Paper deep dive
Test-Time Self-Evolving GUI Visual Grounding via Reflection-Guided On-Policy Self-Distillation
Shiyu Xuan, Zechao Li
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2608.11191v1
- Canonical: https://arxiv.org/abs/2608.11191v1
Trouble viewing inline? Open PDF directly →
Full Text
53,300 characters extracted from source content.
Expand or collapse full text
Test-Time Self-Evolving GUI Visual Grounding via Reflection-Guided On-Policy Self-Distillation Shiyu Xuan 1 , Zechao Li 1 1 School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China, 210094 shiyu_xuan, zechao.li@njust.edu.cn Abstract 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 de- ployment without human-annotated ground truth. It constructs a closed-loop of Exploration, Evaluation, Reflection, and In- ternalization. Specifically, the agent first explores unseen in- terfaces by predicting grounding coordinates for given instruc- tions. 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 con- ditioned self-teacher. Furthermore, we design a Contrastive Calibration method to prevent incorrect auto-regressive pre- fixes from corrupting the supervisory signals during failed explorations. Extensive experiments across six benchmarks demonstrate our framework’s effectiveness, achieving an av- erage 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 capa- bility of GUI agents. The code will be released. Introduction GUI agents capable of navigating and interacting with digital environments have recently emerged as a promising direction for human-computer interaction (Xie et al. 2025). As a fun- damental capability of GUI agents, GUI Visual Grounding aims to localize UI elements according to natural-language instructions. Recent advances have substantially improved grounding performance through large-scale supervised fine- tuning (Qin et al. 2025; Cheng et al. 2024), specialized action heads (Wu et al. 2025; Lin et al. 2025), and reinforcement learning (Luo et al. 2025a; Yuan et al. 2025). Despite these advances, existing GUI grounding models remain static after deployment, with model parameters frozen once training is completed. Consequently, when encounter- ing unseen applications or interface layouts, their perfor- Enabling the model to learn from both successful and failed explorations via R-OPSD. Evaluation Result GUI Screenshot Grounding Model Failed Grounding (e.g., [100, 200]) Incapable of learning after deployment (a) Static Methods Instruction: Click the button “Next” GUI Screenshot Grounding Model [100, 200] Incapable of reflection and cannot instruct the model on why a failure occurred. (b) TTRL Methods Instruction: Click the button “Next” [120, 200] [120, 240] [400, 300] Sparse Reward GUI Screenshot Grounding Model [100, 200] (c) Ours Instruction: Click the button “Next” Reflector Reasoning Reflection Dense Per-token Feedback R-OPSD Figure 1: (a) Static methods operate on frozen parameters af- ter deployment and cannot learn from exploration. (b) TTRL Methods adapt the model using sparse scalar rewards, but lack a mechanism to reason about failures and internalize textual reflections. (c) Test-Time Self-Evolving introduces a Reflector to provide detailed textual reflections, which are translated into dense token-level feedback via R-OPSD, en- abling the model to self-evolve. mance cannot improve through interaction. A recent test- time adaptation method (Du et al. 2026) attempts to alleviate this limitation by optimizing the model online with reinforce- ment learning. This method relies on sparse scalar rewards that merely indicate whether an exploration succeeds. Such feedback provides little information about why a prediction fails or how the model should correct it. arXiv:2608.11191v1 [cs.CV] 11 Aug 2026 A natural solution is to train the model with the reflec- tion on its own explorations. Rather than merely assigning a success or failure signal, reflection explains why a pre- diction is incorrect by identifying the intended UI element, examining the predicted location, and diagnosing the failure. Such reasoning provides richer supervisory information than scalar rewards. However, existing reinforcement learning al- gorithms cannot directly leverage these textual reflections based on scalar rewards. To overcome this, we propose a Test-Time Self-Evolving framework. It constructs a closed-loop that equips the agent with four core capabilities: Exploration in unknown inter- faces, Evaluation of generated coordinates, Reflection upon evaluation results, and Internalization of these reflections into its parameters. Specifically, upon deployment, the agent first explores the interface with instructions by predicting grounding coordinates. To evaluate this exploration, we in- troduce an MLLM-based Reflector that estimates whether the prediction is consistent with the instruction while simul- taneously generating a reflection explaining its evaluation. The remaining challenge is how to internalize the reflec- tion into the model parameters. Unlike scalar rewards, textual reflections cannot be directly exploited by policy optimiza- tion algorithms. To bridge this gap, we propose Reflection- Guided On-Policy Self-Distillation (R-OPSD). Instead of learning from scalar rewards, R-OPSD conditions a self- teacher on the Reflector’s evaluation results and reflections, converting high-level reasoning into dense token-level super- vision over the generated coordinate tokens. However, this dense feedback encounters a bottleneck in the auto-regressive generation of coordinates. During failed explorations, the model predicts incorrect tokens. If the teacher model conditions its subsequent supervision on these prefixes, the supervisory signals gradually become unreli- able. To mitigate this issue, we propose a Contrastive Cali- bration Method. It treats a failed exploration as successful by using an inverse-prompted student. The trajectory is sampled from the model’s own policy, conditioning it on a success prompt amplifies its confidence in this specific sequence. Consequently, at the initial error step, the inverse student assigns a high probability to the incorrect token, while the reflection-guided teacher yields a lower probability. This dis- crepancy produces a negative advantage that suppresses the initial error. As generation continues, the corrupted prefix dominates both models, aligning their output distributions. This decays the advantage to almost zero, preventing cor- rupted feedback from being internalized. Extensive