{"id":"2193d6a6-e563-4973-a64c-1afd0cc2e585","arxiv_id":"2608.11191","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A GUI grounding model can improve itself on unlabeled screenshots by reflecting on its own click predictions and distilling that reflection into its weights, gaining about 7.4% average accuracy across six benchmarks.","lead":"This paper introduces a self-evolving framework for GUI visual grounding: a click-predicting model improves itself after deployment by having a reflection model evaluate its predictions and distilling that feedback back into its weights. On six benchmarks the method improves average accuracy by about 7.4 points over the base model and beats a prior test-time reinforcement learning approach.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on the Reflector's reflection text being genuinely informative, but reflection quality is never measured; only binary evaluation accuracy is reported.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the framework is plausible but requires additional evidence. The reader's weakest assumption was broader (Reflector informativeness plus CC convergence). I focus on the narrower, more directly checkable premise: the reflection text R is unsupervised and unmeasured. This premise is load-bearing because the framework's novelty over test-time RL methods is precisely the use of textual reflection for dense token-level supervision. The proposed ablation isolates the content of R and would either confirm or refute the mechanism. I do not see an internal inconsistency that warrants REJECT; the empirical tables are internally consistent, though code and error bars are missing. Therefore the verdict remains CONDITIONAL, and my stress-test does not change the reader's verdict.","tokens_in":15484,"tokens_out":12865,"duration_ms":113569,"concrete_test":"Run the full adaptation (Qwen2.5-VL-3B, adapt on SSv2 and MMG) under four conditions: (1) current Reflector R; (2) R replaced by an empty string; (3) R replaced by a random reflection from a different sample in the same batch; (4) R replaced by an oracle description of the correct UI element. Keep S and all other hyperparameters fixed. If conditions (2)/(3) match or exceed (1) on the six benchmarks, the reflection content is not load-bearing and the mechanism claimed for R-OPSD is not supported. If (4) substantially exceeds (1), Reflector R quality is the bottleneck. Also report per-seed standard deviation for the headline +7.4% improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline 7.4% improvement is attributed to Reflection-Guided On-Policy Self-Distillation, which 'translates high-level reasoning into dense token-level supervision.' The method's sole stated assumption is that 'the Reflector provides sufficiently informative evaluations and reflections to improve the conditional distribution of the teacher.' However, the Reflector is trained with rewards (Eq. 2) that supervise only the binary label S and output format; the reflection text R is never supervised. Table 4 reports S accuracy (89.5/91.7), but no metric for R quality is given. The ablation showing that adding R helps (Table 2, rows d vs. b) only proves that including the reflection field changes the teacher prompt; it does not show that the content of R drives the gain. If R were generic or even random, the teacher prompt might still shift the output distribution enough to produce useful advantages, and the same empirical gains could appear without the reflection mechanism. Without a sensitivity analysis that degrades R or corrupts S, the central claim that the model learns from diagnostic reasoning is unsupported. The paper's assumption is asserted and illustrated only with selected qualitative examples.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a test-time self-evolving framework for GUI visual grounding. A frozen grounding policy performs rollouts on unlabeled screenshots; an MLLM-based Reflector evaluates each predicted coordinate and outputs a binary score S together with a textual reflection R; then Reflection-Guided On-Policy Self-Distillation (R-OPSD) converts S and R into token-level advantages by comparing a reflection-conditioned self-teacher with the base policy. A Contrastive Calibration method, direction-based advantage clamping, and an optional query-level GRPO term are added. Experiments on ScreenSpot-v2, ScreenSpot-Pro, OSWorld-G, OSWorld-G-Refine, and MMBench-GUI report an average improvement of 7.4% over the base model and favorable comparisons with GUI-RCPO. The appendix includes Reflector accuracy, scalability to 7B/8B models, computational