{"id":"6cf7ed4e-952b-42d6-a044-829cbca30385","arxiv_id":"2611.16408","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"G&F combines Evidence-of-Thought supervised fine-tuning and Relational Transfer Policy Optimization to improve vision-language model reasoning on complex remote sensing scenes.","lead":"The paper introduces Gravity and Flight (G&F), a two-stage framework that fine-tunes vision-language models for fine-grained remote sensing scene understanding with limited supervision. It uses paired images under a shared task and a rubric-based verification that separates visual fidelity checks from reasoning evaluation, improving model judgment accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Single-run harbor evaluation leaves unknown variance from response sampling, random matching, and F2 repeat-change (29.17%); the 4.56-point gain over supervised initialization may not be statistically robust.","rationale":"The reader identified verifier reliability as the weakest assumption. My concern extends to the combined effect of verifier noise, random response sampling, and random matching on the stability of the reported numeric gain. The paper's own reliability study (Table S4) shows non-trivial repeat-change rates, and Section G confirms single-run reporting. However, the extensive ablations, matched controls, and consistent patterns across settings provide converging evidence that the method improves over baselines in a reliable direction. The precise size of the improvement remains uncertain, but this does not invalidate the core contribution of the framework. Therefore I recommend the reader's ACCEPT verdict stand unchanged, though the authors should be expected to add multiple-run statistics in future work to strengthen the quantitative claim.","tokens_in":25606,"tokens_out":4875,"duration_ms":40686,"concrete_test":"Run the full harbor RTPO training with five different random seeds, each producing a final model. Evaluate all models on the same six evaluation sets and report the mean and standard deviation of the overall score. If the mean improvement over M1 is at least 2 points and the standard deviation is less than 1 point, the claim is robust; otherwise the improvement may not be reliably attributable to the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper reports a single run of each experimental setting (Section G: 'One run is reported per setting'). RTPO training involves multiple stochastic sources: (1) random response sampling at temperature 0.6, (2) random one-to-one matching of responses into candidate pairs, (3) verifier score variability as measured by repeat-change rates (F2: 29.17%, Table S4). Without multiple seeds, it is unknown whether the observed 4.56-point improvement over M1 is robust to these choices. The 100-unit training trace (Table S5) shows that 57.9% of pairs with one full-credit response receive a negative shared advantage, indicating that pair composition strongly affects credit assignment; a different random match could shift the advantage distribution and the resulting policy. The verifier's repeat-change rate of 29.17% for F2 means nearly a third of identical inputs yield different scores, adding noise to the paired-adaptation term. The paper acknowledges this variability (b472) but does not bound its impact on final evaluation scores. While ablations show consistent rank ordering (full RTPO > ablations), the magnitude of the gain over M1—a central claim—is not accompanied by any variance estimate. This is the most load-bearing concern because the headline numeric claim (69.12 overall, +4.56 over M1) could shift with run-to-run variation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proposes Gravity and Flight (G&F), a two-stage post-training framework for fine-grained remote sensing scene specialization. Stage 1 (EoT-SFT) uses evidence-of-thought demonstrations to teach the model to describe visual details, analyze task-relevant evidence, and draw supported conclusions. Stage 2 (RTPO) organizes training around scene-matched image pairs, using two rubric-based verifiers—Gravity for visual fidelity and coverage, Flight for evidence use and paired adaptation—whose scores combine with task correctness into a shared pair-level advantage that couples policy updates from both responses. On six harbor evaluation sets, G&F achieves a mean score of 69.12, exceeding the backbone Qwen3.5-4B by 9.76 points and the supervised-initialization baseline by 4.56 points. Ablations, matched controls, and a separate urban instantiation support the contribution of each component. Training uses 1,008 demonstrations and 500 image pairs, and code is released.","tokens_in":26071,"tokens_out":3395,"duration_ms":30064,"significance":"If the reported gains are robust, G&F represents a well-motivated and empirically grounded contribution to