{"id":"c2463cbc-52b1-4f70-ba14-e1894a65ff94","arxiv_id":"2607.10120","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free evidence-selection and topology-preserving packaging pipeline improves frozen VLMs on ultra-high-resolution remote sensing VQA without multi-round search.","lead":"WeaveEarth is a training-free system that helps vision-language models answer questions about huge satellite images by picking a small set of key patches and packaging them with location and layout cues. It matters because ultra-high-resolution remote sensing is bottlenecked by cost and lost detail, and this approach claims better accuracy at lower latency than zoom-search or full-image compression.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The central claim that gains come from organization rather than access is under-supported by the ablations, which never hold visual budget fixed against a non-structured multi-patch baseline.","rationale":"The reader correctly flags that the greedy maximizer of Eq. (2), the default budget of 6, and the hope that VLMs will attend to SEM/TPEB are the weakest assumptions. I sharpen that concern: the manuscript never runs the decisive control that would separate “more relevant high-res patches” from “structured organization.” Fig. 4 and Table 3 are consistent with the organization story but do not rule out the simpler access story. That is a real soft spot for the central slogan, yet it does not overturn the empirical contribution—multi-benchmark, multi-backbone gains and clear efficiency wins over ZoomSearch remain. Hence the verdict stays CONDITIONAL (accept-shaped pending the matched-budget control, fuller encoder/TPEB disclosure, and variance), rather than moving to REJECT or UNCHANGED. Agreement with the reader is partial because we share the same soft spot (proxy quality of encoder similarity + reliance on SEM/TPEB) but I locate the missing experiment more specifically at the organization-vs-access isolation.","tokens_in":20018,"tokens_out":709,"duration_ms":7245,"concrete_test":"On LRS-VQA and XLRS-Bench (OSR/CR), run a matched-budget control: feed the identical MSES patches + global thumbnail to the same frozen Qwen3-VL-8B, but (a) omit SEM entirely and (b) arrange patches in retrieval order or a random grid instead of TPEB. If accuracy remains within ~1–1.5 pp of full WeaveEarth, the organization claim weakens; if it drops by ≥ the SEM/TPEB ablation deltas in Table 3, the claim holds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper’s strongest claim (Tables 1–4, Figs. 3–4) is that WeaveEarth’s gains over passive whole-image and active multi-round methods come from better evidence organization (MSES + SEM + TPEB under GCC), not from expanded visual access. The budget-sensitivity curve in Fig. 4 (peak at 6, then decline) is offered as the main evidence that “more is not better.” That curve only varies the size of WeaveEarth’s own support set; it does not compare, at a matched visual token/patch budget, a plain multi-crop baseline (top-k patches + global thumbnail, no SEM, no topology-preserving board) against the full structured interface. Ablation Table 3 removes modules one at a time from the full system, so every “w/o X” row still benefits from the remaining structure and from the same encoder-selected candidates. Consequently the design never isolates whether the frozen VLM is actually using the metadata and relative topology, or simply benefiting from receiving several high-resolution, question-relevant crops instead of a single downsampled UHR image or a long search trajectory. If a matched-budget unstructured multi-patch condition closes most of the gap, the “organizing better, not accessing more” slogan is overstated and the load-bearing assumption (encoder similarity + hand-chosen budget + SEM/TPEB are what make the support set sufficient) is only partially secured.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"WeaveEarth is a training-free, plug-and-play framework for ultra-high-resolution remote sensing VQA that reformulates the task as structured evidence construction and reasoning under global context constraints. Stage 1 (Global-Aware Evidence Construction) scores overlapping patches with question and global-thumbnail similarity (Eq. 1), expands neighborhoods, and greedily builds a compact Minimal Support Evidence Set via Rel+αCov−βRed (Eq. 2). Stage 2 (Structured Evidence Reasoning) attaches Structured Evidence Metadata and arranges patches into a Topology-Preserving Evidence Board, which, with the global thumbnail, is fed to a frozen VLM. On LRS-VQA, MME-RealWorld, and XLRS-Bench, and across Qwen3-VL-8B, LLaVA-v1.6-7B, and