{"id":"2fef1045-a4b0-4f53-8fea-da9e61f4ad5a","arxiv_id":"2607.28186","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"On-demand spatio-temporal queries guided by ambiguity perception improve high-resolution farmland segmentation stability over single-image and passive multi-image baselines.","lead":"FarmSeeker treats farmland remote-sensing segmentation as an on-demand evidence problem: when the current image is ambiguous, an agent queries other times or a wider spatial context and refines the mask. It matters for global cropland mapping where phenology and truncated fields routinely break single-image models.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Ambiguity localization is too weak to secure the information-gain mechanism behind the headline gains.","rationale":"The reader correctly flags intermediate ambiguity perception as the soft joint in an otherwise solid engineering package. I agree that is the single most load-bearing concern for the central “information-gain agent” claim, not for the narrower fact that the full pipeline improves IoU/Recall over listed baselines. Tables 1–2, 5 and the ambiguous-subset gap are real empirical support; the IB writeup is motivational (Eq. 3 is not optimized). Sharpening the reader’s point: high tool/reasoning accuracies are conditional on detection matches and do not measure whether queried Eq is the right missing evidence. A short oracle/wrong-tool/random-box ablation would settle whether CONDITIONAL should stay (mechanism fragile) or move toward ACCEPT (60% F1 is enough in practice). No grounds for REJECT: comparisons look controlled on evidence-acquisition variants, train/bench separation is stated, and cost/practicality limits are already disclosed. Verdict remains CONDITIONAL pending stronger ambiguity module evidence or the oracle decomposition above.","tokens_in":24754,"tokens_out":710,"duration_ms":45599,"concrete_test":"On the China mixed set and the 649-sample ambiguous subset, freeze Tp/Ts and rerun three RE replacements: (1) GT ambiguity boxes + GT tool labels from the GSFS specialized subsets (oracle); (2) GT boxes + wrong tool; (3) random boxes overlapping base-error pixels + RE tool policy. Report Recall/IoU vs FarmSeeker and vs predefined MT/ST (Table 2 protocol). If oracle ≫ FarmSeeker and (2)(3) collapse toward SegEarth-R1, the F1≈60% bottleneck is load-bearing; if FarmSeeker ≈ oracle, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is not only that FarmSeeker beats SegEarth-R1/predefined fusion on GSFS-Bench (Tables 1–2), but that those gains come from on-demand extra-image evidence that fills identified information gaps (Eqs. 1–3; Think with Extra-Image). That causal story requires the RE to localize the true gap and pick the matching query. Table 6 / Supp. §4.1(4) report ambiguity-perception F1@0.5 = 60.23% and F1@0.75 = 16.69% (Precision 63.5%, Recall 58.0% at 0.5), and the authors note localization errors can cascade into query and refinement. Tool-calling accuracy (90.8%) and collaborative-reasoning accuracy (88.7%) are measured only on matched/constructed regions, so they do not rescue missed or mis-boxed gaps. End-to-end lifts (especially +MT ≈ FarmSeeker on full China Recall 88.61 vs 88.63, with the clear gap only on the 649 ambiguous subset) are therefore compatible with partial hits, SAM2 box-prompted local repair (Ts), or multi-image reasoning on noisy boxes—not only with systematic I(Y;Eq|X) maximization. Without an oracle-gap vs noisy-box decomposition, the mechanism claim is under-anchored relative to the empirical leaderboard claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper argues that high-resolution farmland remote sensing segmentation is limited less by representation capacity than by incomplete single-image evidence, and reframes the task as on-demand acquisition of task-relevant extra spatio-temporal information (Eqs. 1–3). It introduces FarmSeeker, an MLLM agent that (i) produces an initial mask with a base segmenter, (ii) localizes ambiguous regions and chooses temporal/spatial/none queries, (iii) retrieves cropped extra-image evidence, and (iv) refines local masks via collaborative reasoning and SAM2. Training uses FM-Seg69K with progressive General FT → Cold-Start FT → GRPO RFT and task-driven rewards. Evaluation is on GSFS-Bench, a new global multi-temporal/spatial queryable FRSI benchmark. Empirically, FarmSeeker improves over strong intra-image baselines in- and cross-region (Table 