experiments across ScreenSpot and ScreenSpot- v2 (Cheng et al. 2024), ScreenSpot-Pro (Li et al. 2025), MMBench-GUI (Wang et al. 2025), OSWorld-G and OSWorld-G-Refine (Xie et al. 2025) demonstrate the effec- tiveness of our framework. By enabling self-evolving with- out ground truth annotations, our method achieves an av- erage accuracy improvement of 7.4% over the base model. Furthermore, compared to the recent test-time reinforcement learning method GUI-RCPO, our method achieves an aver- age performance gain up to 7.7%. In summary, our main contributions are three-fold: (a) We introduce a Test-Time Self-Evolving framework, which empowers grounding agents with exploration, evaluation, re- flection, and parameter internalization in unseen environ- ments. (b) We propose Reflection-Guided On-Policy Self- Distillation, which converts textual reflections into dense token-level supervision, together with a Contrastive Calibra- tion method that alleviates corrupted supervision caused by incorrect auto-regressive prefixes. (c) Extensive experiments and ablations validate our framework. This is the first work to applying OPSD in test-time scenarios of GUI grounding, establishing a promising direction for GUI grounding. Related Works GUI Visual Grounding. GUI visual grounding locates UI elements based on natural language instructions. Re- cent advancements span three paradigms. Supervised fine- tuning methods establish baseline alignment using large- scale datasets (Qin et al. 2025; Cheng et al. 2024; Gou et al. 2024). To enhance visual perception, models selectively ex- tract UI-related visual tokens (Lin et al. 2025; Ouyang et al. 2026) or employ attention-based action heads to bypass di- rect coordinate generation (Wu et al. 2025). Reinforcement learning methods further enhance the performance. Recent works adapt GRPO (Shao et al. 2024) with rule-based coor- dinate rewards (Luo et al. 2025a; Lu et al. 2026), or design complex reward functions incorporating attention maps, box sizes, and bounding-box distributions (Yuan et al. 2025; Zhou et al. 2025; Tang et al. 2026). Test-time scaling methods al- locate additional inference compute to enhance performance. RegionFocus (Luo et al. 2025b) dynamically zooms into the image for coarse-to-fine UI grounding, a process similarly guided by attention maps in ZoomUI (Liu et al. 2026). Ad- ditionally, GUI-RC (Du et al. 2026) aggregates multiple pre- dictions via region consistency, further refining the model with test-time reinforcement learning (Zuo et al. 2026). On-policy Distillation. To internalize reasoning capabilities into model weights, on-policy distillation (OPD) trains the student model with token-level feedback on its own gener- ation trajectories. Typically, this feedback is derived either from a stronger teacher model (Gu et al. 2024; Agarwal et al. 2024) or a model conditioned on privileged informa- tion (Zhang et al. 2026a; Zhao et al. 2026). However, incor- rect prefixes generated by the student can degrade the quality of the feedback. To address this issue, ESR (Ziheng et al. 2026) proposes an early stopping rollout mechanism, while EOPD (Jin et al. 2026) assesses the reliability of the teacher model using token entropy. Alternatively, RLSD (Yang et al. 2026) and RLCSD (Pan et al. 2026) leverage token-level feed- back to only determine update magnitudes. The potential of OPD has also been explored across various vision-language tasks, such as Video Temporal Grounding (Li et al. 2026), Vi- sual Question Answering (Yuan et al. 2026), and GUI Visual Grounding (Zhang et al. 2026b). Unlike test-time reinforcement learning methods (Du et al. 2026; Zuo et al. 2026) that rely on sparse scalar rewards and fail to leverage reflection reasoning, our framework empow- ers the model to learn from rich reflections via R-OPSD. Furthermore, Contrastive Calibration effectively migrates the corrupted supervision issue caused by incorrect prefixes through suppressing initial errors and decaying the policy gradients of drifted tokens, ensuring that the model can safely internalize knowledge from failed explorations. Methodology Overview Given a screenshot I and a language instruction L, the GUI grounding model π G is required to locate the corresponding UI element B = π G (I,L), where B denotes the coordinates of the bounding box [x 1 ,y 1 ,x 2 ,y 2 ] or the center point (x,y) of the target element. Existing GUI grounding models typi- cally remain static after deployment, with model parameters frozen once training is completed. Consequently, when de- ployed on unseen platforms or applications, the model cannot improve from its own interaction experience. To enable post-deployment adaptation without human an- notations, we propose a Test-Time Self-Evolving framework that constructs a closed-loop consisting of four stages: Ex- ploration in unknown interfaces, Evaluation of its generated coordinates, Reflection upon its evaluation result, and Inter- nalization of these reflections into its parameters. Specifi- cally, this framework executes an episodic loop: • Exploration: The GUI grounding model π G explores the unseen interface, i.e., given the UI screenshot I and in- struction L, B = π G (I,L). • Evaluation & Reflection: To assess the exploration in the absence of ground truth, we introduce a Reflector π R . It takes the screenshot, instruction, and predicted coor- dinates as input to output an evaluation score S and a detailed reasoning process R, S,R = π R (I,L,B),(1) where S ∈ 0, 1 indicates whether the current explo- ration is successful, and R represents the reasoning be- hind this evaluation. Subsequently, S and R are utilized to guide the internalization. • Internalization: The reflection knowledge is then in- ternalized into π G through Reflection-Guided On-Policy Self-Distillation. Specifically, the evaluation score and reasoning process are leveraged as privileged informa- tion to construct a self-teacher that provides per-token feedback on the generated coordinates. Furthermore, to alleviate failed explorations from corrupting the feedback, we propose a Contrastive Calibration Method. In the following parts, we will first introduce the Evalu- ation & Reflection stages, and further propose the Internal- ization stage. The overall framework is illustrated in Fig.2. Evaluation and Reflection with Reflector During test-time adaptation, ground-truth annotations are un- available. Therefore, the model requires an alternative mech- anism to estimate whether its exploration is successful. To