{"id":"90644836-43dd-4ebe-8dc2-abd52cdeb7e8","arxiv_id":"2607.03644","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"MG-MAE pretrained on a new 28-channel lunar map outperforms ImageNet, vanilla MAE, and EO foundation models across six classification, regression, and segmentation tasks.","lead":"Moonstone releases the first 28-channel global multi-modal lunar dataset and MG-MAE, a modality-grouped masked autoencoder that handles missing instruments and uneven coverage. Pretrained features beat ImageNet, vanilla MAE, and Earth-observation models on six lunar science tasks, with data and code public.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper’s strongest claim is comparative and empirical, not a claim of absolute scientific ground truth. All six tasks use fixed, publicly released labels; the same labels are given to scratch, ImageNet, vanilla MAE, EO adapters, and MG-MAE. Consequently, even if majority-vote geology patches or interpolated LP-GRS fields contain scale or noise artifacts, those artifacts affect every method equally and cannot explain the large, consistent margins (+16.2 % geology, +14.7 mIoU craters, near-zero EO transfer). Ablations isolate the lunar-specific mechanisms (missing-modality key masking, coverage-adaptive ratios, spectral continuity) and geographic splits preserve relative gains, further insulating the ranking. The reader correctly flags the label construction as the softest point; it is a limitation of absolute utility, not of the comparative result that the paper actually asserts. No stronger load-bearing flaw (data leakage, non-reproducible splits, missing controls, etc.) appears in the manuscript or supplementary material. Therefore the ACCEPT / high-confidence verdict needs no adjustment.","tokens_in":22601,"tokens_out":531,"duration_ms":4977,"concrete_test":"Re-run the geology linear-probe and finetune evaluations after (a) discarding patches whose majority USGS unit covers <70 % of the 256\times256 window and (b) replacing LP-GRS targets with coarser 0.5° bin averages; if MG-MAE’s absolute accuracy/R^{2} drop by more than ~3–4 points while the gap versus vanilla MAE and scratch shrinks by more than half, the label-artifact concern would become load-bearing. Otherwise the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that MG-MAE domain-specific grouped pretraining yields large, consistent gains over scratch, ImageNet, vanilla MAE, and EO models on the released six-task suite—is well supported by the reported tables, ablations, geographic splits, few-shot results, single-modality baselines, and public artifacts. The reader’s weakest assumption (label noise / scale mismatch for majority-vote 49-class USGS units and ~60 km LP-GRS fields interpolated to 237 m) is real but ordinary for remote-sensing benchmarks and does not undermine the comparative claim: every baseline is evaluated on the identical labels, so relative margins remain informative even if absolute scientific utility is imperfect. No internal inconsistency or experimental flaw that would reverse the ranking was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper introduces Moonstone, the first multi-modal foundation-model benchmark for lunar remote sensing. It assembles a 28-channel global pretraining corpus at 128 ppd (~237 m) from seven instrument families across five missions, proposes MG-MAE (a modality-grouped masked autoencoder with per-group Conv tokenizers, a shared ViT-Base encoder, key-dimension attention masking for missing modalities, coverage-adaptive masking, complementary masking, proportional reconstruction weighting, InfoNCE alignment, and spectral continuity regularization on M3), and defines six downstream tasks (49-class geology, 5-class age, FeO/TiO2 regression, cross-modal thermal prediction, mare and crater segmentation). Across linear probe and finetune protocols, MG-MAE features outperform scratch, ImageNet ViT-B, vanilla MAE, single-modality, and EO-transfer baselines (SatMAE, Prithvi-EO-2.0, TerraMind with linear adapters), with largest gains on geology (+16.2% Acc) and craters (+14.7 mIoU). Ablations, geographic latitude-band splits, few-shot trials, and public data/code support the claims.","tokens_in":22857,"tokens_out":1484,"duration_ms":16543,"significance":"If the reported margins hold under independent re-runs, this is a substantial and timely contribution: it closes a clear gap relative to EO benchmarks (GEO-Bench, PANGAEA) and Mars-Bench by providing a reusable multi-modal lunar pretraining corpus, a domain-specific MAE design that addresses non-uniform instrument coverage, and a fixed six-task evaluation suite. Strengths that raise the bar include public HuggingFace/GitHub release of data, pipeline, and checkpoint; systematic ablations isolating missing-modality masking, grouping, coverage-adaptive masking, and spectral continuity; geographic splits; few-shot stds; and fair EO-transfer adapters that preserve pretrained patch embeddings. Domain-specific pretraining is shown to matter far more than EO transfer (+1.9% geology for the best EO model vs. +16.2% for MG-MAE), which is a useful negative result for planetary ML.","major_comments":[{"comment":"Table 3 reports single-point Acc/R2/mIoU for ImageNet, Scratch, Vanilla MAE, Linear, and Finetune with no multi-seed means or standard deviations (unlike Table 4 few-shot). The central claim of large, consistent margins (e.g., geology 40.1→56.3, craters 0.621→0.768) is load-bearing; without run-to-run variance or at least 3 seeds for the main table, it is hard to judge whether the ranking is stable under training noise. Please add multi-seed statistics for Table 3 (or a compact appendix table) for Scratch, Vanilla MAE, and MG-MAE linear/finetune on at least geology, composition, and craters.","section":null},{"comment":"§3.3 and the composition task definition: LP-GRS FeO/TiO2 have ~60 km effective resolution and are RBF-interpolated onto the 237 m grid, so patch-level R2 largely measures recovery of smooth regional fields rather than fine-scale geochemistry. The paper notes this briefly, but Table 3’s R2=0.924 (linear) is easy to over-interpret as high-resolution composition skill. Please state the resolution mismatch more prominently in the task definition and Discussion (and, if feasible, report a coarse-grid or low-pass evaluation) so absolute R2 is not read as pixel-scale mineral mapping performance. Relative ranking vs. baselines remains informative either way.","section":null},{"comment":"§5.6 / Table 7 EO transfer: the learned linear adapter that maps 28 lunar channels to each EO model’s expected input is a reasonable fairness choice, but the manuscript should state adapter training protocol (frozen vs. joint, data split, whether adapter is trained only on lunar pretraining crops or also on downstream labels) and confirm that EO models are not disadvantaged by an under-trained adapter. A short sensitivity check (e.g., adapter-only vs. full finetune already reported) would make the “EO transfer is near-scratch” claim more airtight.","section":null}],"minor_comments":[{"comment":"Fig. 1 caption and §4.1: complementary masking and coverage-adaptive ratios are described in text and Algorithm 1 (supp.), but the main figure’s “60–85%” range should explicitly cite Eq. (4) / mg = 0.75 + α(cg − c̄) so readers can reproduce the per-group visible-token counts in supp. Table 2.","section":null},{"comment":"Eq. (1)–(2): proportional weights wg = ℓg / Σℓg′ are self-balancing; clarify whether ℓg is the current mini-batch loss, EMA, or epoch-level average, and whether gradients flow through wg (stop-grad is standard for such schemes).","section":null},{"comment":"Table 8 ablation “No spectral continuity reg.” improves composition by +1.3 R2 points in the text but the table shows 0.911 vs 0.924; keep wording and table consistent and specify that SCR applies only to the 8 M3 bands.","section":null},{"comment":"§3.2 Mini-RF log(1+x) and Diviner channel naming (diviner_temp_night as rock-free regolith temperature) are important for reusers; consider a one-sentence pointer in the main text to the supp. quality-corrections subsection.","section":null},{"comment":"Typos/clarity: “Wedefinesix” / “WeproposeMG-MAE” spacing glitches in the introduction PDF; “Geographicgeneralization” and similar concatenated headings in Discussion; arXiv line-break artifacts in URLs.","section":null},{"comment":"Related work: Mars-Bench is cited; a sentence on how Moonstone’s multi-modal pretraining + fixed splits differs from evaluating EO/ImageNet models on Mars-only labels would help position the benchmark.