{"id":"1d60dd23-d20b-45e3-b47d-3551b4f2bf08","arxiv_id":"2508.04389","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"An RL-based method, GuirlVG, achieves state-of-the-art GUI visual grounding with only 5.2K training samples, outperforming SFT trained on 10M samples.","lead":"A new reinforcement learning approach for GUI visual grounding claims to beat supervised fine-tuning with roughly 2,000 times less training data. The method introduces a stabilization technique called the Adversarial KL Factor and reports gains across three benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RL reward proxy may be overfit to ScreenSpot-style benchmarks; out-of-distribution GUI grounding not demonstrated.","rationale":"The reader's weakest_assumption names exactly the same load-bearing condition: the rule-based reward must faithfully capture correct GUI grounding. My concern adds specificity: the reward could be benchmark-aligned rather than generally aligned, and the only available evidence (the abstract) cannot rule this out. Since the paper text is unavailable beyond the abstract, no new evidence can be evaluated, so the appropriate verdict remains UNVERDICTED. No objection requires moving the verdict; the concern reinforces the need for full-text/code verification rather than changing the current verdict.","tokens_in":806,"tokens_out":1188,"duration_ms":15739,"concrete_test":"Obtain the full paper/code and extract the exact reward formulation (e.g., does it use ScreenSpot ground-truth boxes or string matching?). Then run the trained GuirlVG model on a held-out set of GUI screenshots from platforms/apps not represented in the RL training distribution, with human-annotated grounding labels. Compare accuracy against an SFT baseline and against in-distribution ScreenSpot numbers. If the relative improvement largely disappears or reward value increases while human-judged accuracy stays flat, the central claim is weakened. Additionally, report mean±std over at least 5 random seeds for the main comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that 5.2K RL samples beat 10M SFT samples on GUI visual grounding. For this to be true, the rule-based reward used in RFT must be a faithful proxy for real GUI grounding quality. The abstract does not describe the reward function, but the reported numbers are on ScreenSpot/ScreenSpotPro/ScreenSpotV2, which have known evaluation formats. If the reward is designed around these benchmarks (e.g., exact coordinate matching or text-based heuristics), RL can exploit dataset regularities—reward hacking—rather than learning general grounding. The proposed 'Adversarial KL Factor' mitigates reward over-optimization during training, but it cannot fix a fundamentally misaligned reward. A second, related concern is statistical: with only 5.2K training samples and no mention of runs, seeds, or significance tests, a 7.7% improvement could be noise. Neither issue can be resolved from the abstract; they are load-bearing because they determine whether the comparison is meaningful. The reader's weakest_assumption is correct: the reward-signal faithfulness assumption is central and unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GuirlVG, a reinforcement learning-based method for GUI visual grounding. The authors argue that supervised fine-tuning of MLLMs for GUI grounding requires large curated datasets, and that rule-based reinforcement fine-tuning (RFT) can be more efficient. They report a systematic decomposition of RFT components, a novel Adversarial KL Factor for training stabilization, and empirical results claiming that GuirlVG, trained on only 5.2K samples, outperforms SFT methods trained on over 10M samples, with improvements of 7.7% on ScreenSpot, 17.2% on ScreenSpotPro, and 91.9% accuracy on ScreenSpotV2.","tokens_in":1055,"tokens_out":1562,"duration_ms":18124,"significance":"If the reported results hold, the paper would make a strong case that carefully designed RL fine-tuning can replace massive supervised datasets for GUI visual grounding, a practically important capability for GUI agents. The claim of beating 10M-sample SFT with 5.2K samples is striking and would be of broad interest. The proposed Adversarial KL Factor, as a stabilization technique for RFT, could be a useful contribution if it is shown to be effective and general. However, the significance is currently conditional: the abstract does not provide enough detail on the reward design, evaluation protocol, or statistical robustness to assess whether the empirical claim is trustworthy.","major_comments":[{"comment":"The load-bearing assumption is that the rule-based reward used in RFT faithfully captures GUI grounding quality. The abstract does not describe the reward function. If the reward is designed around ScreenSpot-style benchmark formats (e.g., exact coordinate matching or text-based heuristics), RL could exploit dataset regularities rather than learn generalizable grounding. The Adversarial KL Factor mitigates reward over-optimization but cannot fix a fundamentally misaligned reward. The paper must specify the reward definition and provide evidence that optimizing it transfers beyond the evaluation distribution.","section":"Abstract (central claim)"},{"comment":"The abstract reports single-point accuracy improvements (7.7%, 17.2%, 91.9%) with no mention of variance, number