{"id":"3eba9df0-1e27-4504-bedb-cb6a77000fb1","arxiv_id":"2606.20032","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"ReA-OVCD is a training-free reliability-aware method for open-vocabulary change detection that uses semantic change reasoning and boundary-aware refinement to reduce artifacts and improve accuracy on remote sensing datasets.","lead":"This paper introduces a training-free framework called ReA-OVCD for open-vocabulary change detection in remote sensing that first identifies candidate regions from pixel discrepancies and then refines them with semantic and spatial checks. A smart generalist might read it to see how AI can handle flexible text-based queries for land monitoring without needing task-specific training data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No quantitative validation that SCR/BCR reduce net errors rather than trade true positives for fewer false positives","rationale":"The identified weakness is identical to the reader's weakest_assumption and remains load-bearing even after full-text access, because overall dataset F₁ scores alone cannot isolate whether the claimed reliability mechanism is functioning as asserted.","tokens_in":1818,"tokens_out":294,"duration_ms":16722,"concrete_test":"On the LEVIR-CD test split, compute pixel-wise change masks before SCR, after SCR, and after BCR; report precision, recall, and F₁ at each stage against the ground-truth change map. If recall drops by >3 % while precision rises by <5 %, the modules are not demonstrably reliability-enhancing.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central pipeline derives candidate regions from pixel-wise semantic discrepancies, then applies SCR (distributional divergence + response variation) and BCR (interior-pixel support check) to enforce reliability. The headline performance claim (F₁^C gains of 2.13–9.75 %) is only credible if these modules improve precision without systematically discarding genuine changes or retaining boundary artifacts. The manuscript supplies no per-module ablation, no stage-wise precision/recall tables, and no error-rate analysis on the candidate set, so it is impossible to verify that the refinement step is net-positive rather than merely re-labeling the same errors.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes ReA-OVCD, a training-free framework for open-vocabulary change detection in remote sensing imagery. Candidate change regions are first identified via pixel-wise semantic discrepancies; these are then refined by a Semantic Change Reasoning (SCR) module (using distributional divergence and response variation) and a Boundary-aware Change Refinement (BCR) module (validating interior-pixel support) to suppress artifacts while preserving genuine changes. Experiments on LEVIR-CD, WHU-CD, DSIFN and SECOND report consistent F₁^C gains of 2.13–9.75 % over prior SOTA methods together with improved computational efficiency; code is released publicly.","tokens_in":1954,"tokens_out":466,"duration_ms":22821,"significance":"If the refinement modules can be shown to deliver net-positive error reduction, the work would supply a practical, training-free route to flexible OVCD that mitigates the instance-level vs. pixel-level trade-off. Public code release supports reproducibility and is a clear strength.","major_comments":[{"comment":"Methods section describing SCR and BCR: the headline F₁^C gains are credible only if the collaborative refinement demonstrably reduces net errors on the candidate set. No ablation tables, stage-wise precision/recall breakdowns, or error-rate analysis on the initial candidate regions are supplied, leaving open the possibility that the modules trade true positives for fewer false positives or merely re-label boundary artifacts.","section":"Methods (SCR/BCR)"},{"comment":"Experiments section: overall dataset-level F₁^C numbers are reported, yet the absence of per-module contribution metrics or controlled removal of SCR/BCR prevents attribution of the observed gains specifically to the proposed reliability mechanisms.","section":"Experiments"}],"minor_comments":[{"comment":"The symbol F₁^C is used without an explicit definition or reference to the precise formulation of the change-detection F1 metric.","section":"Abstract"},{"comment":"Figure captions and axis labels in the experimental results could be expanded to clarify which curves correspond to which ablation or baseline variant.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback on the validation of the SCR and BCR modules. We agree that additional analyses are required to demonstrate net error reduction and per-module contributions. We address each major comment below and will incorporate the suggested revisions.","responses":[{"response":"We agree that the current manuscript lacks these detailed analyses on the candidate regions. In the revision we will add ablation tables with stage-wise precision, recall and F₁ scores computed on the initial candidate set before and after SCR/BCR. We will also include error-rate breakdowns quantifying net false-positive reduction versus any true-positive loss, using the same four datasets. These additions will directly address the concern that gains may arise from trading true positives or merely re-labeling artifacts.","revision_made":"yes","referee_comment":"[Methods (SCR/BCR)] Methods section describing SCR and BCR: the headline F₁^C gains are credible only if the collaborative refinement demonstrably reduces net errors on the candidate set. No ablation tables, stage-wise precision/recall breakdowns, or error-rate analysis on the initial candidate regions are supplied, leaving open the possibility that the modules trade true positives for fewer false positives or merely re-label boundary artifacts."},{"response":"We acknowledge the absence of per-module metrics and controlled ablations in the experiments section. The revised manuscript will include new experiments that systematically remove SCR, BCR, and both modules, reporting the resulting F₁^C, precision and recall on each dataset. These controlled removals, together with per-module contribution tables, will allow explicit attribution of the reported gains to the reliability-aware components.","revision_made":"yes","referee_comment":"[Experiments] Experiments section: overall dataset-level F₁^C numbers are reported, yet the absence of per-module contribution metrics or controlled removal of SCR/BCR prevents attribution of the observed gains specifically to the proposed reliability mechanisms."