One Rollout Is All You Get: Fully Test-Time Adaptation for GUI Agents

Ziqiang Wang; Li Gu; Zhixiang Chi; Linlian Jiang; Zihuan Jiang; Linqiang Guo; Siobhan Reid; Zhi Liu

cs.LGSep 28, 2026

One Rollout Is All You Get: Fully Test-Time Adaptation for GUI Agents

Authors: Ziqiang Wang, Li Gu, Zhixiang Chi, Linlian Jiang, Zihuan Jiang, Linqiang Guo, Siobhan Reid, Zhi Liu, +1 more

Organizations: Concordia University · Mila – Qu´ebec AI Institute · University of Toronto · Shanghai University

Abstract

GUI agents are deployed with frozen weights and discard everything they experience on the job. Existing ways to update an agent's weights assume something deployment withholds: ground truth, rollouts beyond the single attempt (retries, samples, practice runs), or a learning phase other than deployment. Because GUI actions can be irreversible, a deployed agent gets one attempt per task occurrence, in arrival order, and every attempt counts. No ground truth is available at any point. We define fully test-time adaptation for GUI agents by these constraints and pair it with a minimal weight-space method, SOLO. Auxiliary models read each episode: a judge selects the episodes it deems successful, and a proposer-verifier pair relabels a failed episode's prefix with the subtask that prefix completed. Admitted episodes enter a short sliding window, and each admission updates a small adapter by top-K self-distillation on the agent's own predictions, provided the window holds a judged success. On recurring task streams built from WebArena, VisualWebArena and MobileWorld, SOLO improves on the frozen agent with both UI-TARS-7B and Qwen3-VL-8B, by three to six points of success rate, and exceeds two in-setting memory methods on the web streams.

Figures & tables

Approach	GT supervision	Extra rollouts	Adaptation stage	Retained state	Related work
Demonstration SFT	Yes	No	Pre-deployment	Weights	Wu et al. (2025) ; Qin et al. (2025)
RL / self-evolution	Varies	Yes	Pre-deployment	Weights	Bai et al. (2024) ; Qi et al. (2025) ; Xiao et al. (2025)
Exploration memory	Varies	Yes	Pre-deployment	Memory	Zhang et al. (2025) ; Sun et al. (2026)
Retry-based improvement	Varies	Yes	Deployment	Memory	Shinn et al. (2023) ; Li et al. (2026a) ; He et al. (2026)
Multi-rollout inference	No	Yes	Deployment	None	Snell et al. (2024) ; Yang et al. (2026)
Multi-rollout adaptation	No	Yes	Deployment	Weights	Akyürek et al. (2025) ; Zuo et al. (2025)

Table 1: Where fully test-time adaptation sits among ways to improve a GUI agent. Each row is placed by the resources its learning uses: ground-truth supervision (demonstrations, an environment check or a human; Varies : some members use it, or a trained or VLM evaluator stands in), rollouts beyond the single attempt that counts, the stage at which adaptation happens, and the state retained across tasks. Our setting withholds the first two and admits only the deployment stage, with memory or weights as the retained state. The row marked

\dagger

satisfies it. The last column lists representative works, and Sec. 5 discusses the rest.

Figure 1: Fully test-time adaptation for GUI agents, and Solo . Top: each task occurrence gets one attempt, and a judge decides from the instruction, screens and actions whether it was completed. Bottom: a judged success enters a sliding window as a full episode; for a judged failure, a proposer–verifier pair relabels a completed prefix, which enters the same window. While the window holds a judged success, one top- $K$ self-distillation step updates the LoRA adapter.

	WebArena		VisualWebArena		MobileWorld
Method	UI-TARS-7B	Qwen3-VL-8B	UI-TARS-7B	Qwen3-VL-8B	UI-TARS-7B	Qwen3-VL-8B
Frozen agent	20.8 $\pm$ 0.6	20.1 $\pm$ 0.8	16.7 $\pm$ 1.4	14.8 $\pm$ 0.6	6.1 $\pm$ 1.0	9.7 $\pm$ 1.0
AWM-online	22.7 $\pm$ 1.2	18.7 $\pm$ 1.6	16.3 $\pm$ 1.3	14.7 $\pm$ 1.2	9.7 $\pm$ 1.0	10.8 $\pm$ 1.7
DMS	22.4 $\pm$ 0.6	19.2 $\pm$ 0.5	17.0 $\pm$ 0.9	15.7 $\pm$ 0.9	7.5 $\pm$ 0.8	9.7 $\pm$ 0.5
Solo (ours)	25.8 $\pm$ 1.6	24.9 $\pm$ 0.9	20.0 $\pm$ 0.8	20.9 $\pm$ 1.9	9.2 $\pm$ 0.8	13.3 $\pm$ 2.2