overhead, and an unsupervised training setting.","tokens_in":15744,"tokens_out":13543,"duration_ms":119410,"significance":"If the reported results are robust, the paper would be a meaningful step toward post-deployment adaptation of GUI grounding models without human-labeled ground truth. The framework combines several established ideas in a new configuration: an MLLM Reflector, on-policy self-distillation, and a contrastive mechanism for avoiding corrupted prefix supervision. The experimental coverage is broad, with six benchmarks, two backbone families, component ablations, a separate Reflector reliability study, and an overhead analysis. The main technical elements are clearly described and the design is plausible. The credibility of the central claim, however, rests on two points that are currently under-supported: whether the reflection text R is genuinely informative, and whether the empirical gains are statistically reliable rather than artifacts of single runs and benchmark-level hyperparameter selection.","major_comments":[{"comment":"The central claim that the model learns from the Reflector's diagnostic reasoning is not directly supported, because the quality of the reflection R is never measured. Eq. (2) provides reward only for the binary label S and the output format; the reflection text receives no direct supervision or evaluation metric. Table 4 reports S accuracy, precision, and recall, but no metric for R. The ablation that adds or removes the reflection field (Table 2, rows (d)-(e) and (i)-(j)) changes the teacher prompt as a whole and therefore cannot separate the content of R from the mere presence of an additional text field. Please add a corruption or sensitivity analysis of R (for example, replacing R with templated or generic reasoning, truncating or shuffling the reflection, or randomly swapping reflections between samples) and report the resulting grounding accuracy. Alternatively, provide a direct measure of R informativeness on a held-out set with human annotations. Without this, the statement in the Methodology discussion that \"R-OPSD only assumes that the Reflector provides sufficiently informative evaluations and reflections\" remains a stated assumption, and the attribution of the 7.4% gain to reflection-guided supervision is not established.","section":"Evaluation and Reflection with Reflector; Eq. (2); Table 4"},{"comment":"The headline result is not statistically supported as reported. Every accuracy in Tables 1, 2, 3, and 6 is a single run without standard deviations, confidence intervals, or significance tests, and Fig. 4 reports averages without variance or sample size. Some of the claimed gains are small relative to typical benchmark noise; for example, Table 1 shows 61.2% vs. 60.8% on OSWorld-G-Refine for Qwen3-VL-2B adapted on SSv2. In addition, the integration strength lambda is selected using the evaluation benchmarks themselves in Fig. 5, and the CC and clamping components are ablated and selected on the same benchmarks in Table 2. Please report multi-seed intervals and significance tests, state the number of rollouts used in Fig. 4, and describe a hyperparameter-selection procedure that does not use target benchmark labels. These additions are necessary to support the average improvement claim as a robust effect rather than a tuned point estimate.","section":"Experiments; Tables 1-4; Fig. 4"},{"comment":"The evaluation protocol should be clarified with respect to data provenance. The paper says the model adapts on \"data from ScreenSpot-v2 or MMBench-GUI, strictly omitting their ground-truth annotations,\" and then reports accuracy on the same datasets. Please specify whether the unlabeled adaptation samples are drawn from the test splits of those benchmarks, whether the frozen baselines are evaluated under identical inference conditions, and how the adaptation procedure avoids using evaluation labels for early stopping, model selection, or lambda selection. The current description does not distinguish transductive test-time adaptation from unlabeled domain adaptation, and this distinction is relevant to the \"unseen interfaces\" claim in the abstract and to the fairness of the comparison with GUI-RCPO.","section":"Implementation Details; Table 1"}],"minor_comments":[{"comment":"The phrase \"This is the first work to applying OPSD in test-time scenarios\" should be corrected to \"the first work to apply OPSD.