{"id":"b959e7b3-c3e1-4175-bf4e-3788da05cc72","arxiv_id":"2607.25993","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Training on multi-tool visual reasoning trajectories (zoom, grounding, lines) with an attention-focused RL objective improves UHR remote-sensing VQA accuracy over single-tool zoom-in and larger base models.","lead":"A 7B multimodal model learns to combine zoom-in, object grounding, and drawn auxiliary lines to answer questions about city-scale satellite images, beating much larger models on three ultra-high-resolution remote-sensing benchmarks. The paper also contributes a 13K-sample multi-tool reasoning dataset and an attention-focused reinforcement-learning objective.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) has zero gradient at the stated initialization, so RTAL cannot implement the claimed positive-advantage reinforcement; the +1.8% over GRPO in Table 16 is unexplained by the presented math.","rationale":"The reader's CONDITIONAL verdict focuses on annotation quality, but that concern, while real, is downstream: even perfect traces would not rescue the paper's central methodological claim if RTAL's objective is mathematically inert. The strongest claim explicitly attributes the SOTA to 'multi-tool visual reasoning with tool-attention RL'; RTAL is contribution (3) and the subject of Eq. (3)–(8) and Table 16. If Eq. (3) cannot reinforce positive-advantage tool-attention patterns, the reported 54.2% vs 52.4% (RTAL vs GRPO) is not explained by the proposed algorithm. This is not a question of missing error bars or code; it is an internal inconsistency in the derivation. The JSD gradient vanishes at the initialization point that the paper itself defines (old = pre-update policy). A standard policy-gradient or KL-regularized objective would have nonzero gradient; this one does not. If the authors intended a reverse direction or a separate reference policy, they need to say so, because the current text cannot reproduce the ablation. I therefore recommend REJECT: the central claim about RTAL is unsupported as written, even though the GeoMTVR dataset and the multi-tool SFT recipe may still be useful contributions.","tokens_in":22428,"tokens_out":8554,"duration_ms":81251,"concrete_test":"Compute the gradient of Eq. (3) w.r.t. θ analytically at pθ = pold: ∇θ JSD(pθ∥pold)|_{pθ=pold} = 0, so the attention loss is zero at initialization. Then run the exact RTAL recipe from Tables 12–13 with λ_attn=0 vs λ_attn=1 (three seeds each) on the same SFT checkpoint; if the XLRS-Bench averages differ by more than the seed noise, the implementation must be using an objective different from Eq. (3), and that objective should be stated and re-analyzed.","verdict_should_be":"REJECT","load_bearing_attack":"The central novelty is RTAL, but as written its objective is inert. In Eq. (3), L_tool_AttnRL = Aτ · (1/|C|) Σ_{t∈C} JSD(˜pθ_t ∥ ˜pold_t), with ˜pold_t defined as 'the frozen attention policy before the current update' (§4.2.1). At the start of every update θ is the current policy, so ˜pθ_t = ˜pold_t for all t; the JSD and its gradient w.r.t. ˜pθ_t vanish identically. Therefore the attention term contributes zero gradient regardless of the sign or magnitude of Aτ. A positive advantage cannot 'pull the current tool attention toward successful historical routing' because the minimum of the JSD is already attained at the old policy; a negative advantage cannot 'push away' either, since the gradient is also zero at that point. If PPO-style multiple epochs are used, after the first inner step pθ ≠ pold and the term becomes a trust-region regularizer, not a reinforcement signal: for A>0 it pulls back toward the behavior policy rather than increasing the likelihood of the successful tool-attention pattern. Thus the +1.8 point gain of RTAL over GRPO in Table 16 cannot be produced by Eq. (3) under the stated definitions; it must come from an unspecified implementation choice (e.g., a reversed direction, a non-frozen reference, or an additional surrogate), which the paper does not describe.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GeoMTVR, a 13K-sample ultra-high-resolution remote-sensing VQA dataset with interleaved multi-tool reasoning traces (crop-and-zoom, grounding, auxiliary-line drawing), and proposes Reinforced Tool Attention Learning (RTAL), an attention-level RL objective intended to focus policy updates on tool-call spans. The resulting GeoLens model (Qwen2.5-VL-7B backbone) is reported to achieve 54.2% on XLRS-Bench, 60.7% on LRS-GRO-eval, and 48.8% on RSHR-Bench, surpassing much larger models including Qwen3-VL-235B. The authors claim that multi-tool visual reasoning with tool-attention RL is the key ingredient, supported by a pilot study showing the limits of single-tool zoom-in and by ablations comparing SFT, GRPO, and RTAL.","tokens_in":22802,"tokens_out":8334,"duration_ms":80124,"significance":"If the reported results hold, the paper makes a substantial empirical contribution: it demonstrates that a 7B model can outperform 235B-scale models on UHR RS benchmarks by combining domain-specific tool-use data with RL, and it releases a first-of-its-kind multi-tool trajectory dataset (GeoMTVR). The