Pith. sign in

REVIEW 2 major objections 4 minor 36 references

This paper claims that a general-purpose vision-language model can act as a weak teacher for an astronomy foundation model, improving galaxy morphology classification under limited human labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 09:40 UTC pith:OCFSV5YA

load-bearing objection A legitimate label-efficient distillation result with a useful new benchmark, but the test-time prior correction looks tuned on the evaluation split and the cross-survey schema mapping is unvalidated; both are fixable and worth refereeing. the 2 major comments →

arxiv 2608.02300 v1 pith:OCFSV5YA submitted 2026-08-03 cs.CV

A General-Purpose VLM Can Teach an Astronomy Foundation Model to Better Recognize Galaxy Morphology

Dichang Zhang , Jiaqi Deng , Yixuan Shao , Yuanpeng Liu , Jiali Cui , Zhiqiang Lao , Heather Yu , Liang Peng
show 2 more authors
Simon Birrer Dimitris Samaras
This is my paper
classification cs.CV
keywords galaxy morphologyvision-language modelvisual question answeringknowledge distillationsample selectionfoundation modellabel-efficient learningastronomical surveys
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Crowdsourced labels from Galaxy Zoo cannot scale to the billions of galaxies that LSST and Roman will image, and adapting a pretrained astronomy model like Zoobot to a new survey still demands many human annotations. This paper tries to establish that a general-purpose vision-language model (Qwen) carries enough visual-semantic priors about galaxy shapes to act as a weak teacher: its soft answers, distilled into a frozen Zoobot encoder through a side adapter, improve morphology classification under small human-label budgets. The authors first benchmark three VLMs on a survey-oriented VQA test and find them far below human accuracy (~52–57% overall vs ~69–73% human) but with informative confidence. They then show that distilling those imperfect predictions, selecting galaxies for human labeling with tree-aware uncertainty/diversity, and correcting answer priors at test time yields consistent improvements over no-teacher baselines across Cosmic Dawn and Euclid at 512, 2048, and 8192 labels. A sympathetic reader would take the central claim to be that weak, unreliable VLM predictions are still a useful complementary source of supervision for a domain-specific foundation model. The authors note their two test domains only approximate LSST/Roman conditions and that generalization to future surveys with finer question trees is not guaranteed.

Core claim

On the paper's own terms, the central discovery is that Qwen3.5-9B, a general-purpose VLM never trained for astronomy, produces morphology predictions that are wrong often enough to disqualify it as a standalone annotator but right often enough — and calibrated enough — to serve as a teacher. Zero-shot overall accuracy is 52.19% on Cosmic Dawn and 57.36% on Euclid against a ~34% random baseline, while humans reach 73.17% and 68.96%. When Qwen's per-question soft distributions are distilled into Zoobot with reachability weights (so downstream questions contribute in proportion to the teacher's belief their parent path is active), a frozen encoder plus a zero-initialized residual side adapter,

What carries the argument

The load-bearing machinery is reachability-aware distilled side tuning combined with tree-probability sample selection. The teacher's answer distribution for each question is weighted by reachability ρ = product of the teacher's probabilities for all parent answers along the Galaxy Zoo decision tree, so the teacher can softly route downstream questions instead of committing to a hard path. The student is Zoobot with a frozen encoder and a trainable bottleneck side adapter (zero-initialized output), so the VLM's update is absorbed residually without overwriting the pretrained morphology representation. Sample selection scales a per-galaxy vector of tree-probability features by the teacher's m

Load-bearing premise

The load-bearing premise is that the hand-built mapping of Cosmic Dawn and Euclid survey-specific morphology columns onto a single canonical 13-question Galaxy Zoo schema preserves the meaning of each question and the vote-fraction distributions; the paper gives no validation of that mapping, and if it introduces systematic misalignment, the apparent transfer gains could be an artifact of the standardization rather than of VLM knowledge.

