{"id":"b9bd95f6-668b-4f8f-9ad9-4e5d72f75f20","arxiv_id":"2604.09787","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Overlapping multi-instrument galaxy images plus dual encoders and flow-matching counterfactual generation yield physics latents unconfounded by sensor artifacts.","lead":"A dual-encoder model trained on overlapping telescope images learns to separate a galaxy's true appearance from each instrument's noise and blur. That split enables cleaner property estimates, cross-survey image translation, and artifact-free rare-object search.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Spatial-neighbor instrument conditioning is the load-bearing premise; residual leakage and imperfect pair alignment leave it incompletely stress-tested.","rationale":"The Reader correctly isolates the spatial-neighbor instrument conditioning as the weakest load-bearing assumption. The dual-encoder + counterfactual flow-matching construction is sound given that premise, and the paper supplies multiple corroborating checks (UMAP separation, outlier morphology, cross-survey retrieval, zero-shot ellipticity transfer, progressive SNR traversal). Residual leakage is acknowledged by the authors and partially bounded by the cross-predict baseline, so the concern does not overturn the contribution; it keeps the verdict at CONDITIONAL pending a direct ablation of neighbor quality and fuller quantification of leakage. No stronger internal inconsistency or formal error is present. Agreement with the Reader is therefore full; the recommended verdict remains CONDITIONAL.","tokens_in":20759,"tokens_out":552,"duration_ms":6061,"concrete_test":"Re-train (or ablate at inference) with instrument conditioning replaced by (a) random same-survey images and (b) neighbors drawn from a larger angular radius (e.g. 30–60 arcmin) or deliberately mismatched depth/PSF bins; recompute the physics-vs-instrument R^{2} matrix of Fig. 4 / Table 5 and the UMAP pair-alignment of Fig. 2. If physics-property R^{2} from the instrument latent rises substantially or pair alignment collapses, the spatial-neighbor premise is load-bearing and currently under-supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the instrument encoder, conditioned only on up to five nearest spatial neighbors within 3 arcmin of the same survey (§3 Data Triplets; Application), isolates measurement artifacts without private physics of the target. The paper motivates this by smooth variation of PSF/depth, and Fig. 2 + the R^{2} asymmetry in Fig. 4 / Table 5 are consistent with partial success. However, the same tables show non-trivial residual predictive power of the instrument latent for redshift, stellar mass and morphology, and of the physics latent for some instrument quantities (though below the cross-predict ResNet baseline). If local systematics are not sufficiently shared—or if the 3-arcmin neighbors systematically co-vary with large-scale structure or selection—the architectural bottleneck is incomplete and the claimed unconfounded physics representations are only approximate. The generative objective and dual-encoder design do not by themselves guarantee the separation once the conditioning assumption is violated; the empirical checks are supportive but do not quantify how much residual leakage remains under controlled degradation of neighbor quality.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a dual-encoder architecture with a counterfactual flow-matching objective that uses overlapping multi-instrument observations to disentangle physics-invariant signals from instrument-specific artifacts. Training triplets consist of an anchor image (never seen by the encoders), a same-source different-instrument view for the physics encoder, and same-instrument different-source neighbors for the instrument encoder; the decoder reconstructs the anchor conditioned on both latents (Eq. 1, Fig. 1). On ~100k cross-matched Legacy–HSC galaxy images the method produces UMAP-separated latent spaces with pair alignment (Fig. 2), physics-space outliers free of instrumental pathologies (Fig. 3), asymmetric R^{2} probes against AION-1 and random-ResNet baselines (Fig. 4 / Table 5), instrument-independent nearest-neighbor retrieval (Fig. 5), and usable cross-instrument posterior samples (Figs. 6–7, MSE and pipeline-transfer checks). The authors present the approach as a general recipe for scientific multi-modal pretraining that treats sensor effects as natural augmentations.","tokens_in":21055,"tokens_out":1099,"duration_ms":10141,"significance":"If the disentanglement holds at the claimed level, the work supplies a practical architectural recipe for multi-instrument foundation models that avoids hand-engineered contrastive losses and explicit instrument forward models. The generative decoder additionally yields a data-driven noise model and counterfactual survey translation, both of direct utility for follow-up prioritization and pipeline transfer across surveys. Strengths that raise the contribution above a pure methods paper include the external-catalog probes (PROVABGS spectroscopy, shape catalogs, survey metadata), the cross-predict ResNet control that bounds shared spatial structure, the zero-shot ellipticity pipeline transfer (R^{2} 0.82 vs 0.81), and the public simplified codebase. Residual leakage is acknowledged and partially quantified; the result is therefore a solid, usable advance for astro-ph.IM and related multi-sensor domains rather than a complete solution to causal disentanglement.","major_comments":[{"comment":"§3 Data Triplets / Application: the instrument encoder is conditioned on up to five nearest spatial neighbors within 3 arcmin, under the premise that PSF/depth vary smoothly enough to share essentially the same instrument conditions. Residual R^{2} of the instrument