{"id":"435e086b-68d5-4ddf-b8ac-aee98bad1d4a","arxiv_id":"2505.15810","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GUI-G1-3B shows that for GUI grounding, removing reasoning chains, regularizing box size in the reward, and reweighting GRPO by difficulty yields 90.3% on ScreenSpot and 37.1% on ScreenSpot-Pro.","lead":"This paper analyzes why reinforcement learning with chain-of-thought reasoning hurts GUI visual grounding, and introduces three fixes: a no-thinking prompt, a box-size reward constraint, and a modified GRPO objective. A 3B model trained on 17K samples reaches state-of-the-art accuracy on ScreenSpot and ScreenSpot-Pro.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The GRPO ablation runs on 900 samples with ~1-point gains, but the final 17K model changes template, reward, length normalization, difficulty weighting, and data filtering all at once; no full-pipeline ablation isolates the proposed recipe as the cause of the reported SOTA.","rationale":"Read in good faith, the paper is a useful, well-scoped empirical study with a clear decomposition of the R1-Zero-like pipeline, external benchmarks, and a plausible mechanism. I do not contest the numbers themselves, and I credit the authors for releasing code and reporting the analysis setup. The load-bearing concern is not that the numbers are wrong, but that the paper's central 'understanding' claim—that the proposed template, reward, and objective fixes are what produce the SOTA result—is not established by the experiments as reported. The only component-level ablation (Table 2) is small in scale and effect size; the final model bundles every change with a data filter, so even a fully correct set of numbers does not identify the cause. I checked the reader's weakest assumption (inverse box size as difficulty proxy): it is a real assumption, but it is secondary to the attribution problem, because the difficulty weight is only one of three proposed components and the final result could survive its failure. The more decisive missing experiment is a full-pipeline factorial ablation. This is consistent with the paper's own limitation B.2, which acknowledges that dataset composition and hyperparameters are not fully explored. Therefore I recommend keeping the conditional verdict; the paper needs the scale-level ablation before the causal recipe claim can be accepted.","tokens_in":15265,"tokens_out":11828,"duration_ms":110201,"concrete_test":"Run a 2^3 factorial ablation on the final 17K pipeline: fix the data and base model, toggle (a) Fast vs. Slow Thinking Template, (b) reward = RHit+RIoU+RBox vs. RHit+RIoU, and (c) GRPO objective = proposed vs. original with length normalization and no difficulty weight, evaluating each run on ScreenSpot and ScreenSpot-Pro with at least 3 seeds to obtain confidence intervals. If the simple main effects of the three proposed components are not positive and the cumulative recipe does not outperform the reverted baseline beyond noise, then the SOTA claim is not attributable to the proposed RL fixes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the proposed R1-Zero-like recipe—Fast Thinking Template, box-size reward, length-normalization removal, and difficulty weighting—is a SOTA training method. The weakest link is causal attribution. Table 2, the only ablation of the RL objective, is run on the 900-sample analysis setup (300 per domain, Appendix C.1) and reports 82.3 (Standard GRPO), 83.2 (length fix), and 83.3 (difficulty weight): a roughly one-point cumulative gain with no error bars or multiple seeds. The final GUI-G1-3B is trained on the full 17K set with at least five simultaneous changes relative to prior R1-style agents: (1) a no-thinking template versus the Slow Thinking Template, (2) RHit+alpha*RIoU+beta*RBox versus RHit+RIoU, (3) length normalization replaced by Max_Tokens, (4) a query-level difficulty weight wq, and (5) an 8-shot consistency filter that removes samples the base model answers all-correct or all-wrong. Any of these could shift ScreenSpot accuracy through output format and training distribution alone, independent of the RL objective. The paper's own Limitations (B.2) concede that dataset composition and hyperparameter tuning are not fully explored. Without a factorial ablation at the 17K scale, the headline numbers cannot be traced to the proposed fixes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies R1-Zero-style reinforcement learning for visual grounding in GUI agents, decomposing the pipeline into input template, reward design, and policy update. Through analyses on a 900-sample