Table 2: Main result. Success rate (%) over the stream under the benchmark’s own evaluator; mean

\pm

std over three runs per cell. Rows per stream: WebArena 324, VisualWebArena 411, MobileWorld 120. AWM-online and DMS are the in-setting memory baselines of Sec. 2 , run on the identical streams.

Variant	VisualWebArena	WebArena	Mean
Frozen agent	14.8 $\pm$ 0.6	20.1 $\pm$ 0.8	17.5
Solo (full)	20.9 $\pm$ 1.9	24.9 $\pm$ 0.9	22.9
w/o window (one step per episode)	18.2 $\pm$ 1.2	23.0 $\pm$ 1.2	20.6
w/o top- $K$ self-distillation (one-hot targets)	20.6 $\pm$ 2.6	22.0 $\pm$ 1.5	21.3
w/o hindsight relabeling (success branch only)	17.0 $\pm$ 1.4	21.9 $\pm$ 0.6	19.4

Table 3: Ablations. One component removed at a time from the full method, on the Qwen3-VL-8B agent: the window, replaced by one step per admitted episode; top-

K

self-distillation, replaced by one-hot imitation of the executed tokens; and hindsight relabeling, leaving the success branch alone. Success rate (%), mean

\pm

std over three runs, and the mean over the two streams.

	VisualWebArena		WebArena
Auxiliaries	Success (%)	Judge prec.	Success (%)	Judge prec.
Frozen agent (no auxiliaries)	14.8 $\pm$ 0.6	–	20.1 $\pm$ 0.8	–
gpt-5-mini	20.9 $\pm$ 1.9	0.51	24.9 $\pm$ 0.9	0.66
gpt-5	18.9 $\pm$ 1.4	0.59	23.0 $\pm$ 0.7	0.76
gpt-5-nano	19.1 $\pm$ 0.6	0.43	23.9 $\pm$ 1.7	0.46
the agent’s own model (Qwen3-VL-8B)	17.3 $\pm$ 0.8	0.35	24.2 $\pm$ 1.0	0.54

Table 4: Dependence on the auxiliary models. Solo on the Qwen3-VL-8B agent with the judge, the proposer and the verifier all replaced by the same alternative model. Success rate (%), mean

\pm

std over three runs, and the judge’s precision, the share of the episodes it admitted that the benchmark evaluator scores as successes, averaged over the runs.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Stream	Arm	Wins per round (total)	One-step episodes	Entropy
VisualWebArena	frozen (mean of 3 runs)	22 / 18 / 21 (61)	36 / 35 / 33	–
	entropy min., run 1	18 / 4 / 1 (23)	55 / 90 / 117	0.116 $\to$ 0.014
	entropy min., run 2	22 / 10 / 4 (36)	48 / 53 / 83	0.113 $\to$ 0.012
	entropy min., run 3	20 / 6 / 8 (34)	48 / 87 / 76	0.114 $\to$ 0.014
WebArena	frozen (mean of 3 runs)	20 / 22 / 23 (65)	5 / 6 / 6	–
	entropy min., run 1	21 / 24 / 21 (66)	6 / 2 / 2	0.111 $\to$ 0.016

Appendix

Table 5: Entropy minimization as the test-time signal. Qwen3-VL-8B, both web streams, three runs each, against the frozen agent. Wins per round with the total, episodes ended at their first step per round, and the mean per-episode response entropy in nats over the first quarter of the stream and over the last.