\"","section":"Introduction"},{"comment":"The formatting of Table 2 makes the exact configuration of each row difficult to parse, especially because collapsed rows (a) and (f) are marked with \"×\" in both the component columns and the result columns. Please separate the component settings from the numeric results or spell out the configuration of every row in the caption.","section":"Ablation Study; Table 2"},{"comment":"The statement that R-OPSD with CC \"only takes 151 minutes, reducing the training time by more than 34% compared to GRPO\" applies to the configuration without the query-level advantage; the full framework in row \"R-OPSD (w/ CC & QA)\" actually costs 1.73x of GRPO. This distinction should be stated in the main text so readers do not infer that the final method is cheaper than the GRPO baseline.","section":"Appendix; Table 5"},{"comment":"Figure 4 would be more informative if it reported the number of failed rollouts used for the averages and showed per-token distributions or error bars; as presented, the two bars are aggregate evidence for a mechanism that is claimed to operate at the level of individual drift steps.","section":"Fig. 4"},{"comment":"Because the teacher and the student share the same parameters, please state explicitly in the text that the log-probabilities in the advantage are evaluated under a frozen or stop-gradient snapshot of pi_G before the update, so that the objective is a distribution-matching loss rather than an unrolled gradient through the teacher.","section":"Discussion of R-OPSD, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the empirical effort is substantial, so I do not see grounds for rejection. The main risks are that the reflection mechanism is not directly verified and that the single-run, benchmark-tuned numbers may overstate the robustness of the improvement. Both are fixable with additional experiments and reporting. The novelty claim of being the \"first\" should also be softened in view of existing on-policy distillation work for GUI grounding (e.g., GUI-SD in Table 3), even if the reflection-based variant is new."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution, not a hype job. The framework is new: reflection-conditioned on-policy self-distillation plus contrastive calibration for test-time GUI grounding. The experiments are broad and the ablations are properly isolating. The weakest spot is that the content of the reflection R is never shown to matter; the paper reports only binary evaluation accuracy for the Reflector and gives qualitative examples, but no sensitivity test that degrades R or replaces it with generic text. That is a fixable gap, not a fatal one.\n\nWhat the paper does well. The problem is practical and the setup is honest: no ground-truth labels at test time, six benchmarks, two backbones, and a scaling experiment to 7B/8B. The CC mechanism is genuinely thoughtful—using an inverse-prompted student to suppress the initial wrong token and letting the advantage decay as the prefix drifts is a plausible remedy for a real failure mode in auto-regressive distillation. The ablation structure (with/without reflection, CC, clamping, query-level advantage) supports the headline claim, and the comparison to GRPO with scalar rewards and to GUI-SD's visual-prompt privileged info shows the reflection text is doing something beyond just providing any feedback.\n\nSoft spots. The main one, as the reader's stress-test flags, is that reflection quality is never measured directly. The paper states that R-OPSD only assumes the Reflector gives sufficiently informative evaluations and reflections, but it never verifies that the reflection text itself is informative. The ablation adding R shows a gain, but it does not distinguish useful diagnostics from extra conditioning text that shifts the teacher's distribution. A simple perturbation test—randomizing or paraphrasing R, or replacing it with a fixed generic sentence—would settle whether the content matters. Also, all results are single-run, no error bars; lambda = 0.2 is tuned on the evaluation benchmarks; and the promised code and Reflector dataset are not yet public. None of these are disqualifying, but they make the paper's numbers less robust than they look. The claim that CC decays advantage to zero is supported only by aggregate statistics in Fig. 4; a token-level case study would help.