setup, it reports that longer chain-of-thought reasoning hurts grounding accuracy, that hit-based and IoU-based rewards induce opposite box-size reward hacking, and that GRPO exhibits length and difficulty biases. The authors propose a Fast Thinking Template, a box-size reward RBox added to RHit and RIoU, and a modified GRPO objective with no length normalization and a difficulty-dependent weight. They train GUI-G1-3B, initialized from Qwen2.5-VL-3B-Instruct, on about 17K public samples and report 90.3% on ScreenSpot and 37.1% on ScreenSpot-Pro, claiming state-of-the-art results for models of comparable size and an advantage over the larger UI-TARS-7B.","tokens_in":15516,"tokens_out":4216,"duration_ms":37531,"significance":"If the reported results hold, the paper offers a practical and data-efficient recipe for adapting R1-Zero-style RL to GUI grounding, with valuable negative evidence against reasoning templates for grounding tasks and a documented reward-hacking phenomenon. The use of public external benchmarks, a relatively small fully public training set, and a released repository are strengths that increase the reproducibility value of the work. However, the causal attribution of the final SOTA numbers to the proposed objective changes is not established by the current experimental design, and the difficulty-weight proxy is not validated for this base model; these issues prevent me from treating the recipe claim as fully supported at this stage.","major_comments":[{"comment":"The central attribution of the final gains to the proposed GRPO modifications is not established. Table 2, the only ablation of the RL objective, is run on the 900-sample analysis setup described in Appendix C.1 and reports cumulative gains of 82.3 (standard GRPO), 83.2 (Max_Tokens), and 83.3 (difficulty weighting); the final GUI-G1-3B is trained on the 17K set and changes at least five things simultaneously relative to prior R1-style agents: the no-thinking template, RHit+alpha*RIoU+beta*RBox, replacing |oi| with Max_Tokens, a query-level difficulty weight, and an 8-shot consistency filter. Any of these factors, especially the template and the data filter, could shift ScreenSpot accuracy through output format and training distribution alone. A factorial ablation at the 17K scale, or a deflated claim that the combined recipe is SOTA, is needed to connect the headline numbers to the proposed fixes.","section":"§3.3, §4, Table 2"},{"comment":"The difficulty weight wq is computed from the inverse relative box size, under the claim that smaller boxes are harder. This proxy is plausible for professional GUI datasets but is not validated on this base model or this training distribution; the cited support [17] is about ScreenSpot-Pro statistics, not about Qwen2.5-VL-3B-Instruct's error patterns. Since both RBox and wq depend on box size, the reweighting may simply reinforce output-size regularization rather than capture true per-sample difficulty. Please provide base-model accuracy as a function of lambda_q on the training data, or an ablation with randomized or shuffled wq, to demonstrate that the difficulty weighting is acting for the stated reason.","section":"§3.3, Appendix C.3"},{"comment":"The definition of RBox is inconsistent between Table 1 and the body text. Table 1 lists RBox = 4/(xp1+xp2+yp1+yp2), while the text defines xp1 = 1/(1 - |xhat1 - x1|/image_width), which suggests the reward is based on side-distance matching rather than simply matching box size. As written, the exact functional form and the gradient behavior of RBox are unclear, and this matters because RBox is one of the three core proposed contributions. Please give a single precise formula and state what quantity it is designed to match.","section":"§3.2, Table 1"},{"comment":"The state-of-the-art claims rest on single-run comparisons with numbers taken from prior papers. GUI-G1-3B outperforms InfiGUI-R1-3B by 2.8 points on ScreenSpot and by 1.4 points on ScreenSpot-Pro, margins that are within plausible run-to-run variation for RL-trained models, especially given that the main ablation in Table 2 also reports differences of about one point with no error bars. Please report variance across at least three training or evaluation seeds for the final model and, where possible, for the strongest baselines under the same evaluation protocol, or explicitly hedge the superiority claim.","section":"§4, Tables 4 and 5"}],"minor_comments":[{"comment":"The notation for the difficulty weight is inconsistent: Section 3.3 and Appendix C.3 use wq, while Table 2 and Section 3.4 use wp. Please unify