{"id":"bebd255c-beca-4d44-bbce-782f1a6c6c04","arxiv_id":"2602.22190","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Action-weighted SFT plus KL-regularized GRPO on a curated 81K reasoning dataset lifts open-source GUI agents by 11–22 points on online task-completion benchmarks.","lead":"A new training recipe — action-aware fine-tuning plus KL-regularized reinforcement learning — and an 81K curated reasoning dataset substantially improves open-source GUI agents on web and mobile navigation benchmarks. Small models trained this way match far larger closed systems, though several evaluation choices and missing error bars mean the true effect size still needs replication.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RL grounding reward and SFT/RL filters treat Qwen3-VL-32B-inferred boxes as ground truth (Secs. 4.1.4, 5.5) with no external validation; loose/misaligned boxes would credit off-target coordinates and inflate the grounding-mitigation claim.","rationale":"The central claim is that a fully offline recipe—action-aware SFT plus KL-regularized GRPO with partial verification—makes small native GUI agents match much larger closed systems. The recipe's RL stage depends critically on the step-wise grounding reward, and both that reward and the SFT/RL data filters use Qwen3-VL-32B-generated bounding boxes as ground truth. The paper provides no external validation of these boxes, even though AndroidControl's accessibility trees would permit it. This is the weakest link in the training-signal chain, distinct from the benchmark-construction issue: even the external online gains could in principle be driven by a reward that rewards 'nearby' rather than 'correct' coordinates, which would undercut the mechanistic story about partially verifiable RL and grounding preservation. The reader's weakest assumption identifies this same issue, though it also lists the author-constructed v2 benchmarks; I focus narrowly on the bbox validation because it is more directly tied to the RL mechanism and is testable with existing data. A random-sample validation plus a retraining run with ground-truth boxes would settle whether the concern lands. If the boxes are accurate, the concern dissolves; if not, the RL training signal is an unverified proxy and the quantitative claims about RL-driven grounding gains need qualification. Since the reader already issued a CONDITIONAL verdict based largely on this family of concerns, my read does not change that verdict; it sharpens the required replication step.","tokens_in":58014,"tokens_out":5442,"duration_ms":58127,"concrete_test":"Sample 500 steps from the GUI-Libra RL training set (40K) that are AndroidControl-derived. Use the AndroidControl accessibility tree to obtain ground-truth element bounds (the same source the paper uses for UGround-style evaluation, Appendix C.2). Compute (i) the containment rate of the original point_2d in the Qwen3-VL-32B box, and (ii) the IoU between the Qwen3-VL-32B box and the accessibility-tree element box. Then retrain GUI-Libra-3B with the RL grounding reward recomputed from the accessibility-tree boxes (discarding steps with missing or badly misaligned boxes) and rerun AndroidWorld and WebArena-Lite-v2. If containment/IoU is low or the retrained gains shrink materially, the unvalidated bbox proxy is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec 4.1.4 keeps an SFT step only if the original point_2d falls inside a Qwen3-VL-32B-Instruct-predicted bounding box, and Sec 5.5 defines the RL grounding reward as r_g = 1{u ∈ b*}. These boxes are asserted to be 'reliable bounding-box annotations' but are never validated against accessibility-tree labels or human annotations. This is load-bearing because the RL stage's only spatial credit is that containment test: if b* tends to be loose or systematically shifted, the policy is rewarded for landing anywhere inside a generous envelope rather than on the intended element. The SFT filter also preferentially retains steps that the Qwen3 family can describe, an unmeasured selection effect. The later claim that RL 'fully mitigates grounding degradation' (Sec 6.3, Table 8) then rests on a reward whose geometry is unverified. Independent validation is immediately feasible: AndroidControl has accessibility trees, and the paper itself uses them for UGround-style evaluation (Appendix C.2), yet no such check is reported. If the boxes are accurate, the concern is minor; if not, the RL training signal is a proxy of unknown fidelity and the offline-to-online predictability analysis inherits that uncertainty.