IXC-2.5-7B, the method reports consistent gains over passive whole-image and active multi-round UHR baselines, with ablations (Table 3), efficiency comparisons (Fig. 3), and budget curves peaking at |S|=6 (Fig. 4) offered as evidence that gains come from better organization rather than expanded visual access.","tokens_in":20384,"tokens_out":1440,"duration_ms":16676,"significance":"If the organization-over-access thesis holds, the paper offers a practical and conceptually clean alternative to both costly whole-image adaptation and high-latency multi-round search for UHR RS understanding. Strengths include: (i) training-free transfer across three frozen backbones (Table 4); (ii) multi-benchmark evaluation with sample-size-weighted averages; (iii) component ablations and an accuracy–latency comparison against representative passive and active methods; (iv) a budget-sensitivity curve that rises then falls, consistent with a “minimal yet sufficient” evidence regime; and (v) public code. These make the work useful for deployment-oriented RS-VLM pipelines even if some mechanistic claims need tighter controls.","major_comments":[{"comment":"Central claim isolation (Tables 1–4, Figs. 3–4; §4.3–4.5): The paper’s load-bearing thesis is that gains come from organizing evidence (MSES+SEM+TPEB under GCC), not from expanded visual access. Fig. 4 only varies WeaveEarth’s own |S|; Table 3 removes modules from the full system, so every ablated row still uses encoder-selected multi-patch input and residual structure. There is no matched visual-budget control that feeds the same number of high-resolution, question-relevant crops plus the global thumbnail without SEM and without topology-preserving layout (plain top-k multi-crop). Without that condition, it remains unclear whether the frozen VLM uses metadata/topology or simply benefits from several relevant high-res patches. Please add this control on at least one backbone and two benchmarks; if the unstructured multi-crop closes most of the gap, the slogan and interpretation of Figs.","section":"§4.3–4.5, Tables 3–4, Figs. 3–4"},{"comment":"Eq. (2) and greedy MSES construction (§3.2): Rel(S,q), Cov(S), and Red(S) are named but not defined operationally (feature aggregation for Rel; coverage metric for Cov; overlap/redundancy measure for Red). The claim that a training-free greedy maximizer yields a “minimal yet sufficient” support set therefore cannot be audited or reproduced from the main text alone. Please specify exact formulas, the encoder used for sim(·,·), values of λ/α/β, overlap/grid settings, and either a short justification that greedy is adequate or a small comparison against a stronger combinatorial baseline on a subset.","section":"§3.2, Eqs. (1)–(2)"},{"comment":"Hyperparameter and selection sensitivity (§4.1, Fig. 4): Free parameters include λ, α, β, evidence budget (default 6), patch grid/overlap, and TPEB layout. Only |S| is swept. Given that the weakest assumption is that encoder similarity under a global thumbnail plus hand-chosen budget yields answer-critical evidence, at least a limited sensitivity study for λ and (α,β)—or a clear statement that defaults transfer without retuning across backbones/benchmarks—is needed to support the training-free, plug-and-play claim.","section":"§4.1, Fig. 4"}],"minor_comments":[{"comment":"Table 1: WeaveEarth is listed as 8B while some compared UHR methods use 3B/7B; a short note on parameter fairness (or reporting the same backbone for all UHR methods where possible) would help readers.","section":"Table 1"},{"comment":"Figure 1 / case study (Fig. 5): “ZoomSearth” appears to be a typo for ZoomSearch; please correct consistently.","section":"Fig. 5"},{"comment":"§4.6 Limitation: Counting remains weak because small dense objects are hard in the thumbnail; the planned multi-scale extension is reasonable—consider quantifying residual counting error rates by object size if space allows.","section":"§4.6"},{"comment":"SEM format in Fig. 2 is informative; ensure the exact prompt template that injects SEM+TPEB into each backbone is in the appendix or code for full reproducibility.","section":"§3.3, Fig. 2"},{"comment":"No error bars or multi-seed variance are reported; for a systems paper this is common, but a brief note on run-to-run stability (or deterministic decoding settings) would strengthen Tables 1–4.