this end, we introduce an MLLM-based Reflector π R . Its role is to evaluate the alignment between the predicted coordi- nates B and the language instruction L within the context of the visual screenshot I. Besides producing an estimated evaluation result, it also generates a reflection explaining the reasoning behind its judgment. This reflection serves as priv- ileged information for the subsequent internalization stage. Inference. To achieve this goal, we design a prompt template that forces the Reflector to engage in step-by-step reasoning before outputting a final result: Reflector Prompt Template Evaluate if the given click coordinate correctly executes the user instruction on the provided UI image. Answer with Yes or No. Instruction: L Click Coordinate: B Before providing your final answer, you must think step-by- step. Address the following in your reasoning: 1. Intent Analysis: What specific UI element (text, icon, but- ton) does the instruction intend to interact with? 2. Coordinate Grounding: Look at the exact location of the Click Coordinate [x,y] on the UI. What specific element or visual feature is located at that coordinate? 3. Verification: Does the element at the given coordinate match the intended element from the instruction? Is it the correct tar- get, or did it hit the background, a different element, or an empty space? 4. Conclusion: Explicitly state whether the verification passed (Yes) or failed (No). Output the thinking process in <think> </think> and final an- swer in <answer> </answer> tags. You must strictly format your output as follows: <think> Your step-by-step reasoning and conclusion here </think> <answer>Yes|No</answer> Once receiving the generated response from π R , we parse the output according to the specified tags to ob- tain R and S. Specifically, the text extracted from the <think>...</think> tags serves as the reflec- tion R, The discrete binary label extracted from the <answer>...</answer> tags serves as S. Training. To improve the reliability of the Reflector, we opti- mize it using GRPO (Shao et al. 2024) on an offline collected dataset. Each sample in this dataset comprises a UI screen- shot I, a language instruction L, a sampled click coordinate B, and a ground-truth label S ∗ ∈ 0, 1 indicating whether the coordinate successfully executes the instruction. The con- struction details of the dataset are shown in the Appendix. We employ the format and binary rewards during training, i.e., r = r format + r binary : • r format ensures the model strictly adheres to the requested output template. • r binary supervises discrimination capability: r binary = ( 1, if S = S ∗ and format is valid 0, if S ̸= S ∗ or format is invalid (2) After training, the Reflector remains frozen throughout test- time adaptation. Internalization with Reflection-Guided On-Policy Self-Distillation After obtaining the evaluation result S and reflection R, the final stage is to internalize this knowledge into the ground- Evaluate if the given click coordinate correctly ... think step-by- step ... Yes | No GUI Screenshot [150, 20] Instruction (L): Enable Wi-Fi Exploration Prediction GUI Screenshot Instruction (L): Enable Wi-Fi [150, 20] Evaluation & Reflection <think> User intends to enable Wi-Fi. The click is on the top-right of the screen. This hits a background area. </think> <answer> No </answer> Internalization GUI Screenshot Instruction (L): Enable Wi-Fi S Privileged Information R Reflection-guided Teacher Template Inverse-prompted Student Template v.s. + [150, 20] Self-Evolving Per-token Advantage Figure 2: Test-Time Self-Evolving Framework consists of four core stages: Exploration, Evaluation, Reflection, and Internal- ization. During test-time adaptation, π G explores the environment to predict coordinates. The Reflector π R then assesses the prediction to generate a binary score S and a step-by-step reasoning R. The reflection knowledge is internalized via Reflection- Guided On-Policy Self-Distillation to update the grounding policy. To optimize memory efficiency, both π G and π R share a base model and we alternate between the roles of π G and π R by switching the active LoRA adapters. ing model π G . The remaining challenge is how to transform textual reflection into effective optimization signals. Unlike reinforcement learning, which is based on scalar rewards, re- flection consists of free-form natural language and therefore cannot be directly exploited by standard policy optimiza- tion algorithms. To bridge this gap, we propose Reflection- Guided On-Policy Self-Distillation (R-OPSD), which con- verts reflection into dense token-level supervision. R-OPSD. Given a rollout B ∼ π G (·| I,L) generated by the policy, OPD trains the policy π G on its own rollout with the token-level supervisory signals provided by a teacher model π T . Following the standard policy optimization formulation, the token-level advantage is defined as: a i = log π T (B i | B <i ,I,L) π G (B i | B <i ,I,L) .(3) The optimization objective maximizes the expected return, defined as the sum of token-level advantages: L OPD =− 1 T T X i=1 sg(a i ) logπ G (B i | B <i ,I,L),(4) where sg(·) denotes the stop-gradient operator. To internalize the knowledge from the Reflector, we con- struct a self-teacher by utilizing the π G itself conditioned on the evaluation result S and reflection R as the privileged information. Conditioning on this additional information en- ables the teacher to produce a more informative conditional distribution over the generated coordinate sequence. Specif- ically, the input prompt L(S,R) is formulated as follows, Reflection-guided Teacher Prompt Template If Success (S = 1) The previous prediction B has been verified to correctly fulfill the instruction. Outline the position corresponding to the instruction:L. The output should be only [x1,y1,x2,y2]. If Failure (S = 0) Previous wrong prediction:B Feedback:R Based on the feedback, outline the position corresponding to the instruction:L. The output should be only [x1,y1,x2,y2]. The token-level advantage is then computed as: a i = log π G (B i | B <i ,I,L(S,R)) π G (B i | B <i ,I,L) .