1), and especially on a 649-sample highly ambiguous subset (Table 2: 83.45/78.62 vs SegEarth-R1 67.34/63.95 Recall/IoU), with ablations of training stages, rewards, and on-demand vs predefined MT/SW/ST acquisition.","tokens_in":25201,"tokens_out":1503,"duration_ms":35022,"significance":"If the results hold, the work is a solid systems and benchmark contribution for agricultural remote sensing: GSFS-Bench enables evaluation of dynamic querying that prior FRSI benchmarks lack, and the closed-loop agent shows consistent correction gains across multiple base segmenters (Table 5) plus stronger robustness under phenological and FOV ambiguity. The on-demand vs predefined fusion comparison (Table 2) is practically useful. Strengths include matched base training for baselines, in-/cross-region tables, progressive-training and reward ablations, multi-backbone generality, and public project page. The information-bottleneck framing is mainly motivational rather than a fitted theory, but the empirical package is substantial for the FRSI segmentation community.","major_comments":[{"comment":"The central paradigm claim is that gains come from identifying information gaps and selecting matching extra-image evidence (Eqs. 1–3; “Think with Extra-Image”). Table 6 and Supp. §4.1(4) report ambiguity-perception F1@0.5 = 60.23% and F1@0.75 = 16.69% (Precision 63.51%, Recall 58.03% at IoU 0.5), and the text notes localization errors can propagate to querying and refinement. Tool-calling (90.79%) and collaborative-reasoning (88.72%) accuracies are measured on matched/constructed regions and do not cover missed or mis-boxed gaps. Please add a load-bearing decomposition: (a) oracle ambiguous boxes + oracle tool type, (b) oracle boxes + predicted tools, (c) predicted boxes as deployed, and report end-to-end IoU/Recall on the full China set and the 649 ambiguous subset. Without this, the mechanism story remains under-anchored relative to the leaderboard claim.","section":"Framework Generality; Table 6; Supp. §4.1(4); Eqs. 1–3"},{"comment":"Table 2 shows predefined multi-temporal (+MT) nearly matches FarmSeeker on the full China set (Recall 88.61 vs 88.63; IoU 83.70 vs 84.15), while the large gap appears mainly on the ambiguous subset (78.47/75.61 vs 83.45/78.62). This is compatible with “always fetch multi-temporal crops” explaining much of the average gain, with on-demand selection mattering chiefly in hard cases. Please quantify how often FarmSeeker issues temporal vs spatial vs none queries on China vs ambiguous splits, and report a controlled comparison of on-demand policy vs always-MT restricted to the same retrieved crops/budget so the incremental value of ambiguity-aware selection is isolated from simply having extra temporal images.","section":"Table 2; The Necessity of On-Demand Extra Spatio-Temporal Information Gain"},{"comment":"Ambiguity-perception supervision is built from multi-base-model voting (≥3 models wrong, bbox area and ≥30% error filters; Supp. §2.1). That construction correlates labels with failures of the same model family later used as Tp, raising a mild circularity risk for the “ambiguity” notion and for claimed generalization of the RE. Please report (i) agreement of vote-based boxes with independent human ambiguity annotations on a held-out sample, and (ii) whether RE detections transfer when Tp is swapped (as in Table 5) without redefining the ambiguity GT from the new base model’s errors.","section":"Supp. §2.1 GSFS-Bench Ambiguity Perception Subset; Table 5"}],"minor_comments":[{"comment":"Eqs. 1–3 define an ideal query objective that the paper correctly says is not directly optimized. State earlier (Introduction or FarmSeeker Overview) that I(Y;Eq|X) is explanatory and that the learned policy is a supervised/RL proxy, to avoid readers treating q* as an implemented estimator.","section":"Why Extra-Image…; Eqs. 1–3"},{"comment":"Several in-line spacing artifacts appear throughout (“fromaninformationbottleneck”, “spatio-temporalinformationgain”, “ThinkwithIntra-Image”). Clean typesetting before revision.","section":"Abstract; Introduction"},{"comment":"Table 1 highlights best/second-best but does not report variance or multiple seeds; even a short note on run-to-run stability for FarmSeeker (RFT sampling) would help.","section":"Table 1; Experiments"},{"comment":"Inference cost (~23.9 s/image vs 0.3 s for SegEarth-R1, Table 7) is acknowledged; a brief breakdown (ambiguity RE vs query I/O vs