fine-grained visual reasoning under limited supervision. The separation of visual-fidelity verification (Gravity) from conditional-reasoning verification (Flight), combined with the paired-scene training design (RTPO), is novel and clearly explained. The paper provides extensive ablations (task-only, w/o Gravity, w/o Flight, single-image RL, paired RL w/o F2, shuffled F2, joint-input RL) that show a consistent rank ordering in favor of the full method. The urban specialization experiment demonstrates transfer to a second domain with unchanged rubric dimensions and weights, strengthening the claim of framework generality. The open release of code, data, and evaluation protocols supports reproducibility. The verifier reliability study (Table S4), while indicating moderate agreement for the most complex dimension (F2: 83.33% within one point, 29.17% repeat-change rate), is transparently reported and the limitations are acknowledged.","major_comments":[{"comment":"The paper reports a single run per experimental setting. RTPO training involves multiple stochastic sources: (1) random response sampling at temperature 0.6, (2) random one-to-one matching of responses into candidate pairs, and (3) verifier-score variability (F2 repeat-change rate: 29.17%, Table S4). The 4.56-point gain over M1—a central claim—is not accompanied by any variance estimate. Without multiple seeds or bootstrapped confidence intervals, it is unknown whether this gain is statistically robust. The paper acknowledges the limitation only for cost differences (b615: 'differences below a few percent should be read as run-to-run variation rather than as systematic cost differences'), not for performance. This is load-bearing because the headline numeric result could shift with run-to-run variation.","section":"Section G, sentence b615","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The verifier reliability study reports F2 (paired adaptation) agreement within one point of human ratings at 83.33% and a repeat-change rate of 29.17% on 24 units. The paper acknowledges (b475) that '24 units support an indication of reliability rather than a precise estimate' and (b472) that variability 'can perturb the relative rewards used to compute the advantage.' However, the paper does not bound how this verifier noise propagates to the final evaluation scores or the policy update. Since the reward signal is central to RTPO, the sensitivity of the reported gains to verifier noise should be assessed, e.g., by re-running the evaluation with perturbed scores or by reporting variance across verifier calls.","section":"Table S4 and Section C.2","authors_acknowledge":"Such variability can perturb the relative rewards used to compute the advantage.","authors_acknowledge_p":0.71}],"minor_comments":[{"comment":"The 'Data Scale' column aggregates different types of supervision (pretraining, instruction tuning, paired training) into a single number. For G&F, 2K items includes both demonstrations and pairs, while for GeoChat (318K) the count includes full instruction-tuning data. Clarify what each count covers so that the 'limited supervision' claim is interpretable as an apples-to-apples comparison.","section":"Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The base-model scores on EarthVQA Urban (11.00) and RSVQA HR (15.00) are very low, which makes the relative gains (+20.43 and +28.67 after +RTPO) large in absolute terms but also raises the question of whether the base model is fundamentally unsuited to the urban evaluation protocols. Report whether the evaluation metrics are the same as those used for harbor, and confirm that the base model's low scores are not an artifact of formatting or answer extraction.","section":"Table 3, urban results","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The comparison to general VLMs (LLaVA-1.5, Qwen3.5) is informative but the data-scale gap makes it hard to separate domain specialization from base-model capability. Consider adding a column for domain-specific vs. general model status, or note explicitly that G&F fine-tunes a 4B model while several baselines use 7B parameters.","section":"Section 4.2, Table 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The policy's available reference inputs are listed as 'Own image and shared question with public conditions.' This is stated elsewhere but the table would benefit from a note that the public conditions are part of the shared instruction q.","section":"Table S2, line 'Policy' entry","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The cost measurements (Table 2) are wall-clock time on a specific two-GPU machine with online-verifier waiting. While this is informative, the practical deployment cost is dominated by verifier API calls. Report the approximate number of verifier calls per setting (27,000 per pass is stated in Section D.2; this could be moved into Section G or Table 2).","section":"Section G, cost reporting","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Minor typographical issues: line 'relational reasoning' is hyphenated inconsistently; 'G&F achieves a higher mean score than the backbone' could be 'G&F achieves higher mean scores than the backbone' for grammatical agreement. These do not affect clarity.