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The skeptic’s matched-budget multi-crop control is the right load-bearing ask; if the authors add it and the structured interface still wins clearly, the paper is close to accept. If unstructured multi-crop matches full WeaveEarth, the contribution remains a strong practical multi-patch pipeline but the “organize better, not access more” framing needs to be toned down. Scope fits a multimedia/CV venue well; novelty is incremental relative to ZoomSearch/ZoomEarth/GeoEyes but the training-free structured interface is a useful synthesis. Code link is a plus for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, usable systems paper for UHR remote-sensing VLM inference. The actual contribution is a training-free pipeline that builds a compact Minimal Support Evidence Set under a global thumbnail constraint, then feeds frozen VLMs a topology-preserving multi-crop board plus explicit spatial metadata. That package beats passive whole-image methods and active multi-round search on LRS-VQA, MME-RealWorld, and XLRS-Bench, and it transfers across Qwen3-VL, LLaVA-v1.6, and IXC-2.5 without fine-tuning. Latency vs ZoomSearch is the practical win: roughly 8s vs 50s+ with higher accuracy.\n\nWhat they do well: the framing is honest, the ablations (GCC / MSES / SEM / TPEB) are coherent, the budget curve peaks then falls at |S|=6, and they ship code. Multi-backbone transfer and the spatial-relation / complex-reasoning lifts on XLRS-Bench are the strongest empirical signals. Counting remains weak, which they admit.\n\nSoft spots, in proportion. The stress-test is fair: Table 3 removes modules from the full system, and Fig. 4 only varies WeaveEarth’s own budget. There is no matched-token unstructured multi-crop baseline (top-k patches + thumbnail, no SEM/TPEB). So “gains come from organization, not access” is directionally supported but not fully isolated—some of the lift may just be “several good high-res crops.” Free parameters (λ, α, β, budget, layout) and the similarity encoder are under-specified in the main text; no error bars. None of that sinks the paper; it is a normal systems gap, not a circularity problem.\n\nWho it is for: people building Earth-observation multimodal systems who need better accuracy/latency without training. Not a foundational theory result. I would bring it to reading group, cite it when discussing UHR VLM inference interfaces, and send it to peer review. Ask referees for a matched-budget unstructured multi-patch control and fuller encoder/hyperparameter disclosure. Worth engaging.","headline":"Solid training-free UHR systems paper with real multi-benchmark gains; the “organize better” slogan is only partly isolated by the ablations.","tokens_in":21056,"tokens_out":540,"would_cite":true,"duration_ms":6401,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Ultra-high-resolution remote sensing improves when models organize a small, topology-preserving evidence set under global context, not when they simply see more pixels.","keywords":["Vision-Language Models","Ultra-High-Resolution Remote Sensing","Training-Free Understanding","Structured Evidence Reasoning","Minimal Support Evidence Set","Global Context Constraint","Topology-Preserving Evidence Board"],"falsifier":"On the same UHR benchmarks, fix the visual budget and either (a) replace the greedy MSES selection with random or purely question-only top-k patches of equal size, or (b) strip SEM and the topology-preserving board while keeping the same patches: if accuracy does not fall and the rise-then-fall budget curve disappears, the claim that organization—not access—drives the gains fails.","tokens_in":20864,"feed_emoji":"🛰️","tokens_out":1039,"duration_ms":16911,"temperature":0.7,"pith_summary":"Ultra-high-resolution remote sensing images force vision-language models to hold both a wide scene layout and sparse, answer-critical local details under tight compute budgets. The paper argues that the usual fixes—feeding larger images or running multi-round zoom search—either drop fine detail or fragment context and waste time. WeaveEarth instead treats the problem as structured evidence construction: under a cheap global thumbnail, it selects a compact, low-redundancy, spatially complementary Minimal Support Evidence Set, then weaves those patches with explicit spatial metadata and a topology-preserving board so a frozen model can reason jointly over global layout and local evidence in one pass. Across several UHR benchmarks and multiple frozen backbones, this training-free pipeline beats strong baselines and prior UHR methods, and ablations plus budget curves show the gains come from better organization, not from exposing more visual content. A sympathetic reader cares because the same budget constraint appears in any large-image multimodal setting where the answer lives in a few regions but depends on their place in the whole scene.","feed_headline":"Organize