What would settle it

Re-run the full pipeline on a third Galaxy Zoo-style dataset using its native question tree with no schema mapping; if the VLM-teacher advantage over a no-teacher baseline shrinks or disappears, the reported transfer gains depend on the canonical mapping rather than on complementary VLM knowledge. A cheaper check: have astronomy experts judge semantic equivalence of each mapped question across Cosmic Dawn and Euclid and recompute balanced accuracy only on questions where the mapping is declared exact.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, adapting a morphology foundation model to a new survey can be driven by a large pool of unlabeled images plus a few hundred to a few thousand human labels, with the VLM supplying the bulk of the supervision.
  • Improvements concentrate on balanced and weak-branch accuracy (spiral arm count, bulge size, artifact type), which are exactly the categories most valuable for galaxy-evolution science and most expensive to label.
  • The released benchmark, precomputed VLM logits, and pipeline code make the teacher's inference a reusable public asset, so subsequent groups can train students without rerunning large-model VQA.
  • Because even VLM predictions only modestly above random improve an already strong domain model, the method is forgiving of teacher quality and may remain useful as VLMs continue to improve.
  • The framework positions itself as a practical component for LSST and Roman analytics, where full human inspection of billions of galaxies is impossible.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A gap the paper leaves open is validation of the canonical 13-question schema used to merge Cosmic Dawn and Euclid columns; the authors do not demonstrate that the mapping preserves each question's semantics or vote-fraction meaning, so the headline gains should be rechecked per question before deployment (editorial caution, not a paper claim).
  • The same recipe — soft VLM teacher, tree-structured reachability routing, uncertainty-diversity acquisition — should transfer to other domains with branching label protocols, such as medical image triage or remote-sensing damage assessment; a cheap test would rerun the pipeline on an existing decision-tree dataset outside astronomy.
  • The test-time prior correction has a strength parameter β and contributes to the gains; the paper does not isolate how much of the weak-branch improvement comes from rebalancing answer priors versus from VLM knowledge, so a follow-up should report scores at β=0 and at the best β separately for each survey.
  • All experiments use a single fixed annotation budget selected in one shot; multi-round active learning, where the distilled student's own uncertainty requests the next batch of labels, is a natural extension the paper does not test and could push gains further.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes using a general-purpose VLM (Qwen) as a weak teacher for Zoobot, an astronomy foundation model, to improve galaxy morphology classification under limited human-label budgets. It introduces a survey-oriented VQA benchmark over Cosmic Dawn and Euclid, evaluates three open VLMs, and finds that they are above random but below human performance, with informative confidence estimates. The method then distills the VLM's soft predictions into a frozen Zoobot encoder through a side adapter with reachability-weighted cross entropy, selects galaxies for human annotation using a tree-aware uncertainty-and-diversity criterion, fine-tunes on the acquired labels, and applies a test-time prior-correction readout. The paper reports consistent gains across two surveys and three annotation budgets, with ablations for distillation components, selection rules, and correction strength.

Significance. If the reported results hold under a clean evaluation protocol, the contribution is practically valuable: it demonstrates that a general-purpose VLM carries visual-semantic priors complementary to a domain-specific astronomy foundation model, and it provides a reproducible pipeline with a public benchmark, precomputed logits, and code. The three-seed reporting, component ablations, and the release of inference artifacts are strengths. However, the headline comparison is currently undermined by an apparent test-set fitting of the prior-correction strength, so the central quantitative claim needs verification before the result can be accepted at face value.