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"This is a well-built paper that does what it says on the tin. The two-stage design—EoT-SFT to teach structured observation, then RTPO to refine on scene-matched pairs with decoupled visual and reasoning verifiers—is clearly motivated and carefully ablated. The authors construct their own supervision data (1,008 demos, 500 pairs) with explicit Evidence and Boundary Cards, run matched controls (single-image RL, joint-input RL, shuffled F2 feedback), and include a verifier reliability study. They also demonstrate transfer to urban scenes with the same rubric, which supports their claim of framework reusability. The ablations are the paper's strongest feature: they show that removing Gravity or Flight hurts, that aligned relational feedback beats shuffled feedback, and that the gain over supervised initialization (4.56 points) is not just from more compute. Credit where due—this is serious engineering with thorough documentation.\n\nThe soft spots are real but not disqualifying. The main one is the stress-test concern: a single run per setting, with no variance estimate for the headline score of 69.12 or the +4.56 gain over M1. The paper itself says that one run is reported per setting and that differences below a few percent should be read as run-to-run variation (b615), but that sentence is about wall-time cost, not evaluation scores. Given the stochastic sources (response sampling at temperature 0.6, random one-to-one matching, F2 repeat-change rate of 29.17%), the reader cannot assess whether the central numerical claim is robust. The verifier reliability is moderate—83.33% agreement within one point for F2 on 24 units—and the authors acknowledge that 24 units support an indication rather than a precise estimate (b475). The training data (500 pairs) and evaluation sets are small, and the method's dependence on a single fixed judge (DeepSeek-V4.1-Flash) means the results may not transfer to another verifier. These are all acknowledged, which is to the authors' credit, but they do limit confidence.\n\nThis paper is for researchers working on RL-based fine-tuning of VLMs for specialist domains, particularly remote sensing. It will be cited in that subfield. The method is clearly described and the ablations are informative. The main unresolved issue—single-run variance—is a standard weakness that peer review can address by asking for multi-seed experiments or a bootstrapped confidence interval. It does not undermine the paper's overall structure or claims.\n\nDesk rejection would be too harsh. Send it to review. The paper deserves a serious referee who can ask for a variance estimate and push the authors to bound the run-to-run uncertainty.","headline":"Solid empirical paper with a clean method and honest limitations, but the single-run evaluation leaves the headline improvement unaccompanied by any variance estimate.","tokens_in":26502,"tokens_out":1368,"duration_ms":13343},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Gravity and Flight uses separate visual and reasoning verification to improve fine-grained remote sensing scene reasoning by 9.76 points over the base model.","keywords":["remote sensing","fine-grained reasoning","vision-language model","reinforcement learning","policy optimization","rubric-based verification","paired supervision","evidence-grounded reasoning"],"falsifier":"A controlled experiment where the same training pipeline is run with a different verifier (e.g., an open-source judge of comparable size or a smaller LM) would test whether the gains depend on the specific judge's quality; if performance drops substantially, the claim that the rubric verification itself is responsible is weakened. Alternatively, constructing test pairs where the verifier is known to misjudge the relation (e.g., ambiguous change types) could reveal brittleness in the shared-advantage mechanism.","tokens_in":27163,"feed_emoji":"🛰️","tokens_out":3554,"duration_ms":32147,"temperature":0.7,"pith_summary":"The paper argues that vision-language models reasoning over complex remote sensing scenes often miss fine-grained visual details and draw conclusions beyond what the evidence supports. To address this with limited supervision, the authors propose Gravity and Flight (G&F), a two-stage post-training framework that separates visual verification (Gravity) from reasoning verification (Flight) and couples policy updates