(5) Contrastive Calibration Method. MLLM-based GUI grounding is formulated as an auto-regressive generation of coordinate tokens. Performing R-OPSD on failed explo- rations introduces a challenge. If the teacher model condi- tions its subsequent feedback on these incorrect prefixes, its predicted probabilities become meaningless. To mitigate this, we propose a Contrastive Calibration Method (C) to calibrate the learning advantage based on the state of the generated prefix: it suppresses the initial incorrect token and decays the advantage to zero as the tokens progressively drift. To achieve this, we introduce an inverse- prompted student to contrast against the teacher. Specifically, for failed explorations, we construct a contrastive prompt L(¬S) to misinform the model that the prediction is correct, ... [1 ,201,39 ] Predicted Tokens ... Student Probability ... Teacher Probability ... Inverse-prompted Student Probability ... Almost Zero Per-token Advantage w/ C Suppress the Initial Incorrect Tokens ... Corrupted Feedback Per-token Advantage w/o C Meaningless Probability Figure 3: Illustration of Contrastive Calibration Method. Dur- ing auto-regressive generation, an incorrect prediction, e.g., the initial incorrect token ‘0‘, causes subsequent tokens to drift. Without C, computing advantages on these drifted tokens results in corrupted feedback. By contrast, C lever- ages an inverse-prompted student. This produces a negative penalty to suppress the initial incorrect tokens. As errors accumulate, the conditioned incorrect prefix forces both the teacher and the inverse-prompted student’s distributions to align, decaying the advantage to near zero. Inverse-prompted Student Prompt Template If Failure (S = 0) Previous correct prediction:B Based on the feedback, outline the position corresponding to the instruction:L. The output should be only [x1,y1,x2,y2]. With L(¬S), we reformulate the advantage in Eq.(5) into the contrastive calibration advantage: a i = log π G (B i | B <i ,I,L(S,R)) π G (B i | B <i ,I,L(¬S)) .(6) Note that C is exclusively applied to failed explorations, as successful samples (S = 1) avoid the incorrect prefix issue. As shown in Fig. 3, C addresses the incorrect prefix is- sue across two phases. First, when the prefix B <i is correct, but the current token B i is incorrect, the inverse-prompted student under L(¬S) is misled to treat the prediction as successful, assigning a high probability to B i . In contrast, the teacher under L(S,R) leverages the reflection to pre- dict the true target, yielding a lower probability for B i , producing a negative advantage a i < 0, driving the policy gradient to suppress this token. Second, As more incorrect tokens are generated, both models are conditioned on in- creasingly similar prefixes, leading to similar probabilities Initial Incorrect Token ( ̄a t ) Drifted Last Token (| ̄a t |) −2 0 2 −0.39 1.34 −0.97 0.0033 Average Advantage w/o CCw/ C Figure 4: Quantitative analysis of token-level advantage on failed explorations (S = 0). We report the average advantage ( ̄a t ) at the initial incorrect token, and the average absolute advantage (| ̄a t |) at the drifted last tokens. π G (· | L(S,R)) ≈ π G (· | L(¬S)), driving the advantage to almost zero (a i → 0), thereby gradually reducing the mag- nitude of the token-level advantage. Discussion. To validate C, we analyze token-level advan- tage during failed explorations by calculating the average advantage ( ̄a t ) at the initial incorrect token and the average absolute advantage (| ̄a t |) at the drifted last token. As shown in Fig. 4, without C, R-OPSD applies a marginal penalty at the initial error and generates corrupted advantages| ̄a t | = 1.34 on drifted tokens. In contrast, C generates a strong negative penalty at the initial incorrect token. As tokens progressively drift, the update magnitude decays towards zero, preventing the internalization of corrupted supervision. Direction-based Advantage Clamping. Directly applying token-level feedback of R-OPSD can sometimes destabi- lize training. To mitigate this, the token-level updates should be aligned with the evaluation result, i.e., if a prediction is successful, its probability should be encouraged. If it fails, its probability should be suppressed. Therefore, we apply direction-based clamping to the advantage, a i = max(0,a i ), S = 1 min(0,a i ), S = 0 .(7) In addition, this token-level advantage can be seamlessly integrated with the query-level advantage from GRPO. Given a sampled group of size K, the GRPO reward is defined as r GRPO = r format + r correctness , where r correctness is evaluation S of the predicted coordinates. The query-level advantage ˆ A GRPO [k] is computed by standardizing these rewards: ˆ A GRPO [k] = r GRPO [k]− meanr GRPO [1],...,r GRPO [K] stdr GRPO [1],...,r GRPO [K] . (8) By incorporating the token-level advantage into the GRPO advantage, the final integrated objective is: ˆ A[k] = ˆ A GRPO [k] + λa i ,(9) where λ controls the integration strength. Although the proposed framework relies on the Reflector to estimate grounding correctness, it does not require perfect evaluation. Instead, R-OPSD only assumes that the Reflector provides sufficiently informative evaluations and reflections to improve the conditional distribution of the teacher. As demonstrated in our experiments, this assumption is suf- ficient to improve grounding performance across multiple GUI benchmarks consistently. We also provide quantitative and qualitative evaluations of the Reflector in the Appendix. Experiments Experimental Setup Evaluation Benchmarks and Metrics. We conduct exper- iments across diverse GUI visual grounding benchmarks covering various platforms and instruction complexities: ScreenSpot and ScreenSpot-v2, ScreenSpot-Pro, MMBench- GUI, as well as OSWorld-G and OSWorld-G-Refine. We adopt Element Accuracy as the evaluation metric, where a prediction is correct if the predicted point falls within the bounding box of the target UI element. Implementation Details. Our method is built upon the Qwen2.5-VL (Team 2025) and Qwen3-VL (Bai et al. 2025). To optimize memory efficiency, both the grounding model (π G ) and the Reflector (π R ) share a base model and are fine- tuned using LoRA (Hu et al. 2022). During training, we alternate between the roles of π G and π R by switching the ac- tive LoRA adapters. This parameter-efficient design restricts the GPU memory to approximately 10GB for 3B/2B mod- els and 30GB for 7B/8B models. During the rollout phase, we set the temperature to 1.0 and top-p to 0.95. For the Re- flector π R , we train for 1 epoch via GRPO on 10K samples from GroundCUA (Feizi et al. 2025), with a learning rate of 1× 10 −4 , a group size of 8, and a batch size of 64. For the internalization stage of π G , R-OPSD integrates the GRPO advantage with a strength coefficient λ = 0.2. The model is trained