}],"tokens_in":1444,"tokens_out":412,"duration_ms":18352,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core contribution is a training-free method that pulls candidate regions from pixel-wise semantic differences, then applies Semantic Change Reasoning to check distributional and response signals plus Boundary-aware Change Refinement to test interior support. The abstract positions this as fixing the usual instance-versus-pixel trade-off in open-vocabulary remote sensing change detection, and the reported F1^C lifts of 2-10% across LEVIR-CD, WHU-CD, DSIFN, and SECOND plus the efficiency claim are the concrete results.\n\nThe work does a few things cleanly: it releases code, runs on standard benchmarks, and keeps the pipeline simple enough that others could re-implement the post-processing steps. The two modules target real stated weaknesses (unstable responses and boundary artifacts) rather than just adding another backbone.\n\nThe soft spot is exactly the one flagged in the stress-test note. There are no per-module ablations, no stage-wise precision/recall tables, and no error analysis on the candidate set itself. Without those numbers it is impossible to tell whether SCR and BCR are actually suppressing false positives while keeping true changes or simply trading one error type for another. The overall F1 numbers alone do not settle that question.\n\nThis paper is for the remote-sensing change-detection subgroup that already works with open-vocabulary prompts. A reader in that niche can extract the module ideas and the dataset results for their own experiments. It is not aimed at broader computer-vision audiences.\n\nThe experiments are broad enough and the framing is coherent enough that it should go to peer review rather than a desk reject. A referee can ask for the missing ablations and decide whether the reliability claim holds.","headline":"The paper adds two refinement modules (SCR and BCR) to a training-free OVCD pipeline and shows F1 gains on four datasets, but supplies no module-level ablations or error breakdowns to confirm the refinements are net positive.","tokens_in":2428,"tokens_out":427,"would_cite":false,"duration_ms":11931,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A training-free method spots land cover changes from any text prompt by deriving pixel discrepancies and refining them semantically and spatially.","keywords":["open-vocabulary change detection","remote sensing","semantic refinement","spatial refinement","training-free","land cover change","reliability-aware detection"],"falsifier":"Running the method on a dataset dominated by partial changes or heavy boundary ambiguity and checking whether its F1^C score falls below that of direct pixel comparison baselines.","tokens_in":2725,"feed_emoji":"🛰️","tokens_out":640,"duration_ms":18705,"temperature":0.7,"pith_summary":"The paper aims to show that open-vocabulary change detection can be made reliable without training by first locating candidate regions through pixel-wise semantic differences between image features and text prompts. These candidates are then passed through a Semantic Change Reasoning module that checks distributional divergence and response variation, plus a Boundary-aware Change Refinement module that verifies support from reliable interior pixels. The approach claims to resolve the trade-off between missing fine-grained changes and producing unstable boundaries or false positives. If the method works, it would enable flexible monitoring of arbitrary land cover shifts on standard remote sensing datasets while using less computation than prior techniques.","feed_headline":"Training-free method spots remote sensing changes from any text prompt","feed_subtitle":"Pixel discrepancies are refined by semantic and boundary checks to raise accuracy and cut computation on standard benchmarks.","key_machinery":"The collaborative refinement strategy that uses the Semantic Change Reasoning module to reassess changes via distributional divergence and response variation, together with the Boundary-aware Change Refinement module to validate candidate regions against reliable interior pixels.","core_discovery":"The paper claims that deriving candidate change regions from pixel-wise semantic discrepancies and then applying collaborative semantic and spatial refinement via the SCR and BCR modules produces reliable open-vocabulary change detection, yielding F1^C gains of 2.13% to 9.75% over state-of-the-art methods on LEVIR-CD, WHU-CD, DSIFN, and SECOND with improved efficiency.","pith_inferences":["The same candidate-plus-refinement pattern could be tested on video sequences to track evolving changes over time.","Integration with newer vision-language backbones might further reduce reliance on the initial pixel discrepancy step.","The modules could be applied as post-processing to other open-vocabulary segmentation tasks facing similar ambiguity issues."],"forward_implications":["Outperforms existing approaches by 2.13% to 9.75% in F1^C on four public remote sensing change detection datasets.","Achieves the gains with lower computational cost than trained state-of-the-art models.","Supports detection of changes described by arbitrary text prompts rather than fixed categories.","Suppresses incidental semantic inconsistencies while preserving genuine shifts and reducing boundary artifacts."],"fun_headline_variants":["ReA-OVCD refines open-vocabulary changes with semantic and spatial checks","Training-free framework applies collaborative refinement to OVCD","Pixel discrepancies refined semantically and spatially for reliable detection","Semantic spatial refinement ensures reliable open-vocabulary change detection"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Pixel-wise semantic discrepancies reliably generate candidate regions that the refinement modules can clean without systematically discarding true changes or retaining false positives.","fun_headline_variants_meta":{"raw":{"variants":["ReA-OVCD refines open-vocabulary changes with semantic and spatial checks","Training-free framework applies collaborative refinement to OVCD","Pixel discrepancies refined semantically and spatially for reliable detection","Semantic spatial refinement ensures reliable open-vocabulary change detection"]},"model":"grok-4.3","cost_usd":0.004566,"raw_usage":{"total_tokens":2301,"prompt_tokens":734,"num_sources_used":0,"completion_tokens":51,"cost_in_usd_ticks":45662000,"prompt_tokens_details":{"text_tokens":734,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1516,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":734,"tokens_out":51,"duration_ms":9425,"temperature":1.0,"reasoning_tokens":1516,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T18:22:06.917611+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Running the method on a dataset dominated by partial changes or heavy boundary ambiguity and checking whether its F1^C score falls below that of direct pixel comparison baselines.","supporting_citations":[],"review_version":1}