across scene-matched image pairs. Evidence-of-Thought supervised fine-tuning teaches the model to describe visible details, analyse task-relevant evidence, and draw supported conclusions. Relational Transfer Policy Optimization then refines the policy on matched pairs where the same question applies to two images with fine-grained visual differences, using Gravity to check visual fidelity and coverage and Flight to evaluate evidence use and cross-scene judgment consistency. On six harbor evaluation sets, G&F achieves a mean score of 69.12, exceeding the backbone by 9.76 points and supervised initialization by 4.56 points, and matched controls show gains over single-image and joint-input reinforcement learning. A separate urban specialization also yields improvements, suggesting the framework can transfer to other scene domains.","feed_headline":"Two-stage framework lifts remote sensing reasoning 9.76 points","feed_subtitle":"Gravity and Flight separates visual and reasoning checks to teach VLMs fine-grained scene judgments with limited supervision.","key_machinery":"The key machinery is Relational Transfer Policy Optimization (RTPO), which operates on scene-matched image pairs. For each pair, the model generates responses independently for both images. A rubric-based verifier (DeepSeek-V4.1-Flash) assigns separate scores: Gravity (visual fidelity and description coverage) checked against the image and an Evidence Card; Flight (evidence use and paired adaptation) assessed conditional on the generated descriptions, using the public task and comparison metadata. These scores combine with task correctness into a pair reward. Rewards are group-normalized within each training unit to produce a shared pair-level advantage that updates the policy for both responses, coupling the learning across the two scenes.","core_discovery":"The central claim is that decoupling visual fidelity verification from conditional reasoning verification, and using a shared pair-level advantage derived from both, allows a vision-language model to learn fine-grained, evidence-grounded reasoning in remote sensing scenes with limited supervision (1,008 demonstrations and 500 image pairs). The discovery is that this separation—rather than a single end-to-end reward—provides more targeted training signals: Gravity catches visual fabrications and omissions, while Flight assesses whether the described evidence supports the answer, even when a premise is false. Pairing images with fine-grained differences under a shared task and coupling policy updates through the shared advantage forces the model to attend to task-relevant visual changes, leading to denser, better-supported reasoning in shorter responses.","pith_inferences":["The verification design effectively acts as a learned reward model that is grounded in explicit rubrics, which may make the training more interpretable and less susceptible to reward hacking than a black-box scalar reward.","Paired training with a shared advantage implicitly encourages the model to compare scenes at the level of task-relevant evidence, a capability that could be directly tested on relational reasoning benchmarks outside the training domain.","The reliance on a single fixed verifier (DeepSeek-V4.1-Flash) raises a question of whether the gains would hold with other judges; a sensitivity study across verifiers would help establish robustness.","Because the advantage is normalized within each pair set, the method may be sensitive to the composition of the training units—units with many invariant or easily distinguished pairs could produce weaker training signals, potentially affecting convergence."],"forward_implications":["If the claim holds, remotely-sensed scene understanding can be specialized with far fewer annotated examples than typical instruction-tuning requires, because the paired design extracts more signal per example.","The separation of visual and reasoning verification offers a general template for improving accuracy and faithfulness in vision-language tasks beyond remote sensing, such as medical imaging or autonomous driving scene analysis.","Because the shared advantage couples updates from two responses, the model learns to detect task-relevant visual differences even when the scenes are visually distinct but semantically similar, likely sharpening fine-grained discriminative ability.","The urban specialization experiment (using the same rubric dimensions and reward weights) suggests the framework may transfer across scene types without task-specific engineering, reducing the cost of deploying VLMs in new domains."],"supporting_citations":[{"why":"Provides the CPRS harbor dataset used for training pairs and the baseline benchmarks (HarborEval, OpenEval) that define the main evaluation.","marker":"[13]"},{"why":"Supplies the GRPO algorithm that forms the