major comments (2)
  1. [Section 4.4, Eqs. (12)–(14), Figure 5, Table 4] The test-time prior correction appears to be fitted to the evaluation set. The option prior pi_q is estimated on E_q, the same evaluation examples used to compute the accuracies in Table 4, and Figure 5 sweeps beta with no statement of how beta was chosen for the main table. If the reported 'Ours' rows use the beta that maximizes accuracy on the test set, while the 'No distil + random selection' baselines are decoded with beta=0 (the paper does not state that the same corrected readout is applied to all methods), then the headline gaps (e.g., 66.31 vs 60.30 on Cosmic Dawn at budget 512; 85.48 vs 81.19 on Euclid at budget 512) partly reflect a test-set-fitted scalar rather than the VLM teacher. This directly affects the central quantitative claim. Please fix beta on a validation split and report its value, apply the identical corrected readout to every baseline, and also report a beta=0 c
  2. [Appendix A, Eqs. (15)–(17)] The mapping of survey-specific morphology columns to a canonical 13-question schema is not validated. Because the VQA prompts, distillation targets, and human-label fine-tuning all operate on this standardized schema, a systematic misalignment between Cosmic Dawn/Euclid semantics and the canonical questions would propagate into every comparison and could produce apparent transfer gains that are an artifact of the standardization rather than of VLM knowledge. The paper should provide evidence that the mapping preserves question semantics and vote-fraction meanings, for example by reporting per-question vote-fraction concordance on objects with overlapping annotations, or an expert-reviewed mapping table stating whether any questions were merged or split.
minor comments (4)
  1. [Section 3.2 and References] Table 1 and the text refer to Qwen3.5-9B, InternVL3.5-8B Flash, and Llama3.2-11B Vision, but the cited references [7], [12], and [29] describe earlier model versions. Please update the citations or the model names so the evaluated checkpoints are unambiguous.
  2. [Section 5.2, Table 6] The sentence 'Either diversity or uncertainty only selection is worth than them combined' is not supported by the 512-label row, where Uncertainty-only shows +0.81 Overall and +2.55 Weak relative to the combined criterion. Please qualify or correct the statement.
  3. [Table 4] The 'Finetune VLM + VLM selection' rows are reported as single values without standard deviations, while all other rows show three-seed means; please state the number of seeds and whether the single value is the mean over the same seeds.
  4. [Figure 5] The figure does not state whether the panels are for Cosmic Dawn, Euclid, or a pooled setting, nor whether the bottom-row baseline uses the same beta as 'Ours'. Please label the panels and clarify the comparison.

Circularity Check

1 steps flagged

Main gains partly rest on a test-set-fitted prior correction, but the central VLM-teacher claim also has beta=0 and baseline support.

specific steps
  1. fitted input called prediction [Section 4.4, Eqs. (12)–(14); Figure 5; Table 4]
    "For question q, let E_q be the set of evaluation examples on which question q is scored. ... For each option c in C_q, we first estimate the model-implied marginal option probability using the student predictions: π̃_{q,c} = max(10^{-3}, |E_q|^{-1} ∑_{i∈E_q} s_{θ,q,c}(x_i)) ... The final corrected readout is: ŷ^{(β)}_{iq} = arg max_{c∈C_q} s_{θ,q,c}(x_i) / π̂^β_{q,c}. Figure 5 studies the correction strength β. Mild correction generally improves all accuracy."

    The corrected readout used to compute every reported accuracy is defined using statistics over the same evaluation examples E_q on which accuracy is measured (Eq. 12), and Eq. 14 applies that set-dependent correction to each prediction. Figure 5 sweeps the correction strength β over those evaluation results, and the paper does not state whether β was chosen on a held-out validation split or whether the Table 4 baselines received the identical corrected readout. Therefore the headline gains are partly a self-fitted function of the evaluation set rather than a fixed out-of-sample predictor; the reported “prediction” is not independent of the data used to construct the readout.

full rationale

The central claim—that a general VLM teacher improves Zoobot under limited human labels—is not definitionally circular: the teacher and student are distinct models, distillation is trained on unlabeled pools, sample selection is evaluated against random selection, and the comparison to no-distillation baselines is empirical. There is no load-bearing self-citation chain: the cited Zoobot and Galaxy Zoo works are external, and no uniqueness theorem from the present authors is invoked. The one circularity-adjacent element is the test-time prior correction. Eq. (12) estimates the correction prior on E_q, the same evaluation set used to compute the reported accuracies, and Eq. (14) makes the final prediction depend on that set; Figure 5 sweeps β without disclosing a validation-based selection protocol or whether Table 4 baselines receive the same corrected readout. This makes part of the reported margin a self-fitted function of the test set. However, the paper itself claims the improvement over the baseline persists “across all correction strength,” and the β=0 (uncorrected) comparison provides an internal control, so the central teacher-improvement result is not fully reduced to the test-set fit. The Appendix A schema mapping is unvalidated but is a domain-construction issue, not a circularity. Score 3 reflects one partially self-fitted component while the core comparison retains independent empirical content.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

No new physical entities. The only invented constructs are the tree-probability feature phi_i and tree-aware uncertainty z_i — these are heuristic feature designs, not entities with independent falsifiable handles. The paper's dependency inventory is dominated by the shared-schema mapping, the use of Galaxy Zoo labels as ground truth, and the test-time prior correction using evaluation-set statistics.