pilot-study decomposition into fixed and dynamic failure sets is a useful diagnostic. However, the central algorithmic novelty, RTAL, is not supported by the presented mathematics: Eq. (3) has zero gradient at the stated initialization, so the +1.8-point RTAL-vs-GRPO gain in Table 16 cannot be produced by the objective as written. This is a load-bearing issue that must be addressed before the results can be interpreted.","major_comments":[{"comment":"The RTAL objective L_tool_AttnRL = Aτ * (1/|C|) Σ_{t∈C} JSD(pθ_t ∥ pold_t) is inert at the start of each update. Since pold_t is defined as 'the frozen attention policy before the current update', at the beginning of an update we have pθ_t = pold_t for all t, and the JSD and its gradient vanish identically. Thus a positive advantage cannot 'pull' attention toward successful routing and a negative advantage cannot 'push' it away. If multiple inner PPO steps are used, after the first step the term with A>0 acts as a trust-region regularizer pulling back toward the pre-update policy, not as a reinforcement signal; with A<0 it pushes away. The claimed mechanism, and therefore the +1.8-point gain over GRPO in Table 16, is unexplained by Eq. (3). The authors must either correct the objective/reference-policy definition so that the gradient is non-vanishing and has the stated sign, or report th","section":"§4.2.1, Eq. (3)"},{"comment":"The headline comparisons lack statistical support. Full-tool SFT vs. zoom-only SFT differs by +0.7 points on average (50.3 vs. 49.6); RTAL vs. GRPO differs by +1.8 points (54.2 vs. 52.4). With 13 sub-tasks, many individual sub-task differences go in opposite directions (e.g., Table 16: GRPO is better on OLUC, RLUC, OCC, OCL; RTAL is better on RP, RCCD, CCR). No error bars, confidence intervals, or multiple-seed runs are reported. The inconsistency between Table 5 (Qwen2.5-VL-7B = 47.4) and Table 15 (same model = 44.9) further undermines the ablations. Please report per-seed variance or bootstrap CIs, and reconcile the base-model numbers.","section":"Tables 14–16"},{"comment":"GeoMTVR is built with a staged automatic pipeline: Qwen3-VL-235B coarse cropping, GPT-5.2 trace generation, and GroundingDINO localization. The paper acknowledges that direct MLLM annotation 'often leads to unreliable results', but no human-quality audit is reported for the final GeoMTVR traces. Since SFT and RL are trained on these trajectories, hallucinated or mislocalized tool calls would teach incorrect tool-use behavior. Please provide a human-annotated sample audit (e.g., tool-call correctness, box IoU, line validity, and answer-trace alignment) or another quantitative quality validation.","section":"§3, Fig. 2"},{"comment":"Data deduplication is described only as 'strict image-level deduplication'. Because GeoMTVR is built from SuperRSVQA and LRS-GRO-train, and evaluation is on XLRS-Bench and LRS-GRO-eval, the procedure for detecting near-duplicates or same-source images is critical. If large scene images are cropped differently or the same source tiles appear in both training and test, the reported SOTA could be inflated. Please clarify the deduplication method, including any perceptual-hash or metadata-based checks, and report the number of removed samples.","section":"§3, Tables 10–11"}],"minor_comments":[{"comment":"The pilot experiment text says 'As shown in Table 1' but the table is numbered Table 2. The arrow notation in Tables 14–16 (e.g., '48.9↓+0.7') is confusing; an upward arrow or explicit 'improvement' would be clearer.","section":"§3 / Table 2"},{"comment":"The text says the average resolution of GeoMTVR is 'about 9K × 9K', but Table 3 reports 6,033×6,017. Please correct the inconsistency.","section":"§3"},{"comment":"The column abbreviations (COL, OGD, MRJC, etc.) are not expanded in the caption or main text. Please provide a metric-name table or expand the abbreviations in the caption.","section":"Table 6"},{"comment":"The description says the RTAL 'attention regularization' uses tool-action and response masks, but it does not specify how the JSD term in Eq. (3) is computed for the response mask or what λ_attn is set to. Reporting λ_attn and the masking details would help reproducibility.","section":"Appendix A.2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's citations are heavily self-referential (GeoLLaVA, XLRS-Bench, GeoEyes, and related works from the same group). This is not a disqualifier, but the novelty claim relative to GeoEyes/ZoomEarth should be sharpened. The most urgent issue is Eq. (3): as written, RTAL cannot produce the reported gain. If the authors have an implementation that works, they need to present the actual objective and show its gradient is non-zero at the relevant points. The empirical results are intriguing but need significance testing and a human audit of the dataset before they can be taken as established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the dataset and the pilot study are the real contributions; the RL algorithm is not. Eq. (3) defines L = A * mean JSD(pθ_t || pold_t), with pold_t frozen before the update. At the start of each update pθ = pold, so the JSD and its gradient are identically zero; the loss cannot pull attention toward or push it away from anything. After the first inner step it becomes a trust-region regularizer, not a reinforcement signal. So the +1.8-point gain over GRPO in Table 16 cannot be produced by the math as written. That is load-bearing because RTAL is the paper's second headline contribution. The authors need to specify the actual update—reversed direction, non-frozen reference, or a different surrogate—before I'd believe the RL numbers.