for 2 epochs with a learning rate of 1× 10 −4 and a batch size of 64. To simulate post-deployment self-evolving, we leverage data from ScreenSpot-v2 or MMBench-GUI, strictly omitting their ground-truth annotations. Main Results The evaluation results across six GUI visual grounding benchmarks are presented in Table 1. Overall, our proposed Test-Time Self-Evolving framework consistently enhances the grounding capabilities of the base models, outperforming the existing test-time adaptation method, e.g., GUI-RCPO. Our method demonstrates performance gains across different base models. When deployed on the unseen SSv2 dataset, our framework boosts the average accuracy of Qwen2.5-VL-3B from 50.2% to 57.4%. When applied to the highly capable Qwen3-VL-2B backbone, our self-evolving framework still raises the average accuracy to 69.4% (+3.7%) and 70.3% (+4.6%) when adapting on SSv2 and MMG, respectively. This validates that our framework achieves the self-evolving without requiring ground-truth. Adapting on complex datasets like MMG frequently re- sults in rollout groups where all test-time explorations fail. GRPO used in GUI-RCPO relies on relative rewards, they fail to optimize from failed groups, causing negative transfer. In contrast, R-OPSD overcomes this limitation by leveraging re- flections to provide token-level feedback, guiding the model even from complete failures. Supported by C to filter out corrupted supervision, our framework achieves performance gains (+7.4% and +4.6%) with MMG. Ablation Study Qualitative and quantitative analysis of the Reflector, compu- tational overhead analysis, scalability to larger base models, 00.10.20.30.40.5 80 82 84 86 88 90 Accuracy (%) SSv2 SSv2 MMG 00.10.20.30.40.5 58 62 66 70 MMG SSv2 MMG Figure 5: Sensitivity analysis of the integration strength λ. The subplots indicate the evaluation datasets. The colored lines denote the adaptation datasets. and comparisons with visual bounding box prompting can be found in the Appendix. Effectiveness of Each Component. To validate the effective- ness of each module in our Test-Time Self-Evolving frame- work, we conduct ablation studies in Table 2. We establish (a) as our baseline, which applies OPSD using only evaluation results as privileged information. This naive baseline suf- fers from catastrophic policy collapse. Comparing (d)(e) and (i)(j), incorporating Reflection yields a substantial perfor- mance gain, e.g., when adapting on SSv2, Reflection boosts the accuracy from 24.6% to 28.5% on SSP. Utilizing token- level feedback in auto-regressive generation introduces the incorrect prefix issue. (f) utilizes reflection but lacks C, causing it to collapse. By incorporating C, (g) filters out this corrupted supervision, turning the negative transfer into massive gains, reaching 64.3% on MMG. A similar substan- tial boost is observed in (h) vs. (i), confirming C is impor- tant for learning from failed explorations. (i) demonstrates that direction-based advantage clamping prevents token-level updates from contradicting the evaluation result, providing consistent improvements. Building upon this, (j) shows the benefit of integrating the query-level advantage. Necessity of R-OPSD over Standard RL. To demonstrate the necessity of R-OPSD, we compare it against GRPO us- ing various scalar rewards in Table 3. Following TTRL (Zuo et al. 2026), relying on averaged predicted bounding boxes (Avg B) as pseudo-labels suffers negative transfer on com- plex datasets like MMG. This occurs because when some predictions drift towards visually similar but incorrect UI el- ements, averaging these bounding boxes leads to completely invalid pseudo-labels. The evaluation resultS from Reflector is more robust to this situation, leading to better performance. This also confirms the high reliability of the Reflector. How- ever, GRPO guided by the binary S reward remains limited by the sparsity of scalar signals and cannot instruct the model on how to correct specific coordinate tokens. Furthermore, to verify that our gains stem from the reflec- tion design rather than OPSD, we evaluate GUI-SD (Zhang et al. 2026b), which draws the bounding box in the image as the privileged information. Its severe performance degra- dation confirms that drawing invalid pseudo-labels directly corrupts the visual context. In contrast, R-OPSD leverages both the evaluation result and the detailed textual reflection (S,R), achieving the highest accuracy. Sensitivity Analysis of Integration Strength. We investi- gate the impact of λ, which controls the magnitude of the distillation advantage during policy updates. As shown in Method Adaptation Datasets SSSSv2 SSP OSW-G OSW-GR MMGAvg. GUI-specific Models UGround-V1-2B (Gou et al. 2024)-77.778.822.7---- ShowUI-2B (Lin et al. 2025)-75.1------ GUI-Actor-3B (Wu et al. 2025)-89.791.042.2---- FocusUI-3B (Ouyang et al. 2026)--91.543.853.4--- InfiGUI-R1-3B (Liu et al. 2025)-87.5-35.7---- SE-GUI-3B (Yuan et al. 2025)---35.9---- GUI-G 2 -3B (Tang et al. 2026)-90.891.037.8---- Test-Time Adaptation Qwen2.5-VL-3B (Team 2025)-78.380.420.327.137.757.550.2 + GUI-RCPO (Du et al. 2026)SSv282.385.424.829.338.460.253.4+3.2 + OursSSv285.988.830.532.639.766.657.4+7.2 + GUI-RCPO (Du et al. 2026)MMG79.280.619.825.736.158.249.9-0.3 + OursMMG84.687.530.133.741.768.257.6+7.4 Qwen3-VL-2B (Bai et al. 2025)-83.587.642.847.360.872.265.7 + GUI-RCPO (Du et al. 2026)SSv284.789.346.351.361.372.867.6+1.9 + OursSSv288.591.647.652.161.275.469.4+3.7 + GUI-RCPO (Du et al. 2026)MMG84.286.942.346.861.871.665.6-0.1 + OursMMG88.291.549.352.263.277.670.3+4.6 Table 1: GUI grounding accuracy on six benchmarks including ScreenSpot-V2 (SSv2), ScreenSpot-Pro (SSP), OSWorld-G (OSW-G), OSWorld-G_R (OSW-GR), and MMBenchGUI (MMG). Bold indicates the best results. R C Clamp QA Adapt on SSv2Adapt on MMG SSv2 SSP MMGSSv2 SSP MMG Base Model80.4 20.3 57.580.4 20.3 57.5 (a)× × × × × (b) √ 81.2 21.4 58.780.8 21.1 59.3 (c) √ 79.5 18.0 57.278.6 18.3 57.1 (d) √ √ 84.2 24.6 61.782.3 23.5 61.5 (e) √ √ √ 84.2 25.4 62.682.8 23.4 62.3 (f) √ × × × × × (g) √ √ 84.0 26.8 62.984.0 27.7 64.3 (h) √ √ 82.1 21.5 58.580.2 19.8 57.5 (i) √ √ √ 86.8 28.5 64.286.0 29.9 67.4 (j) √ √ √ √ 88.8 30.5 66.687.5 30.1 68.2 Table 2: Ablation study on each component using Qwen2.5- VL-3B when adapting on SSv2 and MMG, respectively. R: Reflection R; C: Contrastive Calibration Method; Clamp: Direction-based advantage clamping; QA: Integrating query- level advantage.