evidence, don't just zoom: UHR sensing gains","feed_subtitle":"A compact topology-preserving patch set under a global thumbnail beats multi-round search on frozen VLMs.","key_machinery":"Minimal Support Evidence Set (MSES) under Global Context Constraint, plus Structured Evidence Reasoning via Structured Evidence Metadata (SEM) and a Topology-Preserving Evidence Board (TPEB). Together they select a small complementary patch set and re-present it so the model retains spatial grounding and relative layout for joint global-local reasoning.","core_discovery":"The paper claims that for ultra-high-resolution remote sensing, effective understanding is a problem of constructing and organizing the right evidence under global context constraints, not of expanding visual access. Its training-free framework, WeaveEarth, first builds a compact Minimal Support Evidence Set that is relevant, low-redundancy, and spatially complementary, then feeds a frozen vision-language model a unified interface of global thumbnail, structured evidence metadata, and a topology-preserving evidence board, yielding consistent accuracy gains over passive whole-image adaptation and active multi-round search.","pith_inferences":["If encoder similarity under a global thumbnail is the bottleneck, swapping in a stronger or task-adapted cross-modal encoder could raise the ceiling without changing the rest of the pipeline.","The counting limitation the authors note suggests multi-scale or instance-aware evidence units as a natural next module when targets are dense and tiny.","The same ‘organize better, not access more’ principle may transfer to other large-image domains (pathology slides, satellite video frames) where answers depend on sparse regions and long-range layout."],"forward_implications":["UHR remote-sensing VQA can be improved without fine-tuning backbone VLMs by redesigning only the inference input interface.","Passive resolution scaling and multi-round zoom search are not the only viable routes; a single-pass structured evidence interface can match or beat them at lower latency.","Spatial-relation and complex-reasoning subtasks benefit most when local patches retain explicit coordinates, roles, neighbors, and relative layout.","Evidence quantity has a sweet spot: too few patches miss clues, too many dilute them, so minimal-yet-sufficient sets are preferable to always adding more crops.","The same construction-plus-organization pattern can be dropped onto different frozen open-source VLMs with stable relative gains."],"fun_headline_variants":["Organize, don't expand: WeaveEarth reweaves UHR evidence for VLMs","Minimal support patches + topology beat multi-round zoom in UHR RS","Structure evidence, freeze the VLM: WeaveEarth for UHR remote sensing","Global thumbnail and complementary patches unlock UHR without training","Evidence construction over zoom: training-free UHR RS gains on frozen VLMs"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A training-free greedy score of question similarity plus global-thumbnail consistency, coverage, and redundancy, with a fixed small evidence budget and a hand-arranged board layout, is enough for frozen vision-language models to treat the selected patches and metadata as truly minimal yet sufficient support for spatial answers.","fun_headline_variants_meta":{"raw":{"variants":["Organize, don't expand: WeaveEarth reweaves UHR evidence for VLMs","Minimal support patches + topology beat multi-round zoom in UHR RS","Structure evidence, freeze the VLM: WeaveEarth for UHR remote sensing","Global thumbnail and complementary patches unlock UHR without training","Evidence construction over zoom: training-free UHR RS gains on frozen VLMs"]},"model":"grok-4.5","effort":"low","cost_usd":0.00196,"raw_usage":{"total_tokens":916,"prompt_tokens":836,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":19600000,"prompt_tokens_details":{"text_tokens":836,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":0,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":836,"tokens_out":80,"duration_ms":1536,"temperature":1.0,"reasoning_tokens":0,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T14:08:33.499398+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same UHR benchmarks, fix the visual budget and either (a) replace the greedy MSES selection with random or purely question-only top-k patches of equal size, or (b) strip SEM and the topology-preserving board while keeping the same patches: if accuracy does not fall and the rise-then-fall budget curve disappears, the claim that organization—not access—drives the gains fails.","supporting_citations":[],"review_version":1}