{"id":"512b9a1e-5d54-4d8b-b330-8dc29e6578cc","arxiv_id":"2507.05720","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MobileGUI-RL applies online RL with self-generated and filtered tasks plus trajectory-level rewards to mobile GUI agents, reporting improved success rates on AndroidWorld and AITW benchmarks.","lead":"MobileGUI-RL trains phone assistants by letting them practice on automatically generated tasks running in live Android emulators, rather than on pre-recorded examples. If the reported gains hold, this is a cheaper way to make agents that adapt to changing apps, but the paper lacks code and has a risky evaluation setup.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training AVD is named AndroidWorldAvd and self-exploration generates the curriculum from that environment, but the paper never shows the 436 training tasks are disjoint from AndroidWorld eval tasks; if they overlap, the AndroidWorld SOTA and 13.3-point gain are confounded.","rationale":"The reader's primary weakest assumption (unvalidated success oracle O and text-based world model W) is a reliability concern, but the final AndroidWorld numbers are measured with the official evaluator, so a biased oracle does not by itself invalidate the measured gain; it only makes the training signal suspect. The reader's secondary concern about AITW-Gen/AITW-Web manual curation affects two of the three benchmarks but does not touch the AndroidWorld result that anchors the strongest claim. The training/evaluation overlap concern is more direct: Appendix A.1.1 explicitly names the training AVD AndroidWorldAvd, and Section 3.3.1 generates training tasks from random walks in that environment. Because the paper provides no overlap analysis or exclusion statement, the headline AndroidWorld SOTA could be an artifact of training on the evaluation distribution. This does not overturn the reader's CONDITIONAL verdict, but it changes the reason for conditionality: the authors must demonstrate task/app disjointness (or rerun training on a held-out app set) before the central claim can be accepted.","tokens_in":13344,"tokens_out":11805,"duration_ms":141315,"concrete_test":"Extract the 436 training task instructions used for the final curriculum and the AndroidWorld evaluation task set (116 tasks from Rawles et al. 2024). Compute pairwise semantic overlap with an embedding model (e.g., text-embedding-3-large) plus manual review of the top matches, and diff the app/package lists of the training AVDs against AndroidWorld's AVD. If any AndroidWorld eval task is semantically matched by a training task or the app sets substantially overlap, the AndroidWorld results are contaminated; if the task sets are disjoint and app sets differ, the reported 44.8% stands and the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Appendix A.1.1 lists Base AVD Name: AndroidWorldAvd for the training emulator pool, and Section 3.3.1 builds the curriculum by having an exploration agent random-walk the GUI environment and then using GPT-4o to reverse-engineer task instructions from those trajectories. AndroidWorld evaluation uses the official evaluator on the same AndroidWorld emulator/app ecosystem. The paper never states that the 436 curated training tasks are disjoint from, or distributionally different from, the AndroidWorld benchmark's evaluation tasks. If self-exploration traversed AndroidWorld apps and the reverse-synthesized instructions resemble the benchmark's human-written task prompts, then the reported 13.3-point gain over Qwen2.5-VL-32B and the 44.8% vs 34.5% (GPT-4o) and 35.0% (Qwen2.5-VL-72B) comparisons are not evidence of a generalizable RL method; they are evidence of training on the evaluation distribution. This is a missing guarantee rather than a demonstrated failure, but it is the most load-bearing assumption in the paper because the headline claim collapses if the overlap is non-trivial.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces MobileGUI-RL, a framework for training mobile GUI agents via online reinforcement learning. It combines self-exploration to synthesize tasks from random-walk trajectories with GPT-4o reverse instruction generation, a text-based world model that filters tasks for feasibility, and an adapted GRPO objective (MobGRPO) with trajectory-level advantages and a composite reward balancing success, efficiency, and premature-termination penalty. The authors train Qwen2.5-VL-7B and 32B on 436 generated tasks and evaluate on AndroidWorld, AITW-Gen, and AITW-Web. They report substantial gains over base models, including 44.8% vs 35.0% for Qwen2.5-VL-72B on AndroidWorld, and claim state-of-the-art results among compared systems.","tokens_in":13575,"tokens_out":5923,"duration_ms":57869,"significance":"The claimed results are significant if substantiated: they would demonstrate that online RL with self-generated curricula and trajectory-level rewards can improve vision-language GUI agents beyond offline SFT and prior RL methods, and that a 32B model can surpass larger closed/open models on AndroidWorld. The paper is strong in laying out a complete pipeline with public benchmarks and ablations of filtering, curriculum, and reward shaping. However, the significance is conditional on resolving evaluation comparability and supervision-validity concerns; as written, the empirical evidence does not yet support the strongest claims.","major_comments":[{"comment":"The self-exploration training environment