Stream	Agent	Episodes	Succ.	Fail.	No prefix	Abstain	Guards	Verifier	Admitted	Prefix steps
WebArena	UI-TARS-7B	324	96	228	1	42	63 (53)	22	101 (44%)	3.7
WebArena	Qwen3-VL-8B	324	74	250	6	23	69 (51)	19	134 (54%)	3.0
VisualWebArena	UI-TARS-7B	411	94	317	2	67	133 (100)	24	91 (29%)	3.3
VisualWebArena	Qwen3-VL-8B	411	127	284	49	45	100 (83)	20	69 (24%)	3.0
MobileWorld	UI-TARS-7B	120	23	97	0	12	19 (13)	10	56 (58%)	3.9
MobileWorld	Qwen3-VL-8B	120	25	95	0	6	22 (16)	5	62 (65%)	3.6

Appendix

Table 6: What becomes of a judged failure. Per stream and agent, mean over the three Solo runs: episodes, judged successes, judged failures, and of the failures those with no prefix (ended at the first step), those on which the proposer abstained, those rejected by the guards (in parentheses, rejected as a single-action prefix), those rejected by the verifier, and those admitted as relabeled prefixes, with the admitted share of the failures and the mean number of steps in an admitted prefix.

Benchmark	Method	Agent	Run 1	Run 2	Run 3	Mean $\pm$ std	SR (%)
WebArena	Frozen agent	UI-TARS-7B	68 (26/19/23)	69 (25/21/23)	65 (19/20/26)	67.33 $\pm$ 2.08	20.8
	Frozen agent	Qwen3-VL-8B	63 (15/23/25)	64 (23/19/22)	68 (23/23/22)	65.00 $\pm$ 2.65	20.1
	AWM-online	UI-TARS-7B	72 (23/28/21)	78 (27/27/24)	71 (19/28/24)	73.67 $\pm$ 3.79	22.7
	AWM-online	Qwen3-VL-8B	62 (19/21/22)	55 (19/17/19)	65 (20/23/22)	60.67 $\pm$ 5.13	18.7
	DMS	UI-TARS-7B	72 (22/28/22)	71 (25/25/21)	75 (20/27/28)	72.67 $\pm$ 2.08	22.4
	DMS	Qwen3-VL-8B	62 (21/20/21)	61 (22/18/21)	64 (23/18/23)	62.33 $\pm$ 1.53	19.2

Appendix

Table 7: Every run behind Tab. 2 . Wins per run with the round split in parentheses, then mean

\pm

std of wins and success rate (%).

Stream	Variant	Run 1	Run 2	Run 3	Mean $\pm$ std	SR (%)	Updates
VisualWebArena	Frozen agent	64 (24/20/20)	59 (19/18/22)	60 (22/16/22)	61.00 $\pm$ 2.65	14.8	–
	Solo (full)	82 (22/31/29)	95 (30/32/33)	81 (21/27/33)	86.00 $\pm$ 7.81	20.9	196
	w/o window	74 (20/28/26)	80 (23/29/28)	70 (20/25/25)	74.67 $\pm$ 5.03	18.2	176
	w/o top- $K$ self-distillation	75 (19/28/28)	83 (21/30/32)	96 (26/37/33)	84.67 $\pm$ 10.60	20.6	180
	w/o hindsight relabeling	64 (19/18/27)	71 (24/22/25)	75 (24/25/26)	70.00 $\pm$ 5.57	17.0	109
WebArena	Frozen agent	63 (15/23/25)	64 (23/19/22)	68 (23/23/22)	65.00 $\pm$ 2.65	20.1	–

Appendix

Table 8: Every run behind Tab. 3 . Qwen3-VL-8B on both web streams. Wins per run with the round split in parentheses, then mean

\pm

std of wins, success rate (%), and the mean number of updates per run.

Stream	Auxiliaries	Wins per run	Judge precision per run	SR (%)
VisualWebArena	gpt-5-mini	82 / 95 / 81	0.52 / 0.54 / 0.48	20.9
VisualWebArena	gpt-5	76 / 73 / 84	0.59 / 0.59 / 0.60	18.9
VisualWebArena	gpt-5-nano	80 / 76 / 80	0.46 / 0.41 / 0.42	19.1
VisualWebArena	the agent’s own model (Qwen3-VL-8B)	69 / 75 / 69	0.36 / 0.36 / 0.34	17.3
WebArena	gpt-5-mini	80 / 78 / 84	0.63 / 0.67 / 0.67	24.9
WebArena	gpt-5	72 / 76 / 76	0.74 / 0.75 / 0.78	23.0

Appendix

Table 9: Every run behind Tab. 4 . Wins per run, the judge’s precision on that run (share of admitted episodes that the evaluator scores as successes), and the success rate over the runs.