free parameters (1)
  • beta (test-time prior-correction strength) = not stated; ablated in Fig. 5
    Controls how strongly the student's option prior, estimated on evaluation examples, penalizes frequent answers. The paper does not specify a validation split for choosing beta; Fig. 5 suggests performance varies by beta, so reported Table 4 numbers may use a test-tuned beta.
axioms (4)
  • domain assumption Galaxy Zoo vote fractions are reliable ground truth for morphology
    All evaluation, distillation targets, and finetuning labels use aggregated volunteer vote fractions. Any systematic volunteer bias propagates.
  • domain assumption Survey-specific morphology schemas can be faithfully mapped to a shared 13-question canonical tree
    Appendix A maps Cosmic Dawn and Euclid columns to a canonical schema; no validation of mapping correctness or inter-survey semantic equivalence is provided.
  • domain assumption Cosmic Dawn and Euclid are representative proxies for LSST and Roman
    Authors' own limitation (Section 7) states they do not fully reproduce future survey characteristics; the conclusion in Section 6 nevertheless treats them as proxies.
  • standard math Softmax over letter-token logits gives a usable option distribution
    Used everywhere to derive confidence and teacher targets; calibration is only empirically checked with reliability-style curves (Fig. 3), not guaranteed.

pith-pipeline@v1.3.0-daily-deepseek · 20222 in / 9687 out tokens · 81032 ms · 2026-08-04T09:40:27.256475+00:00 · methodology

0 comments
read the original abstract

Existing astronomy foundation models provide strong galaxy representations, but adapting them to new survey conditions and survey-specific morphology recognition tasks still requires substantial human supervision. We show that VLM-based VQA systems contain meaningful visual-semantic priors that can serve as weak supervision for downstream morphology classifiers and improve morphology classification under limited human-label budgets. We first introduce a survey-oriented VQA benchmark spanning two representative imaging regimes and evaluate state-of-the-art VLMs on galaxy morphology questions. The results show that these models capture useful morphology signals and informative uncertainty, but are not sufficiently reliable to replace human annotators. Motivated by this finding, we use a general-purpose VLM as a morphology teacher for Zoobot, an astronomy foundation model pretrained on large-scale Galaxy Zoo annotations. Across two survey domains and multiple annotation budgets, the VLM teacher consistently improves Zoobot's downstream morphology classification. These results demonstrate that a general-purpose VLM provides knowledge complementary to an astronomy foundation model and can teach it to better recognize galaxy morphology under limited human supervision. The resulting pipeline is designed for label-efficient adaptation to forthcoming large-scale surveys, including the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) and the Nancy Grace Roman Space Telescope. The benchmark and code are publicly available at https://github.com/fw-ic/VLM-morphology-teacher.

Figures

Figures reproduced from arXiv: 2608.02300 by Dichang Zhang, Dimitris Samaras, Heather Yu, Jiali Cui, Jiaqi Deng, Liang Peng, Simon Birrer, Yixuan Shao, Yuanpeng Liu, Zhiqiang Lao.

Figure 1
Figure 1. Figure 1: Overview of the proposed VLM-teacher framework for label-efficient galaxy morphology classification. A general [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Euclid and Cosmic Dawn examples with their paths [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Confidence–performance relations. Each point [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the VLM selection space for Cos [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prior correction ablation. The top row compares [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 1 canonical work pages

  1. [1]

    Hiroaki et. al. Aihara. 2018. The Hyper Suprime-Cam SSP Survey: Overview and Survey Design.Publications of the Astronomical Society of Japan70, SP1 (2018), S4. doi:10.1093/pasj/psx066

    doi:10.1093/pasj/psx066 2018

  2. [2]

    Rachel Akeson, Lee Armus, Etienne Bachelet, Vanessa Bailey, Lisa Bartusek, Andrea Bellini, Dominic Benford, David Bennett, Aparna Bhattacharya, Rachel Bohler, et al. 2019. The Wide Field Infrared Survey Telescope: 100 Hubbles for the 2020s.arXiv e-prints(2019). arXiv:1902.05569 [astro-ph.IM]

    Pith/arXiv arXiv 2019

  3. [3]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al

  4. [4]

    Lawrence Zitnick, and Devi Parikh

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. 2015. VQA: Visual Question Answering. InProceedings of the IEEE International Conference on Computer Vision (ICCV). 2425–2433

    2015

  5. [5]

    David Arthur and Sergei Vassilvitskii. 2007. 𝑘-means++: The Advantages of Careful Seeding. InProceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms. 1027–1035

    2007

  6. [6]

    Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal

    Jordan T. Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. 2020. Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds. InInternational Conference on Learning Representations (ICLR)

    2020

  7. [7]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2024. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (C...