\n\nWhat's genuinely good: GeoMTVR is a new resource—13K UHR samples with zoom, grounding, and line trajectories—and the staged annotation pipeline (Qwen3-VL coarse crop, GPT-5.2 trace, GroundingDINO localization) is a sensible response to the hallucination problem, even if no human audit is reported. The pilot study's distinction between localized and dispersed evidence, and the \"moving failure set\" analysis (fixed vs dynamic Quality), is well motivated and is the most interesting empirical content. The SFT ablation (51.7 vs 44.9) shows the dataset itself helps. The 7B model beating much larger baselines is notable if the numbers hold.\n\nSoft spots beyond RTAL: no error bars or significance tests anywhere, and the multi-tool-vs-zoom-only gap is +0.7 points (Table 14) with a strange '↓+0.7' label; Table 5 has a misformatted Qwen3-VL-235B row that looks like shuffled columns; there is no direct comparison to the same group's GeoEyes; and the self-citation pattern to GeoLLaVA, XLRS-Bench, and GeoEyes is heavy. None of these are fatal, but they are exactly what a referee should push on. The annotation-quality concern is real too: no human audit, so the SFT may be imitating biased traces.\n\nBottom line: worth sending to peer review because the dataset and pilot are useful to the community, but RTAL needs to be fixed or removed, and the empirical claims need error bars. I wouldn't cite the RL method as-is, but I'd cite the dataset once it is released.","headline":"GeoMTVR is a genuinely useful dataset and the pilot study is thoughtful, but RTAL's Eq. (3) as written has zero gradient at the start of every update, so the +1.8% headline gain is unexplained.","tokens_in":23347,"tokens_out":2599,"would_cite":false,"duration_ms":26480,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Single-tool zoom-in saturates on hard satellite-reasoning tasks; the paper's answer is a multi-tool loop—crop, ground, and draw lines—trained with attention-focused reinforcement learning, which beats far larger models on three benchmarks.","keywords":["ultra-high-resolution remote sensing","multi-tool visual reasoning","zoom-in","object grounding","auxiliary lines","reinforcement learning","visual question answering","tool-attention learning"],"falsifier":"Run GeoLens's RL stage with tool-call tokens replaced by non-functional placeholders (or with returned observations blanked); if accuracy on XLRS-Bench stays near 54.2%, the reported gains come from the reasoning text rather than the tool calls. Alternatively, have human annotators audit a random subset of GeoMTVR trajectories and measure the fraction of tool calls whose outputs actually localize the queried evidence; if that fraction is low, the training signal is contaminated.","tokens_in":22307,"feed_emoji":"🛰️","tokens_out":8869,"duration_ms":77273,"temperature":0.7,"pith_summary":"Ultra-high-resolution satellite images are too large for a model to answer questions from a single global view, and this paper claims that the usual fix—letting the model zoom in—only solves easy cases. Through a pilot study on the XLRS-Bench benchmark, the authors show that zoom-in saturates on tasks requiring global counting, multi-region comparison, path planning, or evidence scattered across the image. To go beyond zoom, they build GeoMTVR, a dataset of 13,000 satellite question-answer samples annotated with interleaved reasoning traces that call three tools: crop-and-zoom, object grounding, and auxiliary-line drawing. They then train GeoLens in two stages—supervised fine-tuning on those traces, followed by a reinforcement-learning step that concentrates gradients on the attention patterns at tool-call tokens. On three benchmarks, GeoLens outperforms direct-reasoning and zoom-only baselines and, with a 7-billion-parameter model, beats closed models two orders of magnitude larger.","feed_headline":"7B model outdoes 235B on satellite visual QA by using three tools","feed_subtitle":"GeoLens shows that active tool choice, not raw model size, unlocks hard remote-sensing questions.","key_machinery":"The engine of the paper is the combination of GeoMTVR, a supervised dataset of 13,078 ultra-high-resolution VQA samples with interleaved tool-use trajectories, and Reinforced Tool Attention Learning (RTAL). RTAL is the key mechanism: after standard supervised fine-tuning, it extracts the time steps inside tool-call spans, normalizes the model's final-layer attention over the valid causal context to form a 'tool-conditioned causal attention policy,' and applies an advantage-weighted Jensen–Shannon divergence so that high-reward trajectories pull the attention routing at tool tokens toward successful patterns while low-reward trajectories push it away. A