SAM2) would clarify optimization targets for offline mapping use cases.","section":"Table 7"},{"comment":"Clarify whether GSFS-Bench evidence pools and query APIs will be fully released with the project page, since the benchmark’s value depends on reproducible querying.","section":"GSFS-Bench; Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The empirical core and benchmark are publishable; the main risk is overselling a clean information-gain causal story while ambiguity localization is mediocre and always-MT nearly ties average China performance. If the authors deliver the oracle/noisy-box decomposition and tone the mechanism claims accordingly, this is appropriate for a solid accept after revision. Scope fits cs.CV / remote-sensing vision well."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a working closed-loop farmland segmenter that queries extra temporal/spatial tiles on demand, plus GSFS-Bench with actual queryable evidence pools. That package is new enough to matter in agricultural remote sensing even if multi-temporal fusion and RS agents already exist.\n\nWhat they actually ship is useful. FarmSeeker (perception → ambiguity boxes + tool type → crop/query → collaborative reason → SAM2 local refine) beats strong baselines on in-region and cross-region tables, and the gap widens on the 649 hard ambiguous samples. The on-demand vs predefined MT/SW/ST comparison is the right control: dumping all ST evidence is not free, and selective querying wins, especially under ambiguity. Progressive training and reward ablations are clean; multi-backbone correction shows the loop is not glued to one segmenter. They also flag the ~24s vs 0.3s cost and the offline-mapping niche honestly.\n\nSoft spots, in proportion. The IB section is framing, not an optimized objective—Eqs. 1–3 motivate; GRPO + task rewards do the work. The stress-test lands partially: ambiguity F1@0.5 is ~60% and collapses at 0.75, and the authors admit cascade risk. Tool-call and reasoning accuracies are on constructed/matched regions, so they do not fully prove systematic I(Y;Eq|X) maximization. On the full China set, predefined MT almost matches Recall; the clearest separation is the ambiguous subset and IoU. So trust the empirical claim more than the causal “information gain” branding. Mild other notes: ambiguity labels from multi-model vote, dependence on ST archives, no commit-level repro in-text (project page exists).\n\nWho it’s for: people doing high-res cropland mapping, RS agents, or multi-temporal evidence selection—not general CV theory. I’d bring it to a reading group if we care about land-cover agents. It deserves peer review; strengthen the ambiguity module and add an oracle-gap vs noisy-box decomposition and I’d cite it without hesitation. Engage.","headline":"Solid ag-RS agent paper with a real benchmark and clear end-to-end gains; the information-gain story is weaker than the leaderboard because ambiguity localization is only middling.","tokens_in":25843,"tokens_out":537,"would_cite":true,"duration_ms":16082,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Farmland segmentation fails when the current image lacks the evidence; an agent that queries extra time and space on demand is more stable.","keywords":["farmland segmentation","remote sensing","spatio-temporal information gain","vision-language agent","on-demand querying","ambiguity perception","GSFS-Bench","information bottleneck"],"falsifier":"On the highly ambiguous GSFS-Bench subset, if on-demand FarmSeeker failed to beat both the base single-image segmenter and predefined multi-temporal or spatial fusion on Recall and IoU—or if disabling ambiguity-driven querying erased the gain—the central claim would not hold.","tokens_in":25598,"feed_emoji":"🛰️","tokens_out":931,"duration_ms":20132,"temperature":0.7,"pith_summary":"Standard farmland remote-sensing segmentation assumes one high-resolution image already holds enough visual evidence. This paper argues that is often false: the same field looks like bare soil, water, or vegetation at different growth stages, and a tight crop can cut off the boundaries and neighbors that define a parcel. Ambiguity is therefore an observation bottleneck, not only a model-capacity problem. From an information-bottleneck view, the authors recast segmentation as a decision process that should seek task-relevant extra spatio-temporal information only where the current image leaves uncertainty. They build FarmSeeker, an agent