× denotes suffering from policy collapse. Fig. 5, λ = 0.2 leads to the best performance, regardless of the adaptation and evaluation datasets. Conclusion In this paper, we introduce a Test-Time Self-Evolving frame- work for GUI visual grounding that enables post-deployment adaptation without requiring ground truth annotations. Our framework introduces a closed-loop of Exploration, Eval- uation, Reflection, and Internalization, allowing the model to learn from both the outcomes of its explorations and RewardSupervision Adapt on SSv2Adapt on MMG SSv2 SSP MMGSSv2 SSP MMG Base Model80.4 20.3 57.580.4 20.3 57.5 GUI-RCPO-85.4 24.8 60.280.6 19.8 58.2 Binary Avg. B84.3 24.6 61.779.1 21.2 57.5 IoUAvg. B84.7 22.4 61.380.8 20.4 59.4 GUI-G 2 Avg. B83.0 21.2 59.978.6 19.4 55.8 Binary S85.4 25.9 62.384.2 25.7 62.8 GUI-SDAvg. B78.5 16.7 52.3× × × OursS,R88.830.566.687.530.168.2 Table 3: Comparison of the performance between GRPO with various scalar rewards and R-OPSD. The “Supervision” denotes the source of feedback used for optimization: Avg. B indicates that the rewards are derived from aggregated bounding box predictions within a rollout group; S is the evaluation results from the Reflector. the reasons behind its successes and failures. To internal- ize such reflection knowledge, we propose R-OPSD, which transforms natural-language reflections into token-level su- pervision through a self-teacher conditioned on privileged information. Furthermore, we introduce a Contrastive Cali- bration method to alleviate corrupted supervision caused by incorrect auto-regressive prefixes during failed explorations. Extensive experiments across diverse benchmarks and abla- tions demonstrate that our framework can improve the per- formance of the model in a self-evolving manner. References Agarwal, R.; Vieillard, N.; Zhou, Y.; Stanczyk, P.; Ramos Garea, S.; Geist, M.; and Bachem, O. 2024. On- policy distillation of language models: Learning from self- generated mistakes. In ICLR, volume 2024, 21246–21263. Bai, S.; Cai, Y.; Chen, R.; Chen, K.; Chen, X.; Cheng, Z.; Deng, L.; Ding, W.; Gao, C.; Ge, C.; et al. 2025. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631. Cheng, K.; Sun, Q.; Chu, Y.; Xu, F.; YanTao, L.; Zhang, J.; and Wu, Z. 2024. Seeclick: Harnessing gui grounding for advanced visual gui agents. In ACL, 9313–9332. Du, Y.; Yan, Y.; Tang, F.; Lu, Z.; Zong, C.; Lu, W.; Jiang, S.; and Shen, Y. 2026. Test-time reinforcement learning for gui grounding via region consistency. In AAAI, volume 40, 30593–30601. Feizi, A.; Nayak, S.; Jian, X.; Lin, K. Q.; Li, K.; Awal, R.; Lù, X. H.; Obando-Ceron, J.; Rodriguez, J. A.; Chapados, N.; et al. 2025. Grounding Computer Use Agents on Human Demonstrations. arXiv preprint arXiv:2511.07332. Gou, B.; Wang, R.; Zheng, B.; Xie, Y.; Chang, C.; Shu, Y.; Sun, H.; and Su, Y. 2024. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243. Gu, Y.; Dong, L.; Wei, F.; and Huang, M. 2024. Minillm: Knowledge distillation of large language models. In ICLR, volume 2024, 32694–32717. Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3. Jin, W.; Min, T.; Yang, Y.; Kadhe, S. R.; Zhou, Y.; Wei, D.; Baracaldo, N.; and Lee, K. 2026. Entropy-Aware On- Policy Distillation of Language Models. arXiv preprint arXiv:2603.07079. Li, J.; Yin, H.; Xu, H.; Xu, B.; Tan, W.; He, Z.; Ju, J.; Luo, Z.; and Luan, J. 2026. Video-OPD: Efficient Post- Training of Multimodal Large Language Models for Tem- poral Video Grounding via On-Policy Distillation. arXiv preprint arXiv:2602.02994. Li, K.; Meng, Z.; Lin, H.; Luo, Z.; Tian, Y.; Ma, J.; Huang, Z.; and Chua, T.-S. 2025. Screenspot-pro: Gui grounding for professional high-resolution computer use. In ACM M, 8778–8786. Lin, K. Q.; Li, L.; Gao, D.; Yang, Z.; Wu, S.; Bai, Z.; Lei, S. W.; Wang, L.; and Shou, M. Z. 2025. Showui: One vision- language-action model for gui visual agent. In CVPR, 19498– 19508. Liu, Y.; Li, P.; Xie, C.; Hu, X.; Han, X.; Zhang, S.; Yang, H.; and Wu, F. 2025. Infigui-r1: Advancing multimodal gui agents from reactive actors to deliberative reasoners. arXiv preprint arXiv:2504.14239. Liu, Z.; Feng, T.; Kang, B.; Yang, Y.; and Luo, J. 2026. Zoom to Essence: Trainless GUI Grounding by Inferring upon Interface Elements. arXiv preprint arXiv:2603.14448. Lu, Z.; Chai, Y.; Guo, Y.; Yin, X.; Liu, L.; Wang, H.; Xiao, H.; Ren, S.; Zhao, P.; Liu, G.; et al. 2026. Ui-r1: Enhancing efficient action prediction of gui agents by reinforcement learning. In AAAI, volume 40, 17608–17616. Luo, R.; Wang, L.; He, W.; Chen, L.; Li, J.; and Xia, X. 2025a. Gui-r1: A generalist r1-style vision-language action model for gui agents. arXiv preprint arXiv:2504.10458. Luo, T.; Logeswaran, L.; Johnson, J.; and Lee, H. 2025b. Visual test-time scaling for gui agent grounding. In ICCV, 19989–19998. Ouyang, M.; Lin, K. Q.; Shou, M. Z.; and Ng, H. T. 2026. FocusUI: Efficient UI Grounding via Position-Preserving Vi- sual Token Selection. arXiv preprint arXiv:2601.03928. Pan, L.; Tao, S.; Zhai, Y.; Zhang, L.; Liu, Z.; Ding, B.; Liu, A.; and Wen, L. 2026. RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation. arXiv preprint arXiv:2606.11709. Qin, Y.; Ye, Y.; Fang, J.; Wang, H.; Liang, S.; Tian, S.; Zhang, J.; Li, J.; Li, Y.; Huang, S.; et al. 2025. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326. Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open lan- guage models. arXiv preprint arXiv:2402.03300. Tang, F.; Gu, Z.; Lu, Z.; Liu, X.; Shen, S.; Meng, C.; Wang, W.; Zhang, W.; Shen, Y.; Lu, W.; et al. 2026. GUI-G 2 : Gaussian Reward Modeling for GUI Grounding. In AAAI, volume 40, 33214–33222. Team, Q. 2025. Qwen2.5-VL. Wang, X.; Wu, Z.; Xie, J.; Ding, Z.; Yang, B.; Li, Z.; Liu, Z.; Li, Q.; Dong, X.; Chen, Z.; et al. 2025. Mmbench- gui: Hierarchical multi-platform evaluation framework for gui agents. arXiv preprint arXiv:2507.19478. Wu, Q.; Cheng, K.; Yang, R.; Zhang, C.; Yang, J.; Jiang, H.; Mu, J.; Peng, B.; Qiao, B.; Tan, R.; et al. 2025. Gui- actor: Coordinate-free visual grounding for gui agents. arXiv preprint arXiv:2506.03143. Xie, T.; Deng, J.; Li, X.; Yang, J.; Wu, H.; Chen, J.; Hu, W.; Wang, X.; Xu, Y.; Wang, Z.; et al. 2025. Scaling computer- use grounding via user interface decomposition and synthe- sis. arXiv preprint arXiv:2505.13227. Yang, C.; Qin, C.; Si, Q.; Chen, M.; Gu, N.; Yao, D.; Lin, Z.; Wang, W.; Wang, J.; and Duan, N. 2026. Self-Distilled RLVR. arXiv preprint arXiv:2604.03128. Yuan, Q.; Lou, J.; Yu, X.; Lin, H.; Sun, L.; Han, X.; and Lu, Y. 2026. Vision-opd: Learning to see fine details for mul- timodal llms via on-policy self-distillation. arXiv preprint arXiv:2605.18740. Yuan, X.; Zhang, J.; Li, K.; Cai, Z.; Yao, L.; Chen, J.; Wang, E.; Hou, Q.; Chen, J.; Jiang, P.