of runs, seeds, or significance testing. Given the training set is only 5.2K samples, the observed gains could be within run-to-run noise. The authors should report mean and standard deviation over multiple seeds, and ideally a significance test, before claiming superiority over SFT baselines.","section":"Abstract (reported numbers)"},{"comment":"All reported results are on the ScreenSpot family (ScreenSpot, ScreenSpotPro, ScreenSpotV2). The abstract does not present out-of-distribution evaluations or ablation of the reward design. To support the claim that RL with 5.2K samples generalizes GUI grounding, the paper should include evaluations on held-out GUI domains or tasks not seen during training, and at least an analysis of reward hacking or failure modes.","section":"Abstract (experimental scope)"}],"minor_comments":[{"comment":"The 'systematic empirical study' and 'training configurations' are mentioned but not summarized in the abstract. A sentence describing which components were varied and which configuration was optimal would help readers assess the novelty.","section":"Abstract (method description)"},{"comment":"The term 'Adversarial KL Factor' is introduced without explanation. Since it is claimed as a novel contribution, the abstract should give a one-line intuition of the mechanism (e.g., how it dynamically adjusts KL penalty against reward over-optimization).","section":"Abstract (Adversarial KL Factor)"},{"comment":"The SFT methods trained on 10M+ samples are not named. Naming the baselines and clarifying whether they are the same MLLM backbone would strengthen the comparison's credibility.","section":"Abstract (baseline description)"}],"recommendation":"uncertain","confidential_remarks":"This review is based on the abstract only, as the full text was not available. The central empirical claim is plausible but rests on assumptions about reward design and statistical rigor that cannot be verified from the abstract. I would not reject on the basis of disagreement with consensus, but the paper needs a full review of the experimental setup before a verdict. Given the magnitude of the claimed improvement over 10M-sample SFT, I recommend the editor secure the full manuscript and a detailed assessment of the reward function and evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real candidate advance for the GUI-agent subfield, and the abstract is carefully written, but the central claim—5.2K RL samples beating 10M SFT samples—lives or dies on the faithfulness of the rule-based reward, which we cannot see. I'd send it to referees.\n\nWhat's actually new: applying rule-based reinforcement fine-tuning to GUI visual grounding, with a systematic component-wise decomposition of RFT and a proposed stabilization mechanism (Adversarial KL Factor). That's more than a 'first to apply' paper; the authors explicitly report that naive RFT underperforms SFT, which is a useful negative result they then dissect. The three benchmark numbers are consistent and specific.\n\nThe soft spot is the one the reader flagged and the stress-test sharpened: the reward signal. With only 5.2K samples and rewards defined at the level of abstract text, the reported gains could come from exploiting ScreenSpot-style evaluation regularities rather than learning general grounding. The Adversarial KL Factor stabilizes training against reward over-optimization, but it cannot align a misaligned reward. We have no description of the reward function, no ablation of the reward itself, no error bars, and no out-of-distribution test. These are not deal-breakers on their own—many solid RL papers provide all of them in the appendix—but they are exactly the things a referee must check.\n\nI also want to give credit: the abstract's claim that naive RFT underperforms is the kind of honest negative result that increases my priors on the rest of the paper. The numbers are stated precisely enough to be falsifiable. Nothing in the abstract suggests circularity or fitted constants.\n\nBottom line: this is a paper for people working on GUI agents and on RL fine-tuning of MLLMs. The idea is timely and the data-efficiency claim, if real, is practically valuable. It deserves a serious referee; I'd condition acceptance on seeing the full experimental protocol, the reward design, and ideally code or a reproducibility statement. I would bring it to a reading group if the full text was available.","headline":"Plausible, potentially useful RL recipe for GUI grounding, but the headline data-efficiency claim rests on an unverifiable reward-faithfulness assumption.","tokens_in":1502,"tokens_out":1492,"would_cite":true,"duration_ms":16150,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Reinforcement fine-tuning with only 5.2K samples beats supervised GUI grounding trained on 10M+ samples, the paper claims.","keywords":["GUI visual grounding","reinforcement learning","rule-based reward","multimodal large language model","reward over-optimization","KL divergence","ScreenSpot benchmark","supervised fine-tuning"],"falsifier":"Take the trained GuirlVG model and test it on a set of GUI screenshots deliberately designed to be adversarial to the reward rule (e.g., overlapping elements, ambiguous target descriptions, or screenshots from a different OS/theme). If accuracy on these screenshots falls sharply while the reward remains high, the