is based on AndroidWorldAvd, and the generated curriculum is built from random walks in that same emulator/app ecosystem; the paper never states that the 436 curated training tasks are disjoint from, or distributionally different from, the AndroidWorld evaluation tasks. If the reverse-engineered instructions (e.g., calendar, camera, map tasks) overlap with AndroidWorld's test tasks, the reported AndroidWorld SR of 44.8% and the 13.3-point gain over Qwen2.5-VL-32B are confounded by training on the evaluation distribution. Please provide an explicit task-overlap analysis, or change the training environment/apps to be disjoint from evaluation, or otherwise justify that the training distribution does not cover the evaluation tasks.","section":"Section 3.3.1 / Appendix A.1.1 / Table 2"},{"comment":"For AITW-Gen the evaluation set was manually reduced to 300 tasks and for AITW-Web to 150 tasks, but the paper does not state whether the baseline numbers in Table 2 (e.g., OS-Atlas 15.7/17.3, Aguvis 23.0/4.7, Qwen2.5-VL variants) were rerun on these exact curated subsets. If the baselines are taken from previously published full-set evaluations, the comparison is not apples-to-apples and the reported gains on these benchmarks may be inflated. Please rerun all baselines on the identical curated subsets or clearly state which numbers were obtained on which task sets.","section":"Appendix C / Table 2"},{"comment":"Training rewards are based on a single VLM oracle O (Qwen2.5-VL-72B) judging success from the final screenshots, with no reported human validation or agreement rate. Since Eq. (4) defines the RL reward on this binary judgment, any systematic bias in the oracle (e.g., accepting visually plausible but semantically wrong outcomes) directly undermines the learned policy and the ablation conclusions. Please report oracle accuracy against human labels on a sampled set of trajectories, and if possible compare oracle judgment with the official AndroidWorld evaluator.","section":"Section 3.2 / Eq. (4) / Table 3"},{"comment":"All results appear to come from a single training run; no multiple seeds, standard deviations, or significance tests are reported. Several comparisons rely on small margins (e.g., MobileGUI-7B 30.0% vs UI-TARS-7B 33.0% on AndroidWorld; MobileGUI-32B 30.7% vs Qwen2.5-VL-72B 31.3% on AITW-Web). Without variance estimates, the claim of 'consistent and significant gains' is not fully supported. Please add at least 3 seeds for the main results and the ablations in Table 3, or temper the strength of the claims.","section":"Section 4.2 / Table 2"},{"comment":"The text-based world model W is used to decide whether a generated task is admissible, but its fidelity to the real emulator is never validated. If W predicts transitions optimistically, the filter will admit infeasible tasks and the subsequent real-environment rollouts will be misaligned with the simulated feasibility signal. Please validate W on a sample by comparing its predicted next-state descriptions with actual emulator states after the same action, and report the correlation between W-based feasibility and real rollout success.","section":"Section 3.3.2"}],"minor_comments":[{"comment":"Section 1 says 'Experiments on four mobile agent benchmarks', but the abstract and Table 2 describe three benchmarks; please correct the inconsistency.","section":"Introduction"},{"comment":"Section 4.1.1 states a maximum episode length of 25 steps, while Appendix A.2.3 lists 'Maximum Steps per Episode: 15'; please reconcile these numbers.","section":"Section 4.1.1 / Appendix A.2.3"},{"comment":"The text refers to 'appendix X' for input-construction details, but no such appendix section is present; please add the section or remove the reference.","section":"Section 3.2"},{"comment":"Table 1 lists an 'answer' action, but the action-space description in Section 3.2 enumerates only four categories (gestures, text input, system navigation, control); please align the two descriptions.","section":"Table 1 / Section 3.2"},{"comment":"In Table 2, Aguvis-7B and Aguvis-72B are attributed to Huang et al. 2024, but that reference is a survey of LLM planning; the Aguvis paper is Xu et al. 2025. Please correct the citation.","section":"Table 2 / References"},{"comment":"Equation (1) uses the notation ∑G_{t=1}|oi,s| and nested sums over s and t in a way that is hard to parse; please define all indices or simplify the expression.","section":"Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a technical report with no code release and no explicit statement about the number of evaluation episodes per task. The evaluation issues identified in the main report are the key barrier: the potential training/evaluation overlap on AndroidWorld and the curated AITW subsets without rerun baselines would need to be resolved before the headline claims can be trusted. I do not see a novelty-disclosure problem; several related works from the same group are cited, but that is consistent with the topic. The paper fits the cs.LG scope, but the empirical base needs strengthening."