-T.; et al. 2025. Enhancing visual grounding for gui agents via self-evolutionary rein- forcement learning. arXiv preprint arXiv:2505.12370. Zhang, X.; Ding, Z.; Pan, T.; Yang, R.; Kang, C.; Xiong, X.; and Gu, J. 2026a. OPSDL: On-Policy Self-Distillation for Long-Context Language Models. arXiv preprint arXiv:2604.17535. Zhang, Y.; Wu, D.; Shen, H.; Ma, C.; and Zhou, Y. 2026b. Learn where to Click from Yourself: On-Policy Self-Distillation for GUI Grounding. arXiv preprint arXiv:2605.00642. Zhao, S.; Xie, Z.; Liu, M.; Huang, J.; Pang, G.; Chen, F.; and Grover, A. 2026. Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models. arXiv preprint arXiv:2601.18734. Zhou, Y.; Dai, S.; Wang, S.; Zhou, K.; Jia, Q.; and Xu, J. 2025. Gui-g1: Understanding r1-zero-like training for visual grounding in gui agents. arXiv preprint arXiv:2505.15810. Ziheng, Z.; Li, J.; Tang, H.; Wu, Y. N.; and Terzopoulos, D. 2026. Less is More: Early Stopping Rollout for On-Policy Distillation. arXiv preprint arXiv:2605.27028. Zuo, Y.; Zhang, K.; Sheng, L.; Qu, S.; Cui, G.; Zhu, X.; Li, H.; Long, X.; Hua, E.; Qi, B.; et al. 2026. Ttrl: Test-time reinforcement learning. NeurIPS, 38: 131459–131483. Appendix This Appendix provides the details of construction of the Re- flector training dataset, additional qualitative and quantitative analysis of the Reflector, computational overhead analysis, scalability to larger base models, extension of our method to the unsupervised training setting, and comparisons with visual bounding box prompting. Construction of the Reflector Training Dataset To train the MLLM-based Reflectorπ R to accurately evaluate test-time explorations and provide diagnostic reasoning, we construct a training dataset based on the Functional Split of GroundCUA (Feizi et al. 2025). The constructed Reflector training dataset will be released. It is designed to capture a diverse set of predictions while maintaining balanced labels, proceeding as follows: Prediction Generation. For each instruction-image pair in the GroundCUA, we employ the base model to perform infer- ence. To encourage a diverse range of predictions, we set the sampling temperature to 1.0 and sample 8 distinct coordinate predictions per instruction. Label Assignment. Each generated coordinate is evaluated against the ground-truth bounding box. A discrete binary labelS ∗ ∈0, 1 is assigned to each prediction, whereS ∗ = 1 indicates a successful grounding (the predicted point falls within the target bounding box) and S ∗ = 0 indicates a failure. Sample Filtering. We group the 8 predictions by their in- struction and apply a filtering strategy: we exclusively retain the groups that exhibit mixed results. Specifically, an in- struction is kept only if its 8 rollouts contain at least one correct prediction S ∗ = 1 and at least one incorrect predic- tion S ∗ = 0. Balanced Sampling. To ensure a balanced distribution of positive (“Yes”) and negative (“No”) evaluation results, we randomly select exactly one correct prediction S ∗ = 1 and one incorrect predictionS ∗ = 0 from the 8 generated rollouts for each retained instruction. The final Reflector training dataset consists of approxi- mately 10, 160 prediction-label pairs, yielding a 1:1 ratio of positive to negative samples. Discussion on Cold-Start SFT. Additionally, we have explored a supervised fine-tuning cold-start strategy by synthesizing 30K high-quality reasoning trajectories using Qwen3.7-plus. However, results indicated that incorporating this cold-start phase yielded no noticeable performance gain compared to directly applying RL to the base model for the training of the Reflector. More Experimental Results Reliability Evaluation of the MLLM-based Reflector. The efficacy of our Test-Time Self-Evolution framework relies heavily on the quality of the evaluation score S and diagnos- tic reasoning R provided by the Reflector π R . We evaluate the Reflector on a held-out test set consisting of 1,000 ran- domly sampled instruction-prediction pairs from SSv2, SSP, and MMG. We maintain a balanced 1:1 ratio of success- ful and failed groundings. We assess its performance as a binary classifier S ∈ 0, 1 using three metrics: Accuracy, Precision, and Recall. Here, the positive class represents suc- cessful groundings S ∗ = 1. ModelAccuracy Precision Recall Qwen2.5-VL-3B (Zero-shot) 76.572.985.4 Trained Reflector π R 89.586.993.3 Qwen3-VL-2B (Zero-shot)80.578.986.0 Trained Reflector π R 91.788.895.7 Table 4: Performance metrics of the MLLM-based Reflector. As shown in Table 4, after training via GRPO with format and binary rewards, our Reflector achieves an evaluation Ac- curacy of 89.5% and 91.7%, demonstrating its capability in determining whether a predicted coordinate aligns with the user instruction. Computational Overhead and Training Time Analysis. To provide a comprehensive understanding of the computa- tional overhead, we conduct a training time analysis using the Qwen2.5-VL-3B model with 4 NVIDIA A100 40GB GPUs. As shown in Table 5, to compute relative advantages, GRPO samples a group of trajectories K = 8 for every sin- gle query, leading to the total training time of 449 minutes. R-OPSD operates on a different mechanism. By utilizing the Reflector’s feedback as privileged information to construct a self-teacher, R-OPSD performs a token-level knowledge distillation. The model only needs to sample a single tra- jectory per query. Even with the additional forward passes required for the self-teacher and the inverse-prompted stu- dent in Contrastive Calibration (C), the R-OPSD w/ C only takes 151 minutes, reducing the training time by more than 34% compared to GRPO. Our full framework (R-OPSD w/ C & QA) achieves the optimal performance by integrating the token-level advan- tage. Re-introducing QA necessitates returning to the K = 8 group rollout setting, which increases the training time to 778 minutes (1.73×). For resource-constrained deployment environments, R-OPSD w/ C serves as a highly efficient learning method that operates at 1/3 the cost of standard RL. The full framework leverages the combined strengths of both token-level and query-level advantages, offering a computational trade-off. Scalability to Larger Base Models. In the main text, our ex- periments are conducted on the 3B and 2B parameter scales, e.g., Qwen2.5-VL-3B and Qwen3-VL-2B. To further demon- strate the scalability