Stream	Agent	Method	Steps per episode (r1 / r2 / r3)	Budget-ended % (r1 / r2 / r3)
WebArena	UI-TARS-7B	frozen agent	17.0 / 16.5 / 17.5	41 / 40 / 43
		AWM-online	17.4 / 16.4 / 17.5	42 / 36 / 40
		DMS	16.7 / 16.0 / 17.5	41 / 36 / 43
		Solo	17.8 / 15.7 / 16.1	43 / 34 / 35
	Qwen3-VL-8B	frozen agent	9.1 / 9.5 / 9.1	6 / 8 / 8
		AWM-online	9.0 / 8.1 / 8.0	9 / 7 / 5

Appendix

Table 10: Episode length along the stream. Mean steps per episode and share of episodes ending on the step budget, per round, mean over the three runs. Every cell of Tab. 2 on the web streams and every variant of Tab. 3 .

Stream	Agent	Judged	True among judged	Evaluator	Precision	Recall
WebArena	UI-TARS-7B	96	58	84	0.60	0.69
WebArena	Qwen3-VL-8B	74	49	81	0.66	0.60
VisualWebArena	UI-TARS-7B	94	57	82	0.61	0.70
VisualWebArena	Qwen3-VL-8B	127	65	86	0.51	0.76
MobileWorld	UI-TARS-7B	23	10	11	0.44	0.94
MobileWorld	Qwen3-VL-8B	25	16	16	0.63	0.98

Appendix

Table 11: The gpt-5-mini judge on Solo ’s episodes. Mean over the three runs per cell: judged successes, the true successes among them, the evaluator’s successes, precision (true among judged) and recall (judged among true).

Stream	Judge	Wins per run	Judge precision	Success (%)
VisualWebArena	gpt-5-mini (default)	82 / 95 / 81	0.51	20.9 $\pm$ 1.9
VisualWebArena	benchmark evaluator	80 / 78 / 95	1.00	20.5 $\pm$ 2.3
WebArena	gpt-5-mini (default)	80 / 78 / 84	0.66	24.9 $\pm$ 0.9
WebArena	benchmark evaluator	78 / 71 / 78	1.00	23.4 $\pm$ 1.2

Appendix

Table 12: A perfect judge against the default judge. Solo on Qwen3-VL-8B with the judge replaced by the benchmark evaluator, proposer and verifier unchanged. Wins per run, judge precision, and success rate (%) mean

\pm

std over three runs.

Judge	Proposer and verifier	Wins per run	Judge precision	Success (%)
gpt-5-mini	gpt-5-mini	82 / 95 / 81	0.51	20.9 $\pm$ 1.9
Qwen3-VL-8B	gpt-5-mini	71 / 70 / 83	0.36	18.2 $\pm$ 1.8
gpt-5-mini	Qwen3-VL-8B	76 / 69 / 70	0.48	17.4 $\pm$ 0.9
Qwen3-VL-8B	Qwen3-VL-8B	69 / 75 / 69	0.35	17.3 $\pm$ 0.8

Appendix

Table 13: The auxiliary roles swapped one at a time on VisualWebArena. Solo on Qwen3-VL-8B. Wins per run, judge precision and success rate (%) mean

\pm

std over three runs.

Explore similar work

Aug 11, 2026cs.CV

Test-Time Self-Evolving GUI Visual Grounding via Reflection-Guided On-Policy Self-Distillation