latent for redshift, stellar mass and morphology (Fig. 4 / Table 5) and of the physics latent for some instrument quantities leave open how much private physics or large-scale structure co-varies with the neighbors. A controlled ablation that degrades neighbor quality (random same-survey images, larger angular radii, or shuffled sky positions) and re-measures the R^{2} asymmetry and pair alignment would quantify the load-bearing assumption; without it the claim of unconfounded physics representations remains only partially stress-tested.","section":null},{"comment":"§4.1 Parameter inference and Appendix A.5: the cross-predict ResNet baseline shows that physics latents do not capture private instrument information above the shared spatial floor, which is a useful control. However, the instrument latent still retains non-trivial predictive power for physical properties that are only partially recoverable from imaging. The paper should either tighten the instrument bottleneck (capacity study already flagged as future work) or report a quantitative leakage metric (e.g., mutual information or residual R^{2} after regressing out the cross-predict baseline) so that the degree of residual confounding is stated rather than left as a qualitative caveat.","section":null}],"minor_comments":[{"comment":"Fig. 4 and Table 5: the main-text figure is hard to read at print scale; the appendix table is clearer. Consider moving the full numerical table into the main text or enlarging the figure panels.","section":null},{"comment":"§4 Data Preprocessing: the decision to discard the HSC y-band and to upsample Legacy 31×31 to 48×48 via linear interpolation is stated but not ablated; a short note on sensitivity would help reproducibility.","section":null},{"comment":"Eq. (1) and surrounding text: the notation for the variable-length sets of physics and instrument embeddings is dense; a short algorithmic box listing the sampling of j, k, neighbors and the attention conditioning would improve clarity.","section":null},{"comment":"Fig. 3 caption and text: the full 160×160 cutouts are useful, but the main-text figure is small; consider a supplementary gallery of the top outliers with survey labels.","section":null},{"comment":"Related work: the comparison to Audenaert et al. (contrastive dual-encoder) is clear; a one-sentence quantitative note on why the generative objective preserves high-resolution features that contrastive alignment discards would strengthen the positioning.","section":null},{"comment":"Code availability: the simplified public codebase is welcome; stating the exact commit or DOI used for the reported numbers would aid exact reproduction.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The spatial-neighbor assumption is the only genuinely load-bearing soft spot; once the authors add a short ablation or a clearer leakage metric the paper is ready. Fit for a methods-oriented astrophysics or ML-for-science venue is good; novelty relative to the concurrent Audenaert contrastive work is adequately differentiated by the generative objective."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean methods paper that takes the dual-encoder causal idea from Audenaert et al. and replaces contrastive alignment with a counterfactual flow-matching objective. That change is the real novelty: it keeps fine-grained features that contrastive methods tend to discard, and it gives you actual cross-instrument generation for free. On ~100k real Legacy–HSC pairs they show the expected UMAP separation, physics-space outliers free of the usual instrument junk, instrument-independent retrieval (including strong-lens neighbors), and usable counterfactual images that even transfer an ellipticity pipeline almost unchanged (R^{2} 0.82 vs 0.81).\n\nThe empirical checks are better than average for this genre. They probe both latents against PROVABGS spectroscopy, shape catalogs, and survey metadata, include a random ResNet baseline and a cross-predict ResNet control, and report the asymmetry you want to see. The generative diagnostics (PSD, autocorrelation, Z-score calibration) are also there. Code is partially released; that helps.\n\nSoft spots are real but proportionate. The instrument encoder is conditioned on five spatial neighbors within 3 arcmin; that is load-bearing and only motivated by smooth PSF/depth variation. Residual R^{2} of the instrument latent for redshift/mass/morphology (and some physics-to-instrument leakage, though below the cross-predict floor) shows the separation is incomplete. Counterfactuals are informed predictions, not true deeper photons—they say so, but the “survey simulator” language still needs care. Single survey pair and free parameters (token geometry, neighbor radius, preprocessing) are the usual limits.\n\nThis is for people building multi-survey foundation models or needing unconfounded retrieval/generation. It deserves a serious referee; the core claim holds as an approximate, useful architectural recipe. I would engage, cite the method when I need cross-instrument generation, and push for fuller leakage quantification and code.","headline":"Solid generative dual-encoder extension of multi-instrument disentanglement on real Legacy–HSC pairs; useful and mostly well-checked, with residual leakage and neighbor conditioning as the main soft spots.","tokens_in":21679,"tokens_out":490,"would_cite":true,"duration_ms":6074,"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":"Overlapping multi-telescope observations can train dual encoders that separate a galaxy’s true physical signal from instrument-specific distortions.","keywords":["disentangled representation learning","multi-instrument astronomy","counterfactual generation","flow matching","galaxy imaging","instrument systematics","self-supervised pretraining"],"falsifier":"Replace