{"id":"5486602b-050a-4fe7-9bb9-4b0384ba3602","arxiv_id":"2504.16073","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GuidNav improves GUI navigation by reranking top-k VLM action candidates with a learned process reward model, yielding about 3.4% higher static action accuracy and up to around 33% relative dynamic task success gains.","lead":"GuidNav uses a trained process reward model to score and pick among several candidate actions that a vision-language model proposes while navigating phone and web interfaces. It reports moderate accuracy gains on static benchmarks and larger task-completion gains on a dynamic Android test, though the dynamic setup is described only vaguely.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Dynamic gains depend on unvalidated transfer of a static ground-truth-matching reward model; Table 3 is not interpretable without evidence that PRM scores predict dynamic task success.","rationale":"The reader's weakest assumption is exactly the load-bearing point: the paper trains the process reward model on static ground-truth matching labels and then uses it in a dynamic environment without validating that this proxy predicts success. I agree with that framing. The static results are internally coherent: the AitW, GUI Odyssey, and Mind2Web absolute gains roughly average to the reported 3.4%, and the oracle top-k gap (Table 2) shows there is real signal in candidate reranking, so I do not object to the static claim. The dynamic claim, however, is the differentiator of the paper and it is supported only by a table whose environment is unspecified and whose mechanism of transfer is assumed. The concrete test I propose would settle whether the PRM actually discriminates between actions that lead to success in dynamic settings. Since the reader already recommended CONDITIONAL, I keep that verdict; the condition should be the release of dynamic rollouts and the success-prediction correlation/ablation evidence. Without that, the central dynamic claim remains unsupported.","tokens_in":13764,"tokens_out":6264,"duration_ms":63419,"concrete_test":"Release the dynamic rollouts from the 120 AitW tasks, or rerun them in the specified simulator. For every visited state in both successful and failed trajectories, take the top-k candidate actions, score each with the trained PRM, and label whether executing that action led to eventual task success. Compute the per-step AUC of PRM score against eventual success. If the AUC is not significantly above 0.5 (or, alternatively, if re-ranking by these dynamic-outcome labels does not reproduce the Table 3 gap between GuidNav and AR), the static-matching reward proxy has not transferred, and the central dynamic claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central dynamic claim (Table 3: GuidNav 58.1% vs AR 43.7%, a 33% relative gain) rests on using a process reward model trained entirely on a static proxy: Appendix A.1 labels a candidate action correct if it matches the ground-truth action type/coordinate/content, and the reward model is trained with those matching labels (Section 3.1, Eq. 3). Deployment in the dynamic AitW environment is then described only as a 'simulated setting' (Section 4.2), with no transition dynamics, action-execution mechanism, or success criterion specified. The paper never shows that an action which looks like a human demonstration in a static screenshot is the action that advances task completion in an evolving environment. If the dynamic state distribution differs from the static training distribution, the reward model may systematically prefer fluent-looking but ultimately suboptimal actions; the reported +14.4-point improvement over AR and +17.8 over DP would then be unexplained by the proposed mechanism. No confidence intervals, significance tests, inter-annotator agreement, code, or data are provided, so the dynamic result is currently unverifiable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes GuidNav, an inference-time guidance method for GUI navigation agents. A process reward model, trained on human demonstrations and self-play data, assigns scalar rewards to k candidate actions generated by a VLM policy; the highest-reward action is executed at each step. Potential uses include integrating the reward model with trajectory-level reflection and retry. The authors evaluate in static settings on AitW, GUI Odyssey, and Mind2Web, and report a dynamic evaluation on 120 AitW tasks, comparing against DP, TopK, AR, and DigiRL. They report a 3.4% average static single-step action accuracy improvement and an approximately 33% relative dynamic task-success gain over the AR baseline (Table 3: GuidNav 58.1% vs AR 43.7%), with further gains from trajectory refinement.","tokens_in":14015,"tokens_out":6824,"duration_ms":59785,"significance":"GuidNav is a lightweight alternative to fine-tuning or online RL: it leaves the policy VLM untouched and requires only a compact process reward model. The static results across three benchmarks are directionally consistent, and the efficiency analysis in Table 6 addresses practical API cost. If reliable, the dynamic results would be an interesting demonstration that a small supervised reward model can improve task completion without policy updates. However, the dynamic claim is the most novel and currently the least supported component: the reward model is trained on a static ground-truth matching proxy and deployed in an unverifiable simulated setting, and the paper does not release code, data, or checkpoints. The internal numerical inconsistencies further reduce confidence in the reported effect sizes.","major_comments":[{"comment":"The paper reports inconsistent headline dynamic gains. The Abstract states \"around 33% increase in task success rate\"; §5.1 states \"overall improvements of approximately 17.8% and 14.4%\" for GuidNav over DP and AR; and §7 concludes \"around 15% increase in task success rate in dynamic settings.