of our Test-Time Self-Evolution frame- MethodTime (m) Relative Cost GRPO + GUI-RCPO4491.00× GRPO + Binary4471.00× R-OPSD (w/o C)129 ≈ 0.29× R-OPSD (w/ C)151 ≈ 0.34× R-OPSD (w/ C & QA)778 ≈ 1.73× Table 5: Training time Analysis. The Relative Cost is nor- malized against the standard GRPO baseline. Method Adapt on SSv2Adapt on MMG SSv2 SSP MMGSSv2 SSP MMG Qwen2.5-VL-7B86.8 19.9 68.286.8 19.9 68.2 + GUI-RCPO 89.2 25.9 70.388.3 25.5 70.9 + Ours92.731.477.992.532.379.2 Qwen3-VL-8B92.9 53.5 82.792.9 53.5 82.7 + GUI-RCPO93.2 53.8 83.393.5 54.6 83.4 + Ours94.555.884.394.156.384.7 Table 6: Scalability evaluation using Qwen2.5-VL-7B and Qwen3-VL-8B base models. We report the target dataset accuracy when adapting on SSv2 and MMG, respectively. ModelAccuracy Precision Recall Qwen2.5-VL-3B (Zero-shot)76.572.985.4 Trained Reflector π R 91.788.895.7 Qwen2.5-VL-3B + V (Zero-shot) 73.273.578.6 Trained Reflector + V π R 78.378.579.5 Table 7: Performance comparison of the Reflector with visual bounding box prompt and textual coordinate prompt. work, we extend our evaluation to a larger base model. The evaluation results are shown in Table 6. The Qwen2.5-VL- 7B and Qwen3-VL-8B base models exhibit stronger initial grounding capabilities than 3B/2B versions. However, even with this baseline, our framework consistently yields perfor- mance improvements. For instance, adapting the 7B model on MMG using our framework boosts the accuracy from 68.2% to 79.2%. Our method still outperforms the TTRL method GUI-RCPO. These results further demonstrate the scalability of our framework. Comparison with Visual Bounding Box Prompting. Many methods directly draw a bounding box or point on the UI screenshot for representing grounding coordinates (Zhang et al. 2026b; Luo et al. 2025b). To evaluate this design, we conduct an ablation study comparing textual coordinate representation against visual bounding box representation in two critical components of our framework: the Reflector’s evaluation and the teacher model’s privileged information prompt. For the Reflector + V and Ours + V, instead of feeding the textual coordinate token B in the prompt, we draw a red bounding box corresponding to the prediction on the input image I. Table 7 summarizes the evaluation results. Our trained Reflector with textual coordinates achieves a remarkable ac- Method Adapt on SSv2Adapt on MMG SSv2 SSP MMGSSv2 SSP MMG Qwen2.5-VL-3B80.4 20.3 57.580.4 20.3 57.5 + Ours88.830.566.687.530.168.2 + Ours + V77.318.353.277.517.454.7 Table 8: Performance comparison of our method with visual bounding box prompt and textual coordinate prompt. Method GUI-R1 Training Set SSv2 SSP MMG Qwen2.5-VL-3B80.4 20.3 57.5 + Ours87.829.267.2 Table 9: Performance of our method on the unsupervised training setting. curacy of 91.7%, outperforming the zero-shot base model by 15.2%. Conversely, employing visual prompts (+ V), the Trained Reflector + V only reaches 78.3% accuracy. The negative impact of visual prompting is amplified during the parameter internalization stage. As demonstrated in Table 8, leveraging drawn images (Ours + V) as the privileged vi- sual context for the self-teacher in R-OPSD leads to negative transfer. Extension beyond Test-Time Data. While our Test-Time Self-Evolving framework is designed to adapt models to unseen interfaces after deployment, its core internalization mechanism, R-OPSD, inherently serves as an unsupervised training algorithm. Instead of episodic exploration on test sets, we train the base model directly on the training set of GUI-R1 (Luo et al. 2025a), which contains about 3000 samples. To simulate the unsupervised training setting, we only leverage the image and instruction. The results are sum- marized in Table 9. It is shown that our method consistently outperforms the base model across all three datasets, demon- strating its effectiveness in unsupervised training setting. Qualitative Analysis of the Reflector To provide an intuitive understanding of the Reflector’s ca- pabilities, we present qualitative visualizations of its eval- uation process in Fig. 6. The visualization showcases the Reflector’s multi-modal reasoning across two typical test- time exploration scenarios: a successful grounding (S = 1) and a failed grounding (S = 0). As illustrated in the suc- cessful case, the Reflector accurately parses the user’s intent, spatially grounds the agent’s predicted coordinate to the cor- responding UI element, and logically confirms the match. More importantly, in the failed case, the Reflector does not merely output a binary rejection. Instead, it explicitly pin- points the exact cause of the failure within its step-by-step reflection R. Add a new app to the desktop <think> The instruction intends to add a new app to the desktop. The Click Coordinate [102, 59] is positioned near the top left corner of the screen, which typically corresponds to the '+' button used for adding new apps. This is consistent with the action described in the instruction. </think> <answer>Yes</answer> Open gallery <think> The instruction intends to open the 'Gallery' section. The UI shows a menu with various options, including 'Gallery'. The click coordinate [105, 129] corresponds to the 'Gallery' option in the menu. This matches the intended action of opening the Gallery section. </think> <answer>Yes</answer> (a) Successful Exploration (S = 1) Add a new shortcut to the toolbar <think> The instruction intends to add a new shortcut to the toolbar. The Click Coordinate [809, 14] is positioned near the top right corner of the screen, which typically corresponds to the toolbar area. However, there is no visible toolbar or shortcut icon at this exact location. Instead, there are icons for downloading, sharing, and other functions. Therefore, the click coordinate does not correctly execute the user's instruction. </think> <answer>No</answer> View attachment <think> The instruction intends to view an attachment, but the provided click coordinate [151, 465] is not aligned with any visible attachment icon or text. The closest element is a redacted text area, which is likely part of the email content rather than an attachment. Therefore, the click coordinate does not correctly execute the user's instruction to view an attachment. </think> <answer>No</answer> (b) Failed Exploration (S = 0) Figure 6: Qualitative visualizations of the Reflector’s evaluation and reasoning. The red bounding boxes denote the predicted coordinates. The yellow bounding boxes highlight the corresponding UI element.