the five nearest spatial neighbors with random galaxies from the same survey; if instrument-property prediction from the instrument latent collapses and physics-space outliers again become dominated by saturated pixels and noise bands, the shared-local-systematics premise has failed.","tokens_in":21635,"feed_emoji":"🔭","tokens_out":881,"duration_ms":19670,"temperature":0.7,"pith_summary":"Physical measurements always mix the signal of interest with sensor-dependent artifacts that confound inference and block clean multi-instrument analysis. This paper argues that when the same systems are observed by more than one instrument, those overlaps are enough to learn the split without an explicit instrument model. A dual-encoder architecture is trained only to reconstruct an unseen anchor image: one encoder sees the same source under a different instrument, the other sees different sources under the target instrument, and a flow-matching decoder combines the two latents. On roughly 100,000 cross-matched galaxy images from two major surveys, the resulting physics space supports property inference, rare-object search, and similarity retrieval that are not dominated by sensor pathologies, while the instrument space captures observing conditions and enables counterfactual cross-survey generation. The method is offered as a general self-supervised recipe for multi-sensor science: treat instrument differences as natural augmentations and force invariance through counterfactual reconstruction.","feed_headline":"Dual encoders split galaxy physics from telescope artifacts","feed_subtitle":"Overlapping survey pairs train a model that rebuilds unseen views and yields instrument-free embeddings for search and inference.","key_machinery":"Counterfactual generation with dual encoders: the physics encoder is given same-source, different-instrument views; the instrument encoder is given different-source, same-instrument views; a conditional flow-matching decoder reconstructs an anchor image neither encoder has seen, creating an architectural information bottleneck that forces the separation.","core_discovery":"A dual-encoder flow-matching model trained solely with a counterfactual reconstruction objective on overlapping multi-instrument observations learns representations that explicitly separate intrinsic physical signals from sensor-specific distortions and noise. Demonstrated on paired DESI Legacy and Hyper Suprime-Cam galaxy images, the physics latents align matched sources across surveys and yield morphologically meaningful outliers, while the instrument latents cluster by survey and predict local observing conditions, enabling unconfounded inference, instrument-independent search, and cross-instrument generation.","pith_inferences":["Tightening or widening the instrument-encoder bottleneck is a practical dial for residual leakage of physical properties into the instrument space.","Outside dense overlapping footprints the method needs an unpaired extension; otherwise coverage remains limited to shared sky regions.","The same structural split could reduce instrument-driven false positives in multi-mission time-series archives where detector systematics currently dominate anomaly lists."],"forward_implications":["Physics embeddings support outlier detection of rare morphologies without instrument artifacts dominating the ranking.","Counterfactual generation can predict how a low-resolution survey object would appear under a deeper, higher-resolution instrument to prioritize follow-up.","Nearest-neighbor search in physics space retrieves physically similar objects independent of which survey supplied the query.","Analysis pipelines trained on one survey’s images can be applied, zero-shot, to counterfactual images generated from another survey.","The same training recipe—overlap pairs, dual encoders, counterfactual reconstruction—extends to other multi-sensor scientific domains."],"fun_headline_variants":["Dual encoders isolate galaxy physics from telescope artifacts","Counterfactual pairs separate intrinsic signals from sensor noise","Overlapping surveys train instrument-free galaxy embeddings","Model disentangles multi-sensor artifacts via dual-encoder flows","Physics latents align galaxies across surveys without confounds"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Images of other galaxies lying within a few arcminutes of the target are assumed to share essentially the same instrument conditions, so they can teach the model the sensor’s behavior without leaking the target’s own physics.","fun_headline_variants_meta":{"raw":{"variants":["Dual encoders isolate galaxy physics from telescope artifacts","Counterfactual pairs separate intrinsic signals from sensor noise","Overlapping surveys train instrument-free galaxy embeddings","Model disentangles multi-sensor artifacts via dual-encoder flows","Physics latents align galaxies across surveys without confounds"]},"model":"grok-4.5","effort":"low","cost_usd":0.004072,"raw_usage":{"total_tokens":1278,"prompt_tokens":803,"num_sources_used":0,"completion_tokens":77,"cost_in_usd_ticks":40720000,"prompt_tokens_details":{"text_tokens":803,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":398,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":803,"tokens_out":77,"duration_ms":4212,"temperature":1.0,"reasoning_tokens":398,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T23:04:36.036437+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Replace the five nearest spatial neighbors with random galaxies from the same survey; if instrument-property prediction from the instrument latent collapses and physics-space outliers again become dominated by saturated pixels and noise bands, the shared-local-systematics premise has failed.","supporting_citations":[],"review_version":2}