    2024

  8. [8]

    Schlegel, Dustin Lang, Robert Blum, Kaylan Burleigh, Xiaohui Fan, Joseph R

    Arjun Dey, David J. Schlegel, Dustin Lang, Robert Blum, Kaylan Burleigh, Xiaohui Fan, Joseph R. Findlay, Doug Finkbeiner, David Herrera, Stéphanie Juneau, Martin Landriau, Michael Levi, Ian McGreer, Aaron Meisner, Adam D. Myers, John Moustakas, Peter Nugent, Anna Patej, Edward F. Schlafly, Alistair R. Walker, Francisco Valdes, Benjamin A. Weaver, Christop...

  9. [9]

    Willett, and Joni Dambre

    Sander Dieleman, Kyle W. Willett, and Joni Dambre. 2015. Rotation-invariant con- volutional neural networks for galaxy morphology prediction.Monthly Notices of the Royal Astronomical Society450, 2 (2015), 1441–1459. arXiv:1503.07077 [astro- ph.IM] doi:10.1093/mnras/stv632

    Pith/arXiv arXiv 2015

  10. [10]

    Domínguez Sánchez, M

    H. Domínguez Sánchez, M. Huertas-Company, M. Bernardi, D. Tuccillo, and J. L. Fischer. 2018. Improving galaxy morphologies for SDSS with Deep Learn- ing.Monthly Notices of the Royal Astronomical Society476, 3 (2018), 3661–3676. arXiv:1711.05744 [astro-ph.GA] doi:10.1093/mnras/sty338

    Pith/arXiv arXiv 2018

  11. [11]

    Euclid Collaboration, Mike Walmsley, et al. 2025. Euclid Quick Data Release (Q1): First visual morphology catalogue. arXiv:2503.15310 [astro-ph.GA]

    Pith/arXiv arXiv 2025

  12. [12]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. 2024. The Llama 3 Herd of Models.arXiv preprint arXiv:2407.21783(2024). arXiv:2407.21783 [cs.AI]

    Pith/arXiv arXiv 2024

  13. [13]

    Md Abul Hayat, George Stein, Peter Harrington, Zarija Lukić, and Mustafa Mustafa. 2021. Self-supervised Representation Learning for Astronomical Images. The Astrophysical Journal Letters911, 2 (2021), L33. arXiv:2012.13083 [astro- ph.IM] doi:10.3847/2041-8213/abf2c7

    Pith/arXiv arXiv 2021

  14. [14]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network.arXiv preprint arXiv:1503.02531(2015). NIPS 2014 Deep Learning Workshop. arXiv:1503.02531 [stat.ML]

    Pith/arXiv arXiv 2015

  15. [15]

    Huertas-Company and F

    M. Huertas-Company and F. Lanusse. 2023. The Dawes Review 10: The impact of deep learning for the analysis of galaxy surveys.Publications of the Astronomical Society of Australia40 (2023), e001. arXiv:2210.01813 [astro-ph.IM] doi:10.1017/ pasa.2022.55

    Pith/arXiv arXiv 2023

  16. [16]

    Željko Ivezić, Steven M. Kahn, J. Anthony Tyson, Bob Abel, Emily Acosta, Robyn Allsman, David Alonso, Yusra AlSayyad, Scott F. Anderson, John Andrew, et al

  17. [17]

    Laureijs, J

    R. Laureijs, J. Amiaux, S. Arduini, J.-L. Auguères, J. Brinchmann, R. Cole, M. Cropper, C. Dabin, L. Duvet, A. Ealet, et al. 2011. Euclid Definition Study Report. arXiv e-prints(2011). ESA/SRE(2011)12; arXiv:1