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GUI-Libra, a fully offline post-training recipe for native GUI agents. It constructs a curated 81K GUI reasoning dataset from public trajectory corpora by augmenting steps with GPT-4.1-generated chain-of-thought and filtering with Qwen3-VL-based agreement and bounding-box checks. Training consists of action-aware SFT (mixed reasoning-then-action and direct-action data, with token reweighting αa=2, αg=4) followed by KL-regularized GRPO with success-adaptive negative gradient scaling (SNGS). The authors give a formal analysis of partial verifiability (Theorem 5.1, Lemmas E.1–E.2) and report step-wise accuracy on AndroidControl-v2 and MM-Mind2Web-v2 plus task success on AndroidWorld, WebArena-Lite-v2, and Online-Mind2Web. The headline claims are consistent gains over base models and parity with much larger or closed systems, e.g., GUI-Libra-4B/8B reaching 42.6 on AndroidWorld, matching GPT-4o+UGround.","tokens_in":58323,"tokens_out":9086,"duration_ms":88136,"significance":"If the results hold, this is a significant and useful contribution: it demonstrates that careful offline data curation and conservative RL can make small open-weight native GUI agents competitive with much larger or proprietary systems, and it provides open dataset/code/model releases. The paper also gives a clear theoretical framing for why step-wise offline matching can fail to predict online success under partial verifiability, and why KL regularization should help. The proofs of Theorem 5.1 and Lemmas E.1–E.2 are correct under their stated assumptions. The empirical work is extensive: ablations of ASFT, KL coefficient, SNGS, data filtering, and grounding mixing (Tables 9–11, Figs. 9–11), plus decomposed grounding/action metrics in Appendix D.4, strengthen the paper beyond a single benchmark comparison.","major_comments":[{"comment":"Both the SFT coordinate filter and the RL grounding reward treat bounding boxes generated by Qwen3-VL-32B-Instruct as ground truth: a step is retained only if the original point_2d lies inside the predicted box, and r_g = 1{u ∈ b*}. These boxes are never validated against accessibility-tree labels, human annotations, or any independent source, despite the paper itself using accessibility trees for AndroidControl evaluation (Appendix C.2). If b* is loose or systematically shifted, the RL reward credits off-target clicks and the SFT filter preferentially retains steps that the Qwen3 family can describe. This is load-bearing because Section 6.3's claim that RL 'fully mitigates grounding degradation' rests on this reward geometry. Please report agreement statistics (e.g., precision/IoU against accessibility-tree bounds) or a sensitivity analysis with an alternative box source.","section":"Secs. 4.1.4, 5.5 (Eq. 12), 6.3"},{"comment":"The primary offline benchmarks are author-constructed: AndroidControl-v2 is filtered from 500 to 398 samples using Qwen3-VL-32B-Instruct, and MM-Mind2Web-v2 is produced by rewriting symbolic action histories with the same model. No external validation, inter-annotator agreement, or comparison against the original labels is provided. Because the filtering/rewriting model is from the same family as the Qwen3 backbones used for GUI-Libra-4B/8B, this can bias the offline comparison in favor of the proposed models. The headline offline gains (+9.5 to +20.9 Pass@1) are measured on these v2 benchmarks, so the construction needs independent validation or, failing that, results on the original unfiltered benchmarks.","section":"Secs. 6.1, C.2"},{"comment":"AndroidWorld (115 tasks) and Online-Mind2Web (300 tasks) are reported as single runs without standard errors, confidence intervals, or repeated evaluations. The abstract's quantitative claims (+15.6%, +12.2%, +4.0%, +8.7%) are point estimates. Given the high variance typical of live and interactive benchmarks—WebArena-Lite-v2 is averaged over four runs in Table 6—the paper should provide variance information or multiple seeds for at least the main online comparisons before claiming 'consistently improves.'","section":"Tables 5 and 7"},{"comment":"Section 6.3 states that 'our RL stage does not use direct grounding supervision' and attributes the closing of the reasoning/grounding gap to high-level step data. However, Eq. (12) includes r_g = 1{u ∈ b*}, which is