the notation.","section":"§3.3, Eq. (2) and Table 2"},{"comment":"The legend entry \"Golden\" appears to be a typo for \"Ground Truth\".","section":"Figure 3"},{"comment":"The sum of the listed samples is 17,485, which is consistent with \"17K,\" but the paper does not report the number of samples removed by the eight-shot consistency filter; please clarify whether 17K refers to the retained or the original count.","section":"Appendix D.1, Table 6"},{"comment":"Section 2 presents the GRPO objective with KL-divergence regularization, while Section 4 states that no KL divergence regularization is applied in the final training; the analysis experiments in Appendix C.1 also use no KL by default. Please clarify whether the \"Standard GRPO\" baseline in Table 2 uses the same no-KL setting, so that the comparison isolates the objective changes rather than the KL term.","section":"§2 and §4"},{"comment":"For the aggregated rows of Claude Computer Use and Gemini 2.0, the paper reports only an average score; please indicate the evaluation protocol or source for these aggregated values, since the column structure otherwise provides per-domain breakdowns.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's core ideas are interesting and the benchmark numbers are strong, but the causal attribution issue is substantial: the main ablation is at 900 samples while the final model is a 17K-scale bundle of changes. This is fixable with additional experiments or a more careful claim, so I do not recommend rejection. I also note that the difficulty-weighting rationale is plausible but under-validated; if the authors can supply a simple validation plot or randomized-weight ablation, the paper would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What should you know: this is a well-scoped empirical paper on why R1-Zero-style RL needs adjustment for GUI grounding. The three-component decomposition (template, reward, GRPO objective) is genuinely useful, and each piece has its own analysis. The final GUI-G1-3B model is a new SOTA on ScreenSpot and ScreenSpot-Pro at its size, trained on 17K public samples with fewer output tokens. The paper is worth reading for anyone working on RL post-training for visual grounding.\n\nThe genuinely new part is the combination: no-thinking template (Fast Thinking), box-size regularizing reward, and a modified GRPO with length-normalization removed and difficulty weighting. The individual insights borrow from prior work on length bias and reasoning hurting multimodal tasks, but the paper's contribution is showing how they interact in GUI grounding and shipping a working recipe. The analyses of reward hacking—hit rewards shrinking boxes, IoU rewards enlarging them—are clear and well illustrated.\n\nThe soft spot is causal attribution. The ablations in Table 2 run on a 900-sample setup and show about a one-point cumulative gain from the RL objective changes, with no error bars or multiple seeds. The final model changes at least five things at once relative to prior R1-style agents: the template, the reward function, length normalization, difficulty weighting, and an 8-shot consistency filter on training data. So the SOTA numbers cannot be cleanly traced to the proposed recipe. The paper's own limitations section admits dataset composition and hyperparameter tuning are not fully explored. This is a legitimate concern, but not fatal—the recipe is plausible, and the final model's efficiency gains are real. The difficulty-weighting proxy (box size as difficulty) is a reasonable approximation, though not deeply validated. The circularity burden is low: evaluation uses external benchmarks and the weight is a training-time heuristic, not fitted to the test set.\n\nWho is this for: researchers working on GUI agents, RLHF/GRPO variants, and multimodal post-training. It deserves serious peer review and would benefit from a revision that adds factorial ablations at scale and error bars. I'd cite it as evidence that reasoning templates can hurt grounding, and as a strong baseline for efficient GUI grounding.","headline":"A solid empirical study of R1-Zero-style RL for GUI grounding, with a sensible recipe that likely works, but the headline SOTA claim outruns the controlled evidence.","tokens_in":16105,"tokens_out":1770,"would_cite":true,"duration_ms":15787,"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":"Drop the thinking: a 3B GUI grounding model trained without CoT tops 7B rivals.","keywords":["GUI grounding","reinforcement learning","GRPO","R1-Zero","reward hacking","chain-of-thought","ScreenSpot benchmark","visual