contracts retired frames to reduce latency and memory in GUI agents.","ai_keywords":["visual token pruning","KV contraction","trajectory-robust admission","coverage-aware evidence ordering","GUI agents","interaction prior","nested token order"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":2},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[],"acceptLanguages":["en"],"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.10297.md","query":{}}">

Papers

arxiv:2609.10297

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

Published on Sep 9

· Submitted by

Yuhao Wang on Sep 14

Upvote

Authors:

Abstract

TRACE is a training-free framework that ranks visual evidence by future utility and diversity, reserves native tokens for spatial coverage, and contracts retired frames to reduce latency and memory in GUI agents.

Generated by thinkingmachines/Inkling-Small

GUI agents accumulate high-resolution screenshots as the trajectory unfolds, increasing inference latency and memory usage. Training-free visual token pruning can reduce this cost, but cache reuse introduces a fundamental constraint. Once tokens are discarded, the corresponding visual evidence cannot be recovered without re-encoding. Pruning therefore becomes an irreversible admission decision that must remain useful for unknown future targets while preserving coverage of operable regions under tight budgets. To address these challenges, we propose \method{}, a training-free framework for \textbf{Trajectory-robust Admission and Coverage-aware Evidence ordering}. Specifically, we combine a query-independent layout-derived interaction prior with instruction relevance and feature novelty to rank visual evidence according to both potential future utility and diversity. Then, we reserve part of the budget for native visual tokens distributed across the screen, repairing missing spatial coverage without breaking the ordering. Together, these mechanisms produce a nested token order, allowing retained visual evidence to shrink monotonically across budgets while remaining reusable throughout the trajectory. Finally, our monotone KV contraction incrementally contracts retired frames into compact session state, avoiding repeated visual encoding or pruning. Extensive experiments across six GUI benchmarks and diverse models verify the effectiveness of our proposed under tight budgets. The source code will be released.