acceptodds
Under review as a conference paper at ICLR 2027

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

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- 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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
GUI agents, test-time adaptation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SZrwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper defines fully test-time adaptation for GUI agents: no ground truth, one attempt per task, no pre-deployment training, and proposes SOLO uses auxiliary models to extract learning signal from each episode, updating a LoRA adapter via top-K self-distillation anchored by a sliding window. On three benchmarks with two agents, SOLO improves over the frozen baseline by 3–6 points and exceeds two memory baselines on the web streams.

Critical strengths
  1. The paper defines fully test-time adaptation for GUI agents via three deployment-realistic constraints: no ground truth, one attempt per task, and no pre-deployment training phase. This setting is both practically grounded and differentiated from existing paradigms, filling a gap that prior work on agent improvement has not explicitly addressed.

  2. Unlike prior in-setting methods that store learned workflows or experiences in the prompt, SOLO distills them into a lightweight LoRA adapter. This weight-space approach offers better generalization.

  3. Across two agents and three benchmarks, SOLO improves over the frozen baseline by 3–6 points on every stream, and surpasses both in-setting memory baselines on the web.

Critical weaknesses
  1. SOLO continually updates a LoRA adapter via a sliding window, how to ensure that catastrophic forgetting does not occur during the change of tasks. For instance, the adapter's previously learned episodes are not overwritten by subsequent tasks?

  2. Table 4 reveals the judge achieves a precision of only 0.51 on VisualWebArena and 0.66 on WebArena, meaning that half of the episodes are failures. These mislabeled episodes directly enter the sliding window and drive adapter updates, injecting a substantial amount of noisy gradient signal. The paper does not address how to avoid or filter this noisy data.

  3. The paper omits any discussion of deployment-time computational costs. For each task, SOLO requires call to a closed-source model for judging, and up to two additional calls (proposer and verifier) when an episode is judged as a failure. On top of this inference overhead, each admission triggers a gradient update over the sliding window. These costs are not quantified, nor are they compared against the baselines, which are training-free and operate entirely in the prompt.

  4. In ablation auxiliary model dependence study are restricted to Qwen3-VL-8B on the two web streams, omitting MobileWorld Yet the main results in Tab. 2 show that SOLO's gains on MobileWorld with UI-TARS-7B are marginal. Without ablation evidence on MobileWorld, it remains unclear whether each component contributes comparably or whether the method's behavior differs qualitatively in the mobile domain.

  5. Table 4 shows that gpt-5 yields a lower success rate than gpt-5-mini. If higher-precision filtering degrades performance, what exactly is the judge selecting for that benefits the adapter? The mechanism by which judge precision interacts with the quality of the learning signal is not analyzed.

  6. Figure 2 shows that the improvement primarily in rounds 2 and 3, when the same instances reappear and the adapter can exploit previously learned patterns. If deployment scenarios where task recurrence is low or absent, the method's practical benefit may be diminished.

Questions for authors
  1. Please provide additonal experiments to address the proposed concerns.

  2. For the catastrophic forgetting concern, can the author consider incorporating some previously validated high-quality episodes during the adapter update process? I believe this could further enhance the performance of SOLO.

If my concerns are addressed, I will consider improve my score.

2 comments

Thanks for the detailed response. The rebuttal has addressed my main concerns. I will raise my score.

Thank you for the follow-up and for updating your score. We are glad that our rebuttal and revision addressed your concerns, and we appreciate the time and care you put into reviewing our paper. Your feedback has helped us improve it.

Reviewer_G5gVOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes "fully test-time adaptation" for GUI agents (no ground truth, one attempt per task, no separate learning phase) and introduces SOLO, in which a judge selects successful episodes, failed episodes have their prefixes relabeled with the subtask they completed, and a LoRA adapter is updated online through self-distillation over a sliding window. On three recurring task streams it improves success rates by 3–6 points, but only ties the baseline on MobileWorld. The main weaknesses are the lack of a held-out set, so memorization of repeated instances cannot be ruled out; the unmotivated third constraint, with no log-and-retrain baseline; and limited gains that depend on the judge's accuracy.

Critical strengths
  1. A genuinely new and cleanly formalized setting. Codifying action irreversibility as the one-attempt constraint, and organizing prior work (Tab. 1) around which deployment resource each method silently consumes, is a real contribution in itself, and the analogy to Tent's fully test-time adaptation for perception models is well argued.
  2. Careful method engineering with mechanistic ablations, and honest reporting. The hindsight-relabeling branch is shown to carry most of the gain; the diagnosis of one-hot imitation (episode lengthening round by round, three quarters reaching the step budget) as over-commitment is a substantive finding, not a table hash; and the paper openly reports judge precision around 0.51–0.66 and its tie with AWM-online on MobileWorld.
Critical weaknesses
  1. Every instance in the stream is both adapted on and scored, three times over. There are no withheld instances, sites, or templates, and no transfer test. The results therefore cannot separate "adapted to the deployment distribution" from "got better at instances it had already seen." Given the small gain (20.8→25.8 over 324 episodes, at 1–3 gpt-5-mini calls per episode), the latter is the more natural reading. The authors could easily test this by withholding some templates and evaluating on them after adaptation, but this experiment is missing.