grounding"],"falsifier":"A controlled comparison that keeps the Fast Thinking Template and the box-size reward identical but sets $w_q=1$ for every sample would directly test the difficulty-weighting claim; if ScreenSpot-Pro accuracy does not drop, the gain is not caused by the weighting. A complementary check is to run the method on a benchmark where tiny targets are not harder (for example, high-resolution icons with large clear labels), where the inverse-size proxy should lose its effectiveness if it is truly a difficulty signal.","tokens_in":15017,"feed_emoji":"🖱️","tokens_out":8435,"duration_ms":63879,"temperature":0.7,"pith_summary":"This paper argues that the R1-Zero recipe of reinforcement learning plus explicit chain-of-thought reasoning, successful for math and code, is actively counterproductive for GUI grounding. Across three pipeline components it shows: longer reasoning chains hurt grounding accuracy; hit- and IoU-based rewards push models to shrink or inflate predicted boxes; and the standard GRPO objective biases training toward easy examples. The paper's proposed fix combines a no-thinking Fast Thinking Template, a box-size regularization reward, and a difficulty-weighted GRPO objective. The resulting GUI-G1-3B, trained on 17K public samples from Qwen2.5-VL-3B-Instruct, reaches 90.3% on ScreenSpot and 37.1% on ScreenSpot-Pro, surpassing larger models such as UI-TARS-7B. If correct, the paper shows that adapting RL to the geometry of the task, rather than adding reasoning, is what unlocks strong visual grounding.","feed_headline":"Dropping chain-of-thought lifts GUI grounding to 90.3%","feed_subtitle":"A 17K-sample RL recipe with a no-thinking template and box-size rewards beats larger GUI agents.","key_machinery":"The machinery is a three-part training intervention. First, a Fast Thinking Template asks the model to output the bbox coordinates in JSON directly, with no <think> block. Second, the reward is $R_{\\text{Hit}} + 0.25 R_{\\text{IoU}} + 0.125 R_{\\text{Box}}$, where $R_{\\text{Box}} = \\frac{4}{x_{p1}+x_{p2}+y_{p1}+y_{p2}}$ with each term like $x_{p1}=1/(1-|\\hat{x}_1-x_1|/\\text{image width})$; this term penalizes box-size mismatch and counters the opposing hacking pressures. Third, the GRPO objective is multiplied by $w_q = 0.5 + \\frac{1/\\lambda_q - \\min_i(1/\\lambda_i)}{\\max_i(1/\\lambda_i)-\\min_i(1/\\lambda_i)}$, with $\\lambda_q$ the relative box size of the target normalized by image width plus height, and the length normalization $1/|o_i|$ is replaced by a constant. The difficulty weight is the mechanism that shifts gradient mass toward small-target, harder examples.","core_discovery":"The central claim is that the gains attributed to 'thinking' in R1-style GUI agents actually come from the online RL signal, and that explicit chain-of-thought templates reduce grounding quality because grounding depends more on visual tokens than on text tokens. The paper identifies two opposite failure modes in common rewards: hit-based rewards reward point accuracy and therefore shrink predicted boxes, while IoU-based rewards inflate boxes to raise overlap, so optimizing either alone produces reward hacking. It introduces a box-size reward $R_{\\text{Box}}$ that compares each predicted edge to the ground-truth edge scaled by image width, and a modified GRPO objective that replaces the per-response length $|o_i|$ with a constant $\\text{Max\\_Tokens}$ and multiplies the objective by a difficulty weight $w_q$ built from the inverse of the relative box size, so smaller targets receive larger gradients. The claimed result is that this combination, without any reasoning template and with one training epoch on 17K samples, sets a new state of the art on ScreenSpot and ScreenSpot-Pro for models of this size.","pith_inferences":["If the inverse-size difficulty proxy holds, the same weighting could be applied to other perception tasks with scale-imbalanced targets, such as small-object detection in natural images, where similar easy-example bias is observed.","The template comparison suggests that the 'aha moment' celebrated in R1-Zero visual reasoning may not transfer to grounding; a direct comparison on the same base model between a thinking-free and a thinking template would settle whether the template, rather than the data, drives the difference.","A testable extension is to replace the hand-scaled $w_q$ with a learned or adaptive difficulty weight, or to derive the weight from model uncertainty, which would decouple the difficulty proxy from box size and show whether the mechanism is difficulty weighting