optimization backbone for policy updates with group-relative advantage.","marker":"[36]"},{"why":"Specifies the Qwen3.5-4B backbone VLM that is adapted and serves as the base model for all comparisons.","marker":"[43]"},{"why":"Motivates the use of instance-specific rubrics as policy feedback, which is central to the Gravity and Flight verification design.","marker":"[8]"},{"why":"Defines the VRSBench benchmark from which a harbor subset is drawn for evaluation coverage.","marker":"[38]"},{"why":"Defines the RSVQA benchmark used to evaluate visual question answering accuracy on harbor and urban subsets.","marker":"[39]"},{"why":"Defines the VLRS-Bench benchmark for scene reasoning and decision evaluation, contributing one of the six evaluation sets.","marker":"[41]"}],"fun_headline_variants":["Decoupled checks boost fine-grained VLM reasoning 9.76 pts","Gravity and Flight: splitting visual & reasoning supervision","Two-view verification lifts remote sensing VLM scores","Pair-level advantage sharpens VLM scene judgments","Limited supervision, high gain: G&F lifts reasoning 9.76"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The verifier (DeepSeek-V4.1-Flash) can reliably score visual fidelity, coverage, evidence use, and paired adaptation according to the rubrics, so that the policy update receives accurate reward signals; the paper acknowledges that judging errors may be correlated and provides a reliability study showing agreement within one point of human ratings on 83.33–100% of items across dimensions, but with limited sample size (24 units).","fun_headline_variants_meta":{"raw":{"variants":["Decoupled checks boost fine-grained VLM reasoning 9.76 pts","Gravity and Flight: splitting visual & reasoning supervision","Two-view verification lifts remote sensing VLM scores","Pair-level advantage sharpens VLM scene judgments","Limited supervision, high gain: G&F lifts reasoning 9.76"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"cost":0.000171948,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000171948,"upstream_inference_prompt_cost":0.000105588,"upstream_inference_completions_cost":6.636e-05},"total_tokens":1196,"prompt_tokens":959,"completion_tokens":237,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":153}},"tokens_in":959,"tokens_out":237,"cost_source":"openrouter_usage","duration_ms":3655,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":153,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-13T15:23:08.660520+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"A controlled experiment where the same training pipeline is run with a different verifier (e.g., an open-source judge of comparable size or a smaller LM) would test whether the gains depend on the specific judge's quality; if performance drops substantially, the claim that the rubric verification itself is responsible is weakened. Alternatively, constructing test pairs where the verifier is known to misjudge the relation (e.g., ambiguous change types) could reveal brittleness in the shared-advantage mechanism.","supporting_citations":[{"cited_title":"VLRS-Bench: A Vision-Language Reasoning Benchmark for Remote Sensing","cited_arxiv_id":"2602.07045","evidence_quote":"Defines the VLRS-Bench benchmark for scene reasoning and decision evaluation, contributing one of the six evaluation sets."},{"cited_title":"Filling Before Advancing: Capability-Gap-Driven Post-Training for Scenario-Specialized Remote Sensing MLLMs","cited_arxiv_id":"2607.22205","evidence_quote":"Provides the CPRS harbor dataset used for training pairs and the baseline benchmarks (HarborEval, OpenEval) that define the main evaluation."},{"cited_title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","cited_arxiv_id":"2402.03300","evidence_quote":"Supplies the GRPO algorithm that forms the optimization backbone for policy updates with group-relative advantage."},{"cited_title":"Qwen3.5-4B Model Card","cited_arxiv_id":null,"evidence_quote":"Specifies the Qwen3.5-4B backbone VLM that is adapted and serves as the base model for all comparisons."},{"cited_title":"Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains","cited_arxiv_id":"2507.17746","evidence_quote":"Motivates the use of instance-specific rubrics as policy feedback, which is central to the Gravity and Flight verification design."},{"cited_title":"VRSBench: A Versatile Vision-Language Benchmark Dataset for Remote Sensing Image Understanding","cited_arxiv_id":null,"evidence_quote":"Defines the VRSBench benchmark from which a harbor subset is drawn for evaluation coverage."},{"cited_title":"RSVQA: Visual Question Answering for Remote Sensing Data","cited_arxiv_id":"2003.07333","evidence_quote":"Defines the RSVQA benchmark used to evaluate visual question answering accuracy on harbor and urban subsets."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}