GUI Visual Grounding is a fundamental capability for GUI agents. Existing models typically freeze their parameters after deployment, limiting their ability to adapt to unseen interfaces. Although recent methods attempt to adapt models via test-time reinforcement learning, they cannot reflect upon failed exploration. To overcome this, we propose a Test-Time Self-Evolving framework that enables models to improve after deployment without human-annotated ground truth. It constructs a closed-loop of Exploration, Evaluation, Reflection, and Internalization. Specifically, the agent first explores unseen interfaces by predicting grounding coordinates for given instructions. To evaluate these explorations, we introduce an MLLM-based Reflector to assess the generated results and provide the corresponding reasoning reflections. To internalize reflection knowledge into the model weights, we propose Reflection-Guided On-Policy Self-Distillation, which translates high-level reasoning into dense token-level supervision via a conditioned self-teacher. Furthermore, we design a Contrastive Calibration method to prevent incorrect auto-regressive prefixes from corrupting the supervisory signals during failed explorations. Extensive experiments across six benchmarks demonstrate our framework's effectiveness, achieving an average accuracy improvement of 7.4% over the base model. To the best of our knowledge, this is the first work to successfully exploit on-policy self-distillation for test-time adaptation in GUI visual grounding. By filling the gap in post-deployment adaptation, our framework completes the self-evolving capability of GUI agents. The code will be released.

Shiyu Xuan, Zechao Li

School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China, 210094

Graphical User Interface Graphical User Interface Agents Post-Deployment Adaptation Unsupervised On-Policy Self-Distillation

Ask PDF

Aug 12, 2026cs.AI

CoAdapt-GUI: Joint Workflow Context and Policy Adaptation for Unseen GUI Applications

Mobile GUI agents remain brittle when deployed to applications absent from source training. We study novel-app generalization under a limited target interaction budget and without target demonstrations. We introduce CoAdapt-GUI, a test-time adaptation (TTA) framework that jointly adapts structured workflow context and policy from the agent's own target-app rollouts and rewards. The workflow context retains transferable procedures, failure modes, and verification rules while excluding app-bound source details. This separation allows reusable workflow knowledge to guide adaptation without transferring source-interface state. For policy adaptation, task-context-matched group-relative optimization updates a LoRA adapter on a frozen vision-language model. Across two unseen-app evaluations, CoAdapt-GUI reaches 45.0% on AndroidWorld-Generalization, compared with 37.5% for the reported Policy-Only TTA baseline, and raises AndroidWorld Plus performance from 38.6% to 52.9%. These results show that transfer-constrained workflow context provides substantial gains and that joint policy adaptation further improves held-out performance.

Linqiang Guo, Li Gu, Zihuan Jiang +8

Peter · Concordia University · Mila – Québec AI Institute +2

Graphical User Interface Agents Text-Only Adaptation Device Agentic Workflow Design

Ask PDF

Jul 30, 2026cs.AI

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

GUI agents have the potential to become a general purpose executor over existing digital devices. To advance them toward real-world use, we envision agents that operate reliably on real devices, execute workflows across platforms, combine GUI interaction with CLI execution, complete long-horizon tasks, proactively initiate useful services, and autonomously improve their capabilities with minimal human effort. Guided by this vision, we present Qwen-UI-Agent, a real-world centric foundation GUI agent spanning mobile, computer-use, web, and DeepSearch environments. Qwen-UI-Agent combines diverse sandbox environments with a large-scale real-device mobile runtime. Its unified action space interleaves GUI operations with CLI execution and generates batched actions in a single model turn. An AutoResearch-style data flywheel uses agents to construct tasks and environments, diagnose failures, and plan subsequent iterations. Online RL supports training on trajectories exceeding 100 turns, with over 10,000 concurrent environments accelerating rollout. A lightweight harness layer supports proactive service initiation and stateful workflows across mobile and computer. Across a broad suite of evaluations, Qwen-UI-Agent sets state-of-the-art performance on mobile-use benchmarks while delivering competitive performance on computer- and browser-use tasks against frontier models, including Opus 4.8, Gemini 3.1 Pro, and GPT-5.6 Sol. On mobile use, it achieves 82.1% on MobileWorld, 92.2% on MobileWorld-Real, and 97.5% on AndroidDaily. On computer use, it achieves 79.5% on OSWorld-Verified and a 40.0% partial-progress score on OSWorld-v2. On browser use and GUI grounding, it achieves 73.6% on WebArena and 81.5% on ScreenSpot-Pro, respectively.

Hanzhang Zhou, Panrong Tong, Xu Zhang +13

MAI-UI Team Alibaba Token Hub, Alibaba Group

Graphical User Interface Agents Device Technical Report Proactive

Ask PDF