explicit spatial supervision on every click action. This is not a separate grounding-only dataset, but it is still direct grounding reward. The claim as written is inaccurate and should be revised; the paper should also discuss how much of the Table 8 improvement is attributable to this reward component versus the KL/SNGS mechanisms.","section":"Secs. 5.5, 6.3"}],"minor_comments":[{"comment":"The main text says αa=2 and αg=4 by default, but Appendix B reports αa=αg=1 for GUI-Libra-4B. Please clarify which setting was used for each table entry.","section":"Sec. 6.1 vs Appendix B"},{"comment":"The GUI-Libra-4B/8B numbers in Table 11 do not match the corresponding rows in Table 3 (e.g., AC-v2 High: 59.8 vs 62.3 for 4B). If these rows use the no-SNGS variant, the caption should state this explicitly.","section":"Table 11"},{"comment":"The SNGS scaling factor has no lower bound in the displayed formula. If λ0 + κ p̂_g(s) becomes negative, negative advantages would be flipped in sign. Please state the clipping/implementation range (e.g., max(·, 0)).","section":"Eq. (10)"},{"comment":"The theoretical lemmas assume a per-state KL constraint, whereas GRPO applies an expected KL penalty. The paper already labels the bounds informal, but the main text should more clearly state that the formal connection to GRPO is heuristic rather than a direct implication.","section":"Sec. 5.3.2, Lemmas E.1–E.2"},{"comment":"The correlation analysis in Fig. 10 is computed over checkpoints from runs with different KL coefficients and model sizes. Please state the number of checkpoints and whether the correlations are pooled across models; this affects interpretation of the p-values.","section":"Figs. 9 and 10"},{"comment":"Minor typos: 'RL VR-tyle' in the abstract and 'even GPT5' in Section 6.2.1 should be corrected.","section":"Sec. 1 / Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be influential, and the open release of dataset, code, and models is a strong asset. The main risk is the unvalidated use of Qwen3-VL-generated bounding boxes as reward targets and filter criteria; this is fixable with a relatively small validation study against accessibility trees and a sensitivity analysis. The author-constructed v2 benchmarks and the absence of online error bars also need attention before acceptance. If the validation experiments are provided and the claims are adjusted, I would be willing to support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read.\n\nThe paper is GUI-Libra, and the short version is: small open GUI agents (3-8B) trained offline on ~81K curated steps and KL-regularized GRPO match or beat much larger closed systems on three online benchmarks. The 4B matching GPT-4o+UGround on AndroidWorld (42.6 vs 42.6) is the headline, and it holds up internally. This is the strongest data-efficient post-training result I've seen for native GUI agents.\n\nWhat's actually new: not any single ingredient—reasoning augmentation, token reweighting, and step-wise GRPO all have antecedents the paper cites. The contribution is the synthesis and the partial-verifiability analysis. Theorem 5.1 and the lemmas in Appendix E are correct under their stated assumptions, and they don't smuggle in the conclusion. The paper also releases data, code, and models, and the ablations (Tables 9-11, Figures 9-11) are extensive and internally consistent. The authors are honest about what they did and did not test.\n\nSoft spots, in proportion:\n\n1. The RL grounding reward and the SFT filter both trust bounding boxes predicted by Qwen3-VL-32B-Instruct (Secs. 4.1.4 and 5.5). These boxes are never validated against accessibility trees or human annotations, even though the paper uses accessibility trees elsewhere for evaluation. If those boxes are loose or systematically shifted, the RL stage rewards coordinates that may be off-target, and the claim that RL 'fully mitigates grounding degradation' is weaker than it looks. This is the main gap. It is not fatal—the online gains on AndroidWorld are large and the offline step-accuracy metrics are evaluated on accessibility-tree matching—but it does need an external check before I'd trust the exact magnitudes.\n\n2. No error bars on AndroidWorld or Online-Mind2Web. WebArena-Lite-v2 gets four runs, but the other two online results are single runs. Given 115 tasks and 300 tasks, the differences between several models are inside noise. This is a reporting issue more than a logic issue.