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The contribution is real and the experimental package (data, code, ablations, geographic splits) is above average for a first planetary FM benchmark. I do not see a load-bearing flaw that would reverse the ranking of MG-MAE vs. scratch/ImageNet/vanilla MAE/EO. The composition-resolution caveat and missing multi-seed error bars on Table 3 are the only items I would insist on before camera-ready; both are fixable without new missions or redesign. Scope fits a CV / remote-sensing venue that accepts foundation-model + benchmark papers. No citation or novelty concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this is the first shared multi-modal foundation-model setup for the Moon, and the experiments actually show domain-specific pretraining beating ImageNet, vanilla MAE, and EO models by large margins on the same suite.\n\nWhat is new is the artifact set more than any single architectural trick. They assembled a 28-channel global map at ~237 m from seven instrument families, defined six fixed tasks (49-class geology, age, FeO/TiO2, cross-modal thermal, mare, craters), and released data plus code. MG-MAE’s useful pieces are the lunar ones: coverage-adaptive mask ratios, key-only attention masking for genuinely missing groups, and spectral continuity on M3. Grouped tokenizers and multi-modal MAE are inherited; the paper is clear about that.\n\nThe experimental design is the strong part. Ablations isolate missing-modality masking as the biggest hit, geographic latitude-band splits preserve relative gains, few-shot has error bars, single-modality baselines show fusion helps, and EO models (SatMAE, Prithvi, TerraMind) with linear adapters only move geology ~+2% over scratch while MG-MAE is +16%. That EO-transfer result is the most useful scientific claim in the paper. Task-specific U-Net/CNN and Lucey/RF comparisons are fair enough for the genre.\n\nSoft spots, in proportion: the geology labels are majority-vote USGS units on 256×256 patches, and composition is ~60 km LP-GRS interpolated onto a 237 m grid. That is real scale/noise mismatch, but every baseline is scored on the same labels, so the ranking still holds. Main Table 3 has no run-level error bars; hyperparameters (α, λ, μ, mask schedule, M3 band subset) are hand-chosen. Only ViT-Base. None of that reverses the comparative claim.\n\nThis is for planetary ML and anyone building Artemis-era mapping tools. Math is standard MAE/InfoNCE; citations cover EO FMs and lunar sources without weird gaps; data release is complete. I would send it to peer review, cite the benchmark and the EO-transfer negative result, and bring it to reading group if we care about planetary foundation models.","headline":"First multi-modal lunar FM benchmark with real domain-specific gains; engineering is careful, label caveats are ordinary, and EO transfer is honestly weak.","tokens_in":23419,"tokens_out":566,"would_cite":true,"duration_ms":11229,"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":"Domain-specific grouped pretraining on 28 lunar channels beats ImageNet, EO models, and vanilla MAE on six Moon tasks.","keywords":["foundation model","lunar remote sensing","multi-modal learning","masked autoencoder","benchmark","planetary science","modality grouping"],"falsifier":"Re-run the identical linear-probe and finetune protocol after replacing the USGS majority-vote geology labels and the interpolated LP-GRS composition targets with independent higher-resolution ground truth or expert-reviewed maps; if the large MG-MAE gains disappear or reverse, the central claim fails.","tokens_in":23478,"feed_emoji":"🌑","tokens_out":958,"duration_ms":7786,"temperature":0.7,"pith_summary":"Orbital missions have left the Moon with rich multi-modal data—optical, spectral, thermal, radar, gravity, composition—but the archives are fragmented and there has been no shared way to train or compare machine learning models on them. Moonstone supplies the missing pieces: a global 28-channel map at roughly 237 m resolution, a modality-grouped masked autoencoder (MG-MAE) built for incomplete planetary coverage, and six fixed downstream tasks spanning geology, age, composition, thermal cross-prediction, mare, and craters. The central claim is that pretraining with per-group tokenizers, missing-modality attention masks, coverage-adaptive masking, and spectral continuity yields features that beat training from scratch, ImageNet weights, Earth-observation foundation models, and a vanilla MAE on the same lunar data—often by double-digit margins. If the claim holds, lunar science and resource prospecting gain a reusable representation stack instead of one-off