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:2507.05720. The headline — 13.3 points over the 32B base on AndroidWorld, 44.8% SR beating GPT-4o's 34.5% — is exactly the kind of number that makes you want to inspect the eval. I did, and the paper is close to being a solid systems contribution, but there is one load-bearing assumption the authors never surface: the training AVD is literally named AndroidWorldAvd, and the curriculum comes from self-exploring that same emulator/app ecosystem that AndroidWorld evaluates on. The paper never states that the 436 curated training tasks are disjoint from, or distributionally different from, the AndroidWorld eval tasks. That is a missing guarantee, not a demonstrated failure, but it is the first thing a referee needs resolved. If there is non-trivial overlap, the 13.3-point gain is training-on-eval, not evidence of a generalizable RL method.\n\nWhat is actually new: the integration of self-exploration task synthesis, text-based world-model filtering, and trajectory-aware GRPO with a composite reward. Each ingredient is known (OS-Genesis, WebEvolver, DigiRL, DAPO), but the combination is cleanly executed. The ablations are the strongest part — the task-filtering and curriculum ablations show large, consistent effects, and the decaying-reward ablation supports the narrative. The method is plausible and the direction of gains across three benchmarks is coherent.\n\nWhere it gets soft: AITW-Gen and AITW-Web were curated to 300 and 150 tasks, and the paper does not state that any baseline was rerun on those exact subsets. The reported baseline numbers are presumably from prior papers using the full sets, which would inflate the gaps. No error bars or multiple seeds are reported. The oracle (Qwen2.5-VL-72B) and the text-based world model are used without human validation, and the reward hyperparameters (lambda, alpha, beta) are unreported. These are reproducibility gaps, all fixable in a revision. Minor issues: the intro says four benchmarks but the paper evaluates three, and the OS-Genesis reference is cited twice with two different years.\n\nBottom line: this deserves a serious referee, but not acceptance as-is. The method is worth engaging with; the missing contamination check is the make-or-break issue. If the authors can show the training tasks are not drawn from the AndroidWorld eval distribution — or if they can justify why overlap would not matter — the paper becomes a credible contribution. I would send it to review with a specific request for that clarification.","headline":"Promising online RL pipeline for mobile GUI agents, but the training AVD is named AndroidWorldAvd and the paper never shows the 436 training tasks are disjoint from AndroidWorld eval tasks — that missing guarantee has to be resolved before the headline numbers mean anything.","tokens_in":14159,"tokens_out":4060,"would_cite":false,"duration_ms":37077,"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":"MobileGUI-RL claims that online reinforcement learning, with self-explored task curricula and trajectory-level rewards, lifts mobile GUI agents by up to 13.3 points over their base checkpoints.","keywords":["mobile GUI agents","online reinforcement learning","GRPO","curriculum learning","vision-language models","AndroidWorld","task filtering","trajectory-level reward"],"falsifier":"Retrain MobileGUI-32B on the same 436-task curriculum and rollouts but replace the vision-language oracle's success judgments with human annotations; if AndroidWorld success no longer reaches roughly 44.8%, the reported gain depends on evaluator bias rather than genuine task completion.","tokens_in":13112,"feed_emoji":"📱","tokens_out":4585,"duration_ms":51166,"temperature":0.7,"pith_summary":"This paper tries to establish that vision-based mobile GUI agents can be trained effectively through reinforcement learning in an online emulator environment, rather than by fine-tuning on pre-collected offline trajectories. The authors argue that offline training overfits to specific interface templates and fails on unseen screens, while online interaction lets a policy adapt to live GUI dynamics. Their framework, MobileGUI-RL, generates tasks by reverse-engineering natural-language instructions from exploratory trajectories, filters those tasks through a text-based world model to keep only solvable ones, and then optimizes with a GRPO-style algorithm modified for trajectory-level rewards. The central empirical claim is that this yields large, consistent gains: MobileGUI-32B reaches 44.8% success on AndroidWorld, outperforming GPT-4o (34.5%) and the larger Qwen2.5-VL-72B (35.0%), while improving 13.3 points over its own 32B base model. If true, this points to a scalable path for training capable GUI agents without hand-annotated action trajectories.","feed_headline":"Online RL lifts a 32B GUI agent past GPT-4o on AndroidWorld","feed_subtitle":"Agent trained in live emulators with self-made task curricula gains 13.3 points over its base model.","key_machinery":"The load-bearing machinery is MobGRPO, an adaptation of group relative policy optimization (GRPO) for long-horizon GUI navigation. Instead of per-step rewards, it assigns a single normalized advantage to every action in a completed trajectory, computed from the trajectory's composite reward relative to other rollouts of the same task. The composite reward combines a binary success signal from a vision-language oracle, an exponentially decaying efficiency factor that rewards shorter successful trajectories, and a linear penalty that discourages premature termination; degenerate batches where all rollouts fail are filtered out following the DAPO recipe. A second key mechanism is the task-generation and filtering pipeline: a GPT-4o-based reverse task synthesizer converts random exploration trajectories into instructions, and