\n\n3. The two offline 'v2' benchmarks are author-constructed with model-based filtering and rewriting. Selection effects are unmeasured. Since the headline claims rest on the external online benchmarks, this is a minor concern, but it should be flagged in the paper.\n\n4. The 'KL is critical' case rests on one pair of runs with a modest AndroidWorld spread (25.2 vs 21.7). The correlation analysis (Fig 10) is suggestive, not conclusive.\n\nThe math is sound, the data work is reproducible, and the paper is worth a serious referee. I'd send it to review as is, conditional on the authors addressing the bounding-box validation and reporting variance on the online runs. For the reading group, this is a yes; I'd cite it in my own work.\n\nRecommendation: engage with it. This is a legitimately good post-training paper, not a hype paper.","headline":"Solid, internally consistent post-training recipe for small GUI agents; the headline gains look real but unvalidated model-generated boxes and missing error bars keep me from signing off on the exact magnitudes.","tokens_in":58931,"tokens_out":2742,"would_cite":true,"duration_ms":26594,"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 fully offline post-training recipe — action-aware SFT plus KL-regularized, success-adaptive RL — makes small open native GUI agents match closed-source systems on long-horizon web and mobile tasks without online int","keywords":["GUI agents","vision-language models","reinforcement learning from verifiable rewards","partial verifiability","visual grounding","action-aware supervised fine-tuning","KL regularization","data curation"],"falsifier":"Take a random sample of GUI-Libra-81K and the RL subset, and for each step compare the Qwen3-VL-32B bounding box to a human label or accessibility-tree element boundary. If a substantial fraction of original point_2d coordinates fall inside the model box but outside the true element — or vice versa — the reward and filter are biased. A second decisive check: re-train GUI-Libra without the box-based filter/reward, using the original coordinates only; if performance does not drop, the claimed grounding supervision is not the source of the gains.","tokens_in":57830,"feed_emoji":"🤖","tokens_out":8959,"duration_ms":71170,"temperature":0.7,"pith_summary":"The paper is trying to establish that open-source GUI agents lag behind closed systems not because they lack data scale or environment interaction, but because generic post-training pipelines mishandle two GUI-specific problems: long chain-of-thought reasoning degrades visual grounding, and step-wise reinforcement learning is only partially verifiable (multiple actions can be correct, but offline data credits only one). To fix this, it proposes a three-part recipe: a curated 81K reasoning dataset with filtering; action-aware SFT that mixes reasoning and direct-action data and upweights action/grounding tokens; and conservative RL that keeps a KL trust region and downweights ambiguous negative gradients. The payoff: on AndroidWorld, WebArena-Lite-v2, and Online-Mind2Web, 4B–8B models consistently beat their base models and match or exceed much larger native and modular systems — all trained offline. A sympathetic reader would care because it suggests careful data curation and training-design choices can substitute for costly online rollouts and proprietary data.","feed_headline":"Offline-only training lets 4B GUI agents match GPT-4o","feed_subtitle":"A 4B model matches GPT-4o+UGround on AndroidWorld while improving 12–16 points over its base.","key_machinery":"Two mechanisms carry the argument. (1) Action-aware SFT (ASFT): a weighted cross-entropy objective over reasoning, action, and grounding tokens (default αa=2, αg=4) trained on a mixture of reasoning-then-action and direct-action samples, so long CoT traces no longer dilute the action/grounding signal. (2) The partially-verifiable reward analysis: Theorem 5.1 bounds online success J(π) ≥ 1 − H·C(π)·(1 − Moff(π) − η̄π), where C(π) is the occupancy mismatch between the policy's state distribution and the offline data distribution, and η̄π is the unmeasured mass the policy places on valid-but-undemonstrated actions. The theorem identifies these two quantities as the reason offline step-wise metr","core_discovery":"GUI-Libra claims that reasoning and grounding need not conflict in GUI agents if training is action-aware: standard SFT on long CoT traces degrades grounding, so it mixes reasoning-then-action