single-sensor models.","feed_headline":"Lunar pretraining beats ImageNet and Earth models on six Moon tasks","feed_subtitle":"28-channel MG-MAE lifts geology accuracy 16 points and crater mIoU 15 points over scratch","key_machinery":"MG-MAE: a modality-grouped masked autoencoder that uses one multi-channel convolutional tokenizer per physical instrument family, a shared Vision Transformer encoder with key-dimension attention masking for entirely missing groups, coverage-adaptive mask ratios, complementary anchor masking, cross-modal decoder attention, proportional reconstruction loss, InfoNCE alignment, and spectral-continuity regularization on the M3 reflectance bands.","core_discovery":"MG-MAE features pretrained on the assembled 28-channel lunar dataset outperform scratch, ImageNet-pretrained, vanilla MAE, and transferred Earth-observation foundation-model baselines on every one of six fixed downstream tasks, with the largest reported gains on 49-class geology classification and crater segmentation; Earth models transfer only marginally, so domain-specific grouped pretraining is required.","pith_inferences":["The same coverage-adaptive and ghost-token machinery should transfer to other airless bodies that also have sparse multi-mission archives (Mercury, asteroids) once comparable global stacks exist.","Because composition linear probing already exceeds finetuning, further gains may come more from better spectral regularization or multi-scale fusion than from simply scaling the ViT.","The near-zero transfer from Earth models suggests that any future multi-body foundation model will need explicit planetary domain adapters rather than pure continual pretraining on mixed EO–lunar data."],"forward_implications":["New instruments can be added as extra modality groups without redesigning the encoder, because missing-modality masking already handles heterogeneous coverage.","Polar and other under-covered regions become usable test beds rather than automatic out-of-distribution failures, since the model is trained to remain well-defined when groups contribute zero tokens.","Few-shot and low-label regimes for geology and age improve substantially, lowering the barrier for resource-prospecting maps that currently rest on a handful of sample-return sites.","Task-specific lunar models that rely on hand-crafted spectral indices or single-sensor U-Nets are dominated by the multi-modal pretrained features on the shared benchmark."],"fun_headline_variants":["MG-MAE lunar features beat ImageNet and Earth models on six tasks","28-channel Moon pretraining tops scratch and transfer baselines","Domain lunar pretraining needed as Earth models transfer poorly","Moonstone MG-MAE leads on geology classification and crater mIoU","First lunar multimodal benchmark shows grouped MAE gains over MAE"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The six constructed labels—especially majority-vote 49-class geologic units on 256-pixel patches and composition targets interpolated from roughly 60 km native resolution onto a 237 m grid—are accurate and scale-matched enough that the measured gains reflect real scientific utility rather than label artifacts.","fun_headline_variants_meta":{"raw":{"variants":["MG-MAE lunar features beat ImageNet and Earth models on six tasks","28-channel Moon pretraining tops scratch and transfer baselines","Domain lunar pretraining needed as Earth models transfer poorly","Moonstone MG-MAE leads on geology classification and crater mIoU","First lunar multimodal benchmark shows grouped MAE gains over MAE"]},"model":"grok-4.5","effort":"low","cost_usd":0.003122,"raw_usage":{"total_tokens":1095,"prompt_tokens":766,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":31220000,"prompt_tokens_details":{"text_tokens":766,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":261,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":766,"tokens_out":68,"duration_ms":2634,"temperature":1.0,"reasoning_tokens":261,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:58:07.547754+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the identical linear-probe and finetune protocol after replacing the USGS majority-vote geology labels and the interpolated LP-GRS composition targets with independent higher-resolution ground truth or expert-reviewed maps; if the large MG-MAE gains disappear or reverse, the central claim fails.","supporting_citations":[],"review_version":1}