\" Given Table 3 (GuidNav 58.1, AR 43.7), the relative gain over AR is 14.4/43.7 = 32.9%, so the 15% figure in §7 is not explained. The static gain is also presented as 3.4% in the Abstract but \"near 5%\" in §7 without clarifying that the latter refers only to AitW. These headline numbers must be reconciled.","section":"Abstract vs. §5.1 and §7"},{"comment":"Several reported row averages are not the arithmetic means of the per-subset values. In Table 2, DP should be (30.3+39.1+36.2+34.4)/4 = 35.0, not 34.3; GuidNav should be (35.5+41.4+40.9+38.5)/4 = 39.1, not 38.9; and Pass@N should be (43.4+48.4+48.8+42.3)/4 = 45.7, not 46.8. In Table 3, DP should be 38.9, not 40.3; AR should be 41.9, not 43.7; GuidNav should be 56.4, not 58.1; and Integration should be 69.0, not 71.6. These discrepancies, up to 2.6 points, are comparable to or larger than several reported method-versus-baseline gaps, so the central comparative claims need to be recomputed and verified.","section":"Tables 2 and 3"},{"comment":"The dynamic evaluation is described only as \"a simulated setting\" with 120 sampled AitW tasks. The manuscript does not specify the simulator, the transition dynamics, how actions are executed, the task-success criterion used by the two human annotators, or inter-annotator agreement, and it provides no confidence intervals or significance tests. As a result, Table 3 is not interpretable as a reproducible benchmark result and the headline dynamic gain is currently unverifiable. The paper should report the full evaluation protocol and, ideally, release the code and simulator configuration.","section":"§4.2 and Table 3, dynamic evaluation"},{"comment":"The process reward model is trained on static screenshots with labels derived from ground-truth-matching rules (Appendix A.1), then deployed in a dynamic environment where no ground truth exists and the correct action depends on the evolving state. The paper provides no evidence that PRM scores transfer to human-judged task success in the dynamic setting. A model trained on a static visual-matching proxy may systematically prefer actions that look like demonstrations but do not advance the task under changing environment dynamics. Please provide a direct validation, such as the correlation between PRM scores and eventual task completion on held-out dynamic episodes, or an offline analysis of whether the top-scoring actions would be allowed by the dynamic environment.","section":"§3.1, Eq. (3), Appendix A.1, §4.2"}],"minor_comments":[{"comment":"The phrase \"a around 33% increase\" should be corrected to \"an around 33% increase\" or reworded.","section":"Abstract"},{"comment":"The legend entry \"TopK+AP\" should read \"TopK+AR\" to match the text and the other abbreviations.","section":"Figure 2"},{"comment":"The case study refers to \"Figure ??\", but the flight-search figure appears to be missing or its number is unresolved.","section":"§6.3"},{"comment":"The phrase \"For other actions: For other actions:\" is duplicated and should be condensed.","section":"Appendix A.1"},{"comment":"The denominator is typeset as \"1PN...\" in a garbled way, and the notation switches between r^(i)_t,true and r^(i)_t,anno for the same quantity; please unify the notation and fix the equation formatting.","section":"Eq. (3)"},{"comment":"The DigiRL row contains dashes for Google_apps and Install without explanation, and the statement that GuidNav \"maintains superiority\" over DigiRL is only partially supported because DigiRL is not evaluated on those subsets.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is useful and the static results are directionally consistent, but the systematic arithmetic errors in the main tables and the unverifiable dynamic evaluation are serious. I would request a full re-run of all aggregations, a complete specification of the dynamic simulator and success criterion, and additional statistical support (CIs or significance tests) before this paper can be accepted. If code and data cannot be released, the dynamic evaluation protocol must at least be described in sufficient detail to be replicable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll give you the short version. The paper's real contribution is static: a process reward model trained on ground-truth matching labels that reranks top-k candidate actions from a black-box VLM. Across AitW, GUI Odyssey, and Mind2Web, and across three policy models, this gives a consistent 2–5 point accuracy improvement. That is a useful, plausible result for applied GUI agents, and the oracle upper-bound control suggests the reward model is genuinely picking better actions, not just gaming the metric.