model is over-optimizing the reward rather than learning genuine grounding.","tokens_in":734,"feed_emoji":"🎯","tokens_out":1525,"duration_ms":19383,"temperature":0.7,"pith_summary":"The paper argues that for graphical user interface visual grounding (GUI-VG), rule-based reinforcement fine-tuning (RFT) can replace massive supervised fine-tuning (SFT) datasets. The authors show that naive RFT underperforms SFT, then systematically decompose RFT into its components—reward design, optimization, and stabilization—to find a working recipe. Their method, GuirlVG, uses only 5.2K training samples yet outperforms SFT models trained on over 10M samples, with reported gains on three benchmark suites. The central message is that intelligent use of a cheap reward signal can substitute for expensive data curation in GUI agent training.","feed_headline":"5.2K samples beat 10M for GUI grounding","feed_subtitle":"Reinforcement fine-tuning with a rule-based reward outperforms massive supervised datasets on three benchmarks, the paper claims.","key_machinery":"Rule-based reinforcement fine-tuning (RFT) with a hand-designed reward for correct GUI element grounding, plus an Adversarial KL Factor—a dynamic regularizer that adjusts the KL penalty during training to keep the policy from hacking the reward. The machinery is the empirical decomposition of RFT into reward formulation, KL stabilization, and training configuration, which the paper uses to identify why naive RFT fails and how to fix it.","core_discovery":"GuirlVG is a reinforcement-learning recipe for GUI visual grounding built from an empirical decomposition of rule-based reinforcement fine-tuning. The paper's central claim is that, with a properly stabilized RFT setup, a model trained on just 5.2K samples achieves better GUI grounding accuracy than supervised fine-tuning on datasets orders of magnitude larger: a 7.7% improvement on ScreenSpot, a 17.2% improvement on ScreenSpotPro, and 91.9% accuracy on ScreenSpotV2. To make RFT work, the authors introduce an Adversarial KL Factor that dynamically counters reward over-optimization, which otherwise causes naive RFT to fall behind the SFT baseline.","pith_inferences":["The paper leaves open whether the 5.2K sample advantage transfers when the evaluation distribution shifts to unseen GUI styles or languages; a stress test on out-of-distribution screenshots would clarify whether the reward generalizes beyond its rule-based scope.","The Adversarial KL Factor might be interpreted as a form of early stopping that adapts over time; if so, a simpler schedule-based KL penalty may achieve similar gains, which is a testable simplification.","If the reward is the true driver, then the method's ceiling is set by the expressiveness of the rule-based reward; combining it with a learned reward model could push beyond the reported margins.","The paper's decomposition methodology could be reused to diagnose RFT failures in other dense-prediction or localisation tasks, not just GUI grounding."],"forward_implications":["If correct, GUI agent training no longer needs to collect and curate millions of labeled screenshots; a few thousand samples with a rule-based reward suffice for grounding.","The Adversarial KL Factor could be a general stabilizer for other rule-based RL fine-tuning tasks beyond GUI grounding, where reward over-optimization is a known failure mode.","The result suggests that increasingly capable pretrained MLLMs already encode GUI knowledge, and post-training should focus on aligning behavior via rewards rather than re-teaching grounding from data.","The reported benchmark numbers provide a concrete target: any new GUI-VG method should compare against 5.2K-sample RFT, not only against SFT baselines."],"supporting_citations":[],"fun_headline_variants":["RL beats supervised training for GUI grounding with 5.2K samples","GuirlVG: 5.2K samples outperform 10M for GUI grounding","Reinforcement fine-tuning outguns supervised fine-tuning in GUI grounding","Adversarial KL fix lets RL beat SFT for GUI grounding"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The rule-based reward used during reinforcement fine-tuning faithfully measures whether a model has correctly grounded a GUI element, so that improving the reward improves real-world grounding rather than gaming a proxy.","fun_headline_variants_meta":{"raw":{"variants":["RL beats supervised training for GUI grounding with 5.2K samples","GuirlVG: 5.2K samples outperform 10M for GUI grounding","Reinforcement fine-tuning outguns supervised fine-tuning in GUI grounding","Adversarial KL fix lets RL beat SFT for GUI grounding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1307,"prompt_tokens":798,"completion_tokens":509,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":542,"tokens_out":509,"duration_ms":6281,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:00:12.157815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained GuirlVG model and test it on a set of GUI screenshots deliberately designed to be adversarial to the reward rule (e.g., overlapping elements, ambiguous target descriptions, or screenshots from a different OS/theme). If accuracy on these screenshots falls sharply while the reward remains high, the model is over-optimizing the reward rather than learning genuine grounding.","supporting_citations":[],"review_version":1}