time-step mask restricts backpropagatio","core_discovery":"The paper's central claim is that evidence acquisition in ultra-high-resolution remote sensing is a multi-tool problem, not a resolution problem. The authors report that a single zoom-in tool resolves tasks whose evidence is locally recoverable, such as object classification, regional counting, and attribute recognition, but leaves a 'moving failure set' behind: as training solves easy failures, the remaining hard cases require global coverage, comparison across distant regions, and structured spatial operations that one crop cannot provide. The proposed solution is to let the model also call an object-grounding tool and an auxiliary-line tool, and to train the policy with an attention-level","pith_inferences":["Editorial inference: the 'moving failure set' result likely generalizes beyond zoom-in; any domain with a single narrow evidence-gathering operation, such as OCR on documents or depth maps in robot navigation, should show the same pattern of easy wins followed by saturation as tasks require heterogeneous evidence.","Editorial inference: RTAL's success implies a testable prediction about internal representations—models trained with RTAL should show measurable differences in attention entropy and cross-step routing at tool tokens compared with token-level RL models, and these differences should correlate with tool-call quality.","Editorial inference: because the annotation pipeline is semi-automatic, a cheap quality check is to run the same staged pipeline with and without the coarse-crop stage and measure downstream task accuracy; if traces are sensitive to that stage, dataset noise is likely concentrated in localization, pointing to where further human cleaning would pay off.","Editorial inference: explicit tool calls produce inspectable intermediate evidence—crop boxes, object boxes, drawn lines—which could make GeoLens-style systems easier to audit than end-to-end black-box answers, but only if the tool outputs themselves are trustworthy."],"forward_implications":["If correct, model scale is not the dominant factor for ultra-high-resolution remote-sensing QA: a 7-billion-parameter model trained on multi-tool reasoning outperforms models with far more parameters on the same benchmarks.","Zoom-in should be treated as one tool among several; tasks requiring global coverage, multi-region comparison, or long-range spatial reasoning need grounding and line-drawing operations to gather dispersed evidence.","Attention-level reinforcement is a viable credit-assignment mechanism for tool-use policies: concentrating optimization on tool-call spans improves accuracy and yields more decisive, less redundant tool trajectories than token-level RL.","Tool-use trajectories can be generated at scale by a staged semi-automatic pipeline that separates coarse question-region cropping from trace generation and object localization, then used to teach active multi-tool reasoning.","UHR remote-sensing MLLMs should be designed as active, task-adaptive agents that interleave reasoning with visual tool calls rather than as single-pass image understanders."],"fun_headline_variants":["Multi-tool reasoning beats zoom-in alone on UHR satellite QA","7B GeoLens uses three tools to outscore 235B on remote-sensing","Zoom-in isn't enough: GeoLens teaches MLLMs to pick tools","From zoom to toolkit: GeoLens conquers hard UHR visual QA","Tool planning lifts small model past 235B on satellite tasks"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The staged annotation pipeline produces correct, non-hallucinated tool-use trajectories; if the coarse crops, generated traces, or box localizations are systematically wrong, supervised fine-tuning teaches the model bad tool behavior and the reinforcement-learning gains are built on that flawed foundation, and the paper does not report a human-quality audit of these traces.","fun_headline_variants_meta":{"raw":{"variants":["Multi-tool reasoning beats zoom-in alone on UHR satellite QA","7B GeoLens uses three tools to outscore 235B on remote-sensing","Zoom-in isn't enough: GeoLens teaches MLLMs to pick tools","From zoom to toolkit: GeoLens conquers hard UHR visual QA","Tool planning lifts small model past 235B on satellite tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000647,"raw_usage":{"total_tokens":2857,"prompt_tokens":841,"completion_tokens":2016,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":1916}},"tokens_in":585,"tokens_out":2016,"duration_ms":13122,"temperature":1.0,"reasoning_tokens":1916,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:53:06.238337+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GeoLens's RL stage with tool-call tokens replaced by non-functional placeholders (or with returned observations blanked); if accuracy on XLRS-Bench stays near 54.2%, the reported gains come from the reasoning text rather than the tool calls. Alternatively, have human annotators audit a random subset of GeoMTVR trajectories and measure the fraction of tool calls whose outputs actually localize the queried evidence; if that fraction is low, the training signal is contaminated.","supporting_citations":[],"review_version":1}