\n\nThe dynamic claim is where it falls apart. The reward model is trained on static screenshots where 'correct' means matching a human demonstration. It is then deployed in a 'simulated setting' in AitW, but the paper never specifies the transition dynamics, action-execution mechanism, or success criteria, and never validates that an action that matches ground truth in a static screenshot is the action that advances task completion in an evolving interface. Without that, Table 3's double-digit gains are unexplained. The numbers are also inconsistent: the abstract says 33% dynamic gain, the conclusion says 15%, and the averages in Tables 2 and 3 don't match the per-subset values. These are load-bearing bookkeeping errors, not typos. The limitations section is generic and doesn't mention this transfer issue.\n\nThere is no code, no data, no confidence intervals, no significance tests. For a paper whose headline is about success-rate improvements, that's a low bar for verification. The TopK baseline is weak too—it just returns the first of k candidates, so the comparison is partly a prompt effect. The citation pattern is fine; the self-citation for TopK is legitimate.\n\nThe static result is probably real and worth refereeing. The dynamic result, as presented, is not. This paper deserves a serious referee, but it needs major revision first: either drop the dynamic claim or properly validate the transfer, fix the arithmetic, and release code and data. I'd want the authors to explain the average mismatches before I'd trust any of the headline numbers.","headline":"Useful static reranking result, but the dynamic gains are unverified and the paper's own numbers don't add up.","tokens_in":14545,"tokens_out":7007,"would_cite":false,"duration_ms":56322,"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":"A process reward model that scores each candidate action at every GUI navigation step improves static action accuracy by 3.4% and raises dynamic task success by about a third.","keywords":["process reward model","GUI navigation","inference-time guidance","action reranking","visual language model agent","Android-in-the-Wild","GUI Odyssey","Mind2Web"],"falsifier":"Run the reward model on states collected from dynamic rollouts in the simulated Android environment: for each step, record the reward it assigns to the action that was actually executed, then compare with whether the trajectory eventually succeeds. If step-level reward accuracy on dynamic states is near chance, the static-trained reward model's transfer to dynamic environments is refuted, removing the basis for the headline 33% gain.","tokens_in":13595,"feed_emoji":"📱","tokens_out":6879,"duration_ms":56388,"temperature":0.7,"pith_summary":"The paper proposes GuidNav, an inference-time method that uses a process reward model to choose which action a visual-language-model agent should take at each GUI navigation step. Instead of fine-tuning the acting model or waiting until the end of a trajectory to judge success, a separately trained reward model scores the agent's top candidate actions on the current screen, and the highest-scoring one is executed. The paper reports that this step-level reranking improves single-step action accuracy by an average of 3.4% across Android-in-the-Wild, GUI Odyssey, and Mind2Web static settings, and raises task success in a simulated dynamic Android environment from 43.7% to 58.1%, a roughly 33% relative gain over trajectory-level autonomous refinement. When reward-guided step selection is combined with trajectory reflection and retry, success reaches 71.6% after up to three retries. The interest of the claim is that it promises better GUI agents with black-box commercial policy models, without fine-tuning.","feed_headline":"A learned reranker for GUI agents lifts task success 33%","feed_subtitle":"A step-level reward model picks the best action per screen, improving static accuracy by 3.4% on three benchmarks.","key_machinery":"The load-bearing object is the process reward model R(x, h_t, s_t, a_t), which maps the instruction, a language summary of the navigation history, the current screen, and a candidate action to a scalar reward. It is trained by mean squared error against step-level annotations, where the reward for an action is 1 if it matches ground truth and 0 otherwise under the dataset-specific matching rules. At inference it selects a* = argmax R(...) among k actions proposed by the policy VLM. The reward model's role is to reclaim the correct action that the policy already knows but does not rank first; the paper's oracle upper bound shows that gain is available in the candidate set.","core_discovery":"The central discovery claimed is that a process reward model trained on ground-truth-matched actions can serve as a reranker at inference time. Given a user instruction, a summary of prior steps, and the current screenshot, the policy VLM proposes k candidate next actions; the reward model assigns a scalar to each candidate in context, and the agent executes the argmax. Trained on human demonstrations plus VLM self-play trajectories, with rewards derived from matching the action type and target to ground truth (including a distance tolerance for clicks and direction matching