that spots ambiguous regions, chooses temporal or spatial queries, reasons over the returned images, and locally refines the mask. On a new global high-resolution benchmark that supports such querying, on-demand extra evidence beats both strong single-image segmenters and fixed multi-image fusion, especially on hard ambiguous samples.","feed_headline":"Farm maps improve when the model asks for missing time and space","feed_subtitle":"An agent queries extra images only where one snapshot is ambiguous, beating fixed multi-image fusion.","key_machinery":"Task-relevant information gain of a query: I(Y; E_q | X) = H(Y|X) − H(Y|X, E_q). FarmSeeker approximates choosing the query with largest gain via ambiguity perception, tool choice (temporal, spatial, or none), collaborative multi-image reasoning, and local mask refinement, trained progressively with task-driven rewards.","core_discovery":"When the current farmland image is incomplete, actively acquiring the missing extra-image evidence that maximizes task-relevant information gain—rather than fusing all available images or only refining inside the image—produces more stable high-resolution farmland masks across regions and under phenological and spatial ambiguity.","pith_inferences":["Other land-cover tasks with strong seasonality or edge truncation (flood mapping, crop type, field boundaries) may need the same on-demand extra-image loop rather than only multi-temporal fusion.","Raising fine-grained ambiguity localization is the practical bottleneck: coarse boxes may still help, but tighter region models would likely multiply the value of each query.","If evidence pools are incomplete or misaligned in production archives, the method’s advantage should shrink toward the base segmenter—deployment needs reliable temporal/spatial retrieval, not only a better reasoner."],"forward_implications":["Single-date high-res farmland maps should treat hard parcels as queryable cases, not only as harder pixels for a bigger backbone.","Indiscriminate stacking of all temporal and spatial neighbors can hurt; matching evidence to the local information gap matters more than image count.","GSFS-Bench enables measuring not only mask quality but ambiguity perception, tool calling, and multi-image reasoning across global agricultural regions.","The same agent loop can sit on top of different base segmenters and still lift their masks, so the gain is not locked to one backbone.","The approach fits offline mapping and omission-sensitive verification more than pure real-time single-pass inference, given the extra compute on ambiguous regions."],"fun_headline_variants":["FarmSeeker queries extra images only where farmland views stay ambiguous","Spatio-temporal info gain steers on-demand queries for stable farm masks","Agent seeks missing time-space evidence instead of fusing every image","Ambiguous farmland regions trigger targeted extra-image reasoning","Beyond one snapshot: dynamic queries yield steadier high-res farm maps"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The agent must correctly find ambiguous regions and pick the right kind of extra image; if those boxes or tool choices are wrong, the new evidence may not fix the mask and can pass the error downstream.","fun_headline_variants_meta":{"raw":{"variants":["FarmSeeker queries extra images only where farmland views stay ambiguous","Spatio-temporal info gain steers on-demand queries for stable farm masks","Agent seeks missing time-space evidence instead of fusing every image","Ambiguous farmland regions trigger targeted extra-image reasoning","Beyond one snapshot: dynamic queries yield steadier high-res farm maps"]},"model":"grok-4.5","effort":"low","cost_usd":0.003681,"raw_usage":{"total_tokens":1178,"prompt_tokens":738,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":36808000,"prompt_tokens_details":{"text_tokens":738,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":347,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":738,"tokens_out":93,"duration_ms":5981,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T15:16:25.229781+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the highly ambiguous GSFS-Bench subset, if on-demand FarmSeeker failed to beat both the base single-image segmenter and predefined multi-temporal or spatial fusion on Recall and IoU—or if disabling ambiguity-driven querying erased the gain—the central claim would not hold.","supporting_citations":[],"review_version":1}