a text-based world model W simulates whether a proxy agent can solve each candidate task, admitting only feasible tasks into the curriculum and ordering them by estimated complexity.","core_discovery":"The paper's central discovery is that online reinforcement learning can be made to work for mobile GUI agents if three design choices are combined: self-exploration for task discovery, a lightweight text-based world model for task filtering and curriculum construction, and a trajectory-level advantage with a composite reward that rewards success, penalizes premature termination, and distinguishes efficient from inefficient successful trajectories. The authors adapt GRPO to GUI navigation, calling their variant MobGRPO, and show through ablations that each component contributes: removing task filtering drops 32B AndroidWorld success from 44.8% to 41.0%, removing curriculum learning drops it to 34.0%, and removing the decaying efficiency reward drops it to 35.5%. The resulting agents improve over their base models on all three evaluated benchmarks, with the largest gains on the more diverse Android-in-the-Wild tasks.","pith_inferences":["A testable extension is to replace the proprietary LLM used for reverse task synthesis with the agent's own policy, which would remove the last external dependency and test whether the curriculum can be fully self-generated.","The text-based world-model filter may generalize to web or desktop GUI environments, but the paper only demonstrates it on Android emulators, so that transfer is an open question.","Because the paper does not report human validation of the oracle's success judgments, a direct audit of oracle accuracy on held-out trajectories would clarify how much of the measured gain is genuine task completion rather than evaluator bias."],"forward_implications":["If correct, a 32B open-weights model can surpass both its 72B sibling and a leading closed-source model on AndroidWorld without task-specific reward engineering or human action annotations.","The same recipe transfers to 7B models, where AndroidWorld success rises from 22.0% to 30.0% and AITW-Gen from 49.0% to 65.3%.","Removing the task filter, the curriculum, or the decaying reward each causes clear degradation, so all three components are load-bearing for the reported gains.","Training curves show steady online improvement, suggesting the framework could support continual adaptation as new tasks or UI layouts appear."],"supporting_citations":[{"why":"Supplies the AndroidWorld benchmark and the official evaluation code used to measure success rate.","marker":"Rawles et al., 2024"},{"why":"Provides the DigiRL offline-to-online RL paradigm, the VLM-based success evaluation idea, and the adapted AITW evaluation scripts.","marker":"Bai et al., 2024"},{"why":"Establishes reverse task synthesis, the method of generating task instructions from GUI trajectories that the self-exploration pipeline builds on.","marker":"Sun et al., 2025"},{"why":"Introduces GRPO, the group-relative policy optimization algorithm that MobGRPO adapts with trajectory-aware advantages.","marker":"Shao et al., 2024"},{"why":"Establishes the DeepSeek-R1 reinforcement-learning paradigm for LLM reasoning that motivates applying RL to agent policies.","marker":"Guo et al., 2025"},{"why":"Contributes the DAPO-style degenerate batch filtering used to avoid zero-gradient updates when all rollouts in a batch fail.","marker":"Yu et al., 2025"},{"why":"Shows multi-turn online RL with binary success signals for web agents, a baseline and motivation for the online RL approach.","marker":"Wei et al., 2025"},{"why":"Identifies GPT-4o, which is used both as a closed-source baseline for AndroidWorld and as the reverse task synthesizer for generating task descriptions.","marker":"Hurst et al., 2024"}],"fun_headline_variants":["Online RL lifts mobile GUI agent 13.3 points on AndroidWorld","Self-exploring GUI agent surpasses GPT-4o via online RL","MobGRPO: Online RL for GUI agents beats GPT-4o on Android","Online RL with self-curriculum boosts GUI agent by 13.3","GUI agent trained in live emulator outperforms GPT-4o"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The success signal and task-feasibility filter are trusted: a vision-language oracle decides whether a trajectory completed a task, and a text-based world model decides which tasks are learnable, with no human validation reported.","fun_headline_variants_meta":{"raw":{"variants":["Online RL lifts mobile GUI agent 13.3 points on AndroidWorld","Self-exploring GUI agent surpasses GPT-4o via online RL","MobGRPO: Online RL for GUI agents beats GPT-4o on Android","Online RL with self-curriculum boosts GUI agent by 13.3","GUI agent trained in live emulator outperforms GPT-4o"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1331,"prompt_tokens":878,"completion_tokens":453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":355}},"tokens_in":494,"tokens_out":453,"duration_ms":5245,"temperature":1.0,"reasoning_tokens":355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:19:27.289700+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain MobileGUI-32B on the same 436-task curriculum and rollouts but replace the vision-language oracle's success judgments with human annotations; if AndroidWorld success no longer reaches roughly 44.8%, the reported gain depends on evaluator bias rather than genuine task completion.","supporting_citations":[],"review_version":1}