for scrolls), the reward model achieves 78.8% step-wise accuracy on the AitW evaluation. The paper argues that this selection mechanism is what yields the gains: simply generating top-k actions does not help, but selecting among them with a learned reward does.","pith_inferences":["A testable extension the authors do not run is to separate two possible sources of gain: candidate quality (whether the correct action is in the top-k set) and reranking quality (whether the reward model finds it). Reporting both Pass@K and reward-selection accuracy on the same states would show how much of the 3.4% is due to each.","The dynamic result rests on an unvalidated transfer: the reward model is trained on static ground-truth matching, yet deployed where correctness depends on state evolution. A direct check would score the reward model on states encountered in dynamic rollouts and compare its rankings with whether the trajectory ultimately succeeds.","If the transfer holds, the same reranking recipe could be applied to any action space with cheap automatic labels, such as web navigation or desktop control, where the paper's own limitations section notes broader generalization is untested."],"forward_implications":["In static GUI benchmarks, reranking the policy's top-k actions with a learned process reward improves mean single-step action accuracy by 3.4% across Android-in-the-Wild, GUI Odyssey, and Mind2Web.","In a simulated dynamic Android setting, the same reward-guided selection raises task success to 58.1%, roughly 33% higher than the 43.7% trajectory-level autonomous-refinement baseline.","Combining process-reward selection with trajectory reflection and retry raises success further, to 71.6%, showing the two levels of supervision are complementary.","Because the acting VLM is only prompted, the recipe works with commercial black-box policy models and avoids expensive fine-tuning or online reinforcement learning.","The gap between reward-model selection and the oracle top-k upper bound indicates headroom: if candidate generation and reward scoring both improve, the ceiling is higher."],"supporting_citations":[{"why":"It supplies the Android-in-the-Wild dataset and the static action-accuracy and ground-truth matching protocol used for AitW evaluation.","marker":"Rawles et al., 2024b"},{"why":"It supplies the GUI Odyssey cross-app benchmark and its evaluation metrics.","marker":"Lu et al., 2024"},{"why":"It supplies the Mind2Web web-navigation benchmark and the element-accuracy and step-success metrics used for evaluation.","marker":"Deng et al., 2023"},{"why":"It defines the Autonomous Refinement baseline that GuidNav is compared against and integrated with for trajectory reflection and retry.","marker":"Pan et al., 2024"},{"why":"It is the reinforcement-learning baseline; GuidNav claims superior dynamic success without online RL training.","marker":"Bai et al., 2024"},{"why":"Reflexion is the reflection framework underlying the AR baseline and the retry mechanism.","marker":"Shinn et al., 2024"},{"why":"Set-of-Mark prompting is used to number screen elements so actions refer to labelled regions instead of needing precise coordinates.","marker":"Yang et al., 2023b"},{"why":"CogVLM2 is the open-source VLM that is fine-tuned into the process reward model in the experimental setup.","marker":"Hong et al., 2024a"}],"fun_headline_variants":["Process rewards rerank GUI actions: +3.4% static, +33% task success","Inference-time reward model steers VLM agents to +33% success","Reranking with process rewards boosts GUI agent success by 33%","Step-level reward model lifts GUI agent accuracy and success","GUI agents guided by process rewards gain 33% more task success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reward model is trained on static screenshots with labels derived from matching the predicted action to ground truth, and this matching proxy is assumed to transfer to dynamic environments where ground truth is absent and correctness depends on the evolving state.","fun_headline_variants_meta":{"raw":{"variants":["Process rewards rerank GUI actions: +3.4% static, +33% task success","Inference-time reward model steers VLM agents to +33% success","Reranking with process rewards boosts GUI agent success by 33%","Step-level reward model lifts GUI agent accuracy and success","GUI agents guided by process rewards gain 33% more task success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000523,"raw_usage":{"total_tokens":2502,"prompt_tokens":895,"completion_tokens":1607,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":1510}},"tokens_in":511,"tokens_out":1607,"duration_ms":10839,"temperature":1.0,"reasoning_tokens":1510,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:11:09.388702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the reward model on states collected from dynamic rollouts in the simulated Android environment: for each step, record the reward it assigns to the action that was actually executed, then compare with whether the trajectory eventually succeeds. If step-level reward accuracy on dynamic states is near chance, the static-trained reward model's transfer to dynamic environments is refuted, removing the basis for the headline 33% gain.","supporting_citations":[],"review_version":1}