per se or merely upweighting small objects.","The reward combination suggests a general principle: when multiple metrics are complementary and competing, add an explicit constraint on the object being optimized (here box size) rather than relying on tuned linear combinations of the metrics alone."],"forward_implications":["GUI-G1-3B's numbers on ScreenSpot (90.3%) and ScreenSpot-Pro (37.1%) become the reference points for resource-efficient GUI grounding, beating InfiGUI-R1-3B and the larger UI-TARS-7B.","The inverse relation between reasoning length and grounding accuracy implies that GUI grounding should be treated as a fast, perception-driven task, so future R1-style GUI agents should not inherit thinking templates from general RL recipes.","Removing the length normalization from GRPO and weighting by difficulty should generalize beyond GUI grounding to other perception tasks where output length is not a proxy for effort.","The box-size reward provides a concrete recipe for preventing size-based reward hacking in any grounding RL setup, not just GUI agents.","The small training budget (17K samples, one epoch, no KL regularization) suggests that much of the capability is unlocked from pretrained knowledge via the RL signal rather than from new task-specific data."],"supporting_citations":[{"why":"Supplies the GRPO algorithm that is the base RL objective throughout the paper and the baseline 'Standard GRPO' in Table 2.","marker":"[33]"},{"why":"Provides the R1-style baseline (slow-thinking template, RHit+RIoU rewards) whose reasoning and reward dynamics are analyzed and whose ScreenSpot/ScreenSpot-Pro numbers are beaten.","marker":"[24]"},{"why":"Provides the ScreenSpot benchmark and the grounding-as-language-generation formulation, and is the main evaluation set for the training analyses.","marker":"[7]"},{"why":"Provides ScreenSpot-Pro, the harder high-resolution benchmark, and the relative-box-size difficulty proxy used to build w_q.","marker":"[17]"},{"why":"Provides the length-bias and difficulty-bias analysis of GRPO that motivates the two objective modifications (Max_Tokens and difficulty weighting).","marker":"[25]"},{"why":"Initiates the R1-Zero paradigm of rule-based RL without SFT that the whole paper examines and adapts.","marker":"[11]"},{"why":"Supplies the Web and Desktop training data (OS-Atlas) and is the strong SFT baseline compared in both benchmark tables.","marker":"[38]"},{"why":"Supplies the Mobile training domain (UI-BERT) used in the 17K mixed dataset.","marker":"[3]"},{"why":"Provides the base model Qwen2.5-VL-3B-Instruct used for training and evaluation, and the Qwen2.5-VL-7B baseline in the comparison tables.","marker":"[4]"}],"fun_headline_variants":["No CoT: RL hits 90.3% on GUI grounding","Drop thinking, add RL: GUI agents hit 90.3%","RL beats CoT for GUI grounding, 90.3%","GUI-G1: RL-only recipe nails 90.3% grounding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that smaller ground-truth boxes are genuinely harder grounding examples, so the inverse-size weight $w_q$ correctly identifies which samples should receive larger gradients during policy updates.","fun_headline_variants_meta":{"raw":{"variants":["No CoT: RL hits 90.3% on GUI grounding","Drop thinking, add RL: GUI agents hit 90.3%","RL beats CoT for GUI grounding, 90.3%","GUI-G1: RL-only recipe nails 90.3% grounding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1366,"prompt_tokens":1091,"completion_tokens":275,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":707,"tokens_out":275,"duration_ms":3149,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:11:25.134785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled comparison that keeps the Fast Thinking Template and the box-size reward identical but sets $w_q=1$ for every sample would directly test the difficulty-weighting claim; if ScreenSpot-Pro accuracy does not drop, the gain is not caused by the weighting. A complementary check is to run the method on a benchmark where tiny targets are not harder (for example, high-resolution icons with large clear labels), where the inverse-size proxy should lose its effectiveness if it is truly a difficulty signal.","supporting_citations":[{"cited_title":"Cheng, Q","cited_arxiv_id":null,"evidence_quote":"Provides the ScreenSpot benchmark and the grounding-as-language-generation formulation, and is the main evaluation set for the training analyses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Web and Desktop training data (OS-Atlas) and is the strong SFT baseline compared in both benchmark tables."}],"review_version":1}