  2. "No ground truth" and "one attempt per occurrence" are well motivated, but "no learning phase other than deployment" is not. Real deployments already log trajectories, run the same judge offline, batch-finetune, validate, and redeploy, using exactly the data and judge SOLO uses. The paper excludes this route by definition and compares only against baselines excluded with it. It also never identifies when online adaptation would beat log-and-retrain. Meanwhile SOLO adds per-episode judge calls and unvalidated live weight updates in production, for a 3–6 point gain on an author-constructed stream. A log-and-retrain baseline using the same data and judge is needed to justify the setting.

Questions for authors
  1. Signal-specificity control (clarification): Under the same window, schedule, and learning rate, replace judged admissions with random admissions of the agent's own episodes. How much of the 3–6 point margin survives? As the results stand, a generic small-LoRA self-distillation effect cannot be ruled out as the cause of the gain.
  2. Construct a stream that withholds a subset of templates/instances, adapt on the rest, and evaluate performance on the withheld tasks the agent never encountered. The current design trains and tests on the same 27–39 templates, so I cannot tell adaptation apart from familiarity with the specific instances.
1 comment

Dear Reviewer G5gV,

With the discussion period ending soon, we would like to check in on our response to your review. We have run the three experiments you requested: the held-out template experiment (W1, Q2), the log-and-retrain baseline with the same data and judge (W2), and the random-admission control (Q1). All three are in App. G of the revised PDF, marked in blue (Tab. 18, Tab. 16 and Tab. 17).

Please let us know if you have any further questions; we will be glad to answer them.

Thank you again for your time and your review.

Reviewer_f6A7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The work proposes a test-time adaptation method that lets GUI agents improve their own weights on the job. A judge model picks which episodes the agent should learn from, and each update takes a LoRA step toward the policy’s own renormalized top-4 tokens.

Critical strengths

None

Critical weaknesses
  • I find it concerning that the judge’s quality does not seem to affect the agent’s final performance. For example, gpt-5 is a stronger judge than gpt-5-mini (higher precision on both streams in Tab. 4), yet SOLO with gpt-5 as the judge achieves a lower success rate. Likewise, replacing the judge with the benchmark’s own evaluator (Tab. 12) gives no improvement over gpt-5-mini. This suggests the gain may come mostly from the top-K self-distillation itself rather than from the judge’s selection.
  • The decoding strategy is not reported, and it matters for interpreting the results. If decoding is greedy, the top-K update preserves the ranking of tokens, so the argmax at the trained contexts cannot change, and any gain would have to come from generalization to other contexts. If decoding uses sampling, the gain might simply come from sharpening the distribution, which is similar to lowering the temperature. A frozen agent with greedy or top-4 decoding would then be a necessary baseline.
Questions for authors

None

1 comment

Dear Reviewer f6A7,

As the discussion period closes soon, we would like to follow up on our response to your review. For your two points, we added a control that removes the judge while keeping top-K self-distillation (W1), and a baseline using the frozen agent with top-4 sampling (W2). Both are in the revised PDF, marked in blue (Tab. 17 in App. G and Tab. 11 in App. D). The revision also states that decoding is greedy (Sec. 4.1) and further discusses why a higher judge precision does not necessarily yield a higher success rate (Sec. 4.4).

Please let us know if you have any further questions; we will be glad to answer them.

Thank you again for your time and your review.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Discussion

We thank the reviewers for their reviews and summarize the rebuttal and the discussion below.

Recognition of the work. The reviews recognize a genuinely new and cleanly formalized setting (G5gV) that is practically grounded and differentiated from existing paradigms (SZrw); careful method engineering with mechanistic ablations, a substantive diagnosis of one-hot imitation, and honest reporting (G5gV); and gains on every stream (SZrw).

Initial concerns and our responses. We answered each concern with a new experiment or analysis, all in the revised manuscript and marked in blue.

  1. Does the gain reflect adaptation, or familiarity with instances already seen? (G5gV, SZrw) On templates withheld from every update, SOLO still gains over the frozen agent: a success rate of 26.7% against 20.1% on WebArena, and 15.7% against 12.6% on VisualWebArena (Qwen3-VL-8B, Tab. 18).
  2. Motivation of "no learning phase other than deployment", and no log-and-retrain baseline (G5gV). Log-and-retrain without a validation step is inside the setting: the interval between updates is a design choice. We ran it with the same judge and a retrain after every round, and it keeps less than half of SOLO's gain (Qwen3-VL-8B, Tab. 16).
  3. The gain may come from the self-distillation itself, not from the judge's selection (f6A7, G5gV, SZrw). With random admissions in place of judged ones, SOLO's 6.1-point gain over the frozen agent on VisualWebArena falls to 0, and its 4.8-point gain on WebArena to 1.4 points (Qwen3-VL-8B, Tab. 17). Sec. 4.4 now explains why a more precise judge does not necessarily give a higher success rate.
  4. Decoding strategy not reported (f6A7). Decoding is greedy, now stated in Sec. 4.1, and top-4 sampling does not lift the frozen agent to SOLO's level (UI-TARS-7B, Tab. 11).
  5. Forgetting, cost, and ablations without MobileWorld (SZrw). SOLO repeats a larger share of its wins in the next round than the frozen agent (Tab. 10). The revision also adds a cost comparison with the baselines (Tab. 15) and MobileWorld to the ablation and auxiliary-model studies (Tab. 3, Tab. 4).

Discussion. Reviewer SZrw acknowledged our responses and raised the rating to 3 (weak accept). For Reviewers f6A7 and G5gV, every requested experiment has been run, and their concerns (items 1 to 4) are addressed in our responses and the revised manuscript.

Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.