and direct-action data and reweights tokens so action/grounding dominate. Its second claim is that step-wise RL is partially verifiable — multiple actions can be correct but only the demonstrated one is credited — and that this ambiguity, together with distribution shift, breaks offline-to-online predictability. A theorem (Thm 5.1) ties online success to occupancy mismatch and off-demo validity mass; KL regularization controls both. With moderate KL (β=0.001–0.005) and success-adaptive negative-gradien","pith_inferences":["A natural testable extension is to validate the Qwen3-VL-32B bounding boxes against human or accessibility-tree labels; if they are loose, the same recipe with human-verified boxes would likely improve further and reduce filter bias.","The partial verifiability bound is domain-agnostic: it suggests any offline-trained interactive policy (web navigation, desktop control) should use conservative KL-regularized updates and success-adaptive negative scaling to keep offline metrics meaningful.","The paper's finding that adding grounding data to RL improves grounding but degrades navigation hints at a competition between objectives; an extension would weight the two losses dynamically or add a navigation-specific reward to the grounding mix.","The two refined benchmarks (AndroidControl-v2, MM-Mind2Web-v2) are author-constructed; an independent human study of label quality would establish whether the reported gains reflect real improvement or partially the revised evaluation."],"forward_implications":["If the recipe is right, small open models (3B–8B) can substitute for much larger closed systems on long-horizon GUI tasks, lowering the barrier for building competent computer-use agents.","Data curation and filtering, not just raw scale, drive post-training gains: a filtered 81K-step dataset outperforms larger noisier corpora on both offline and online benchmarks.","KL regularization, which RLVR pipelines often discard, is essential for partially verifiable multi-step tasks; removing it causes policy collapse and weakens offline-to-online correlation (Pearson 0.63 vs 0.89 with KL).","Explicit reasoning traces become more valuable after RL: ablations show removing CoT at training time hurts online generalization far more than on offline metrics.","Offline step-wise metrics can be reliable predictors of online success, if the training process respects a KL trust region — useful for safely validating agents before deployment."],"fun_headline_variants":["Action-aware SFT fixes the reasoning-grounding tradeoff in GUI agents","4B GUI agent matches GPT-4o using offline-only action-aware training","KL trust region turns partially verifiable RL into reliable GUI training","Action-aware supervision unlocks offline GUI training on par with GPT-4o"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The training signal rests on bounding boxes predicted by Qwen3-VL-32B-Instruct being treated as ground truth for both reward and data filtering; if those boxes are loose or systematically misaligned, the RL reward credits off-target coordinates and the filtered data becomes biased.","fun_headline_variants_meta":{"raw":{"variants":["Action-aware SFT fixes the reasoning-grounding tradeoff in GUI agents","4B GUI agent matches GPT-4o using offline-only action-aware training","KL trust region turns partially verifiable RL into reliable GUI training","Action-aware supervision unlocks offline GUI training on par with GPT-4o"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001045,"raw_usage":{"total_tokens":4290,"prompt_tokens":862,"completion_tokens":3428,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":3350}},"tokens_in":606,"tokens_out":3428,"duration_ms":24635,"temperature":1.0,"reasoning_tokens":3350,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:45:57.948436+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of GUI-Libra-81K and the RL subset, and for each step compare the Qwen3-VL-32B bounding box to a human label or accessibility-tree element boundary. If a substantial fraction of original point_2d coordinates fall inside the model box but outside the true element — or vice versa — the reward and filter are biased. A second decisive check: re-train GUI-Libra without the box-based filter/reward, using the original coordinates only; if performance does not drop, the claimed grounding supervision is not the source of the gains.","supporting_citations":[],"review_version":1}