\n\nBottom line: this paper is for people building test-time adaptation for GUI agents or other structured auto-regressive outputs. It deserves serious peer review, and with the reflection-perturbation experiment, error bars, and code release it should pass. My recommendation: engage, conditional on those additions.","headline":"A genuine new combination for test-time GUI grounding, with broad experiments and a thoughtful contrastive calibration mechanism; the main gap is that the reflection text's content is never directly verified.","tokens_in":16212,"tokens_out":3120,"would_cite":true,"duration_ms":27095,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a GUI grounding model can keep improving after deployment, with no human labels, by reflecting on its own click predictions and distilling that reflection into dense per-token training feedback.","keywords":["GUI visual grounding","test-time adaptation","on-policy self-distillation","contrastive calibration","multimodal language model","reflection","self-evolving agents","GUI agents"],"falsifier":"On a held-out set of failed rollouts, measure the average absolute token-level advantage on the final drifted tokens: the paper reports it decays to 0.0033 with Contrastive Calibration, so a replication finding values above 0.1 would refute the claim that corrupted prefixes are filtered. A second check is to feed the Reflector's binary verdicts with shuffled text reflections; if gains persist, the reasoning text is not the load-bearing signal.","tokens_in":15306,"feed_emoji":"🖱️","tokens_out":6332,"duration_ms":49576,"temperature":0.7,"pith_summary":"The paper sets out to close the post-deployment adaptation gap for GUI visual grounding: models that localize UI elements from natural-language instructions typically freeze after training, so they never get better on new apps or layouts. The authors propose a self-evolving loop in which the model explores an unseen interface, an MLLM-based Reflector scores whether each predicted coordinate satisfies the instruction and writes a short reasoning about why, and a Reflection-Guided On-Policy Self-Distillation step converts that reasoning into per-token supervision. A Contrastive Calibration trick prevents the auto-regressive generation of wrong coordinate tokens from poisoning the training signal. Across six benchmarks the framework reports an average accuracy gain of 7.4 points over the frozen base model and up to 7.7 points over a recent test-time reinforcement learning baseline, suggesting that reflection-grounded self-distillation could let deployed GUI agents learn from their own experience.","feed_headline":"Deployed GUI agents improve 7.4% by reflecting on their own clicks","feed_subtitle":"Self-distillation turns the model's own commentary into per-token training signal, no ground truth needed.","key_machinery":"The central object is Reflection-Guided On-Policy Self-Distillation (R-OPSD) combined with Contrastive Calibration. R-OPSD builds a self-teacher by conditioning the same grounding model on the privileged information (the Reflector's binary score S and free-text reflection R), and defines the token-level advantage as the log-probability ratio between this teacher and the student's own policy on the student's generated coordinate tokens. Contrastive Calibration replaces the student in the denominator with an inverse-prompted student (which is told the failed prediction was correct), so that at the first incorrect token the teacher assigns lower probability to the wrong token than the deluded student, producing a negative advantage, while on drifted prefixes the two distributions converge and the advantage shrinks toward zero, preventing corrupted feedback.","core_discovery":"The central claim is that textual reflection on a failed or successful click, produced by a frozen multimodal Reflector, can be internalized into a GUI grounding model's weights through on-policy self-distillation, turning high-level natural-language reasoning into dense token-level advantages. The paper argues that the existing test-time RL baseline relies on sparse scalar rewards and therefore cannot explain or correct failures, whereas the reflection-conditioned self-teacher provides a superior gradient signal even on fully failed rollout groups that defeat relative-reward methods like GRPO. To keep that signal clean, the Contrastive Calibration method computes the advantage against an inverse-prompted student that treats the failed rollout as correct, yielding a negative penalty at the first wrong token and an advantage that decays to near zero once the prefix has drifted. The paper's evidence is the consistent gains across six benchmarks and two base model families, plus ablations showing that removing reflection, calibration, or direction clamping causes collapse or negative transfer.","pith_inferences":["The same Exploration-Evaluation-Reflection-Internalization loop could transfer to other coordinate-output tasks, such as robot manipulation or document layout, whenever a verifier can produce textual diagnostics.","The framework's ceiling is set by the Reflector's quality; a natural stress test is to measure how the grounding gains scale with Reflector accuracy, or to inject deliberately misleading reflections while keeping the binary score fixed.","Contrastive calibration is a general cure for drifted auto-regressive prefixes and could be applied to other on-policy distillation settings beyond GUI grounding, such as long-horizon agent trajectories.","The approach implicitly assumes the reflection text, not just the binary verdict, carries information; an ablation that shuffles or empties the reflection while keeping S would isolate where the gain actually comes from."],"forward_implications":["Deployed GUI agents can adapt to unseen applications and layouts without any human-annotated ground truth, learning from both successes and failures.","Failed explorations, which are typically discarded or merely penalized by sparse rewards, are converted into dense per-token training signal.","The framework consistently improves accuracy across six GUI benchmarks and on both 2B/3B and 7B/8B base models, and outperforms the test-time RL baseline GUI-RCPO.","The shared base model with alternating LoRA adapters keeps the memory footprint low (about 10 GB for 3B/2B models, 30 GB for 7B/8B models), making self-evolution feasible on a single GPU.","R-OPSD without the query-level GRPO advantage already runs at roughly one-third the training time of standard GRPO, so the core distillation loop is the cheap and effective part."],"supporting_citations":[{"why":"The test-time reinforcement learning baseline (GUI-RCPO) that the framework must beat; the comparison shows gains of up to 7.7 points.","marker":"Du et al. 2026"},{"why":"Provides the ScreenSpot and ScreenSpot-v2 benchmarks used for adaptation and evaluation.","marker":"Cheng et al. 2024"},{"why":"Provides the OSWorld-G and OSWorld-G-Refine benchmarks used for evaluation.","marker":"Xie et al. 2025"},{"why":"Provides the MMBench-GUI benchmark used for adaptation and evaluation.","marker":"Wang et al. 2025"},{"why":"Provides the ScreenSpot-Pro benchmark used Evaluation for evaluation.","marker":"Li et al. 2025"},{"why":"Supplies the GRPO algorithm used to train the Reflector and to form the query-level advantage.","marker":"Shao et al. 2024"},{"why":"The GroundCUA dataset from which the Reflector training data is constructed.","marker":"Feizi et al. 2025"},{"why":"GUI-SD, the on-policy self-distillation baseline for GUI grounding compared in Table 3.","marker":"Zhang et al. 2026b"},{"why":"The on-policy distillation formulation that R-OPSD builds on.","marker":"Agarwal et al. 2024"}],"fun_headline_variants":["GUI agents self-improve 7.4% at test time via reflection","Reflection-guided self-distillation boosts deployed GUI agents by 7.4%","No ground truth: GUI agents reflect and self-evolve, gaining 7.4%","Failed clicks teach GUI agents: reflection-guided adaptation yields 7.4% gain","Self-evolving GUI agents: reflection-distilled on-policy learning, +7.4% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole loop works only if the Reflector's binary verdict and free-text reasoning are informative enough to improve the teacher's conditional distribution; the paper also assumes, without proof, that on drifted prefixes the inverse-prompted student and the reflection-guided teacher converge so that corrupted advantages decay to zero.","fun_headline_variants_meta":{"raw":{"variants":["GUI agents self-improve 7.4% at test time via reflection","Reflection-guided self-distillation boosts deployed GUI agents by 7.4%","No ground truth: GUI agents reflect and self-evolve, gaining 7.4%","Failed clicks teach GUI agents: reflection-guided adaptation yields 7.4% gain","Self-evolving GUI agents: reflection-distilled on-policy learning, +7.4% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2810,"prompt_tokens":1002,"completion_tokens":1808,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1697}},"tokens_in":618,"tokens_out":1808,"duration_ms":10175,"temperature":1.0,"reasoning_tokens":1697,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:17:02.606700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of failed rollouts, measure the average absolute token-level advantage on the final drifted tokens: the paper reports it decays to 0.0033 with Contrastive Calibration, so a replication finding values above 0.1 would refute the claim that corrupted prefixes are filtered. A second check is to feed the Reflector's binary verdicts with shuffled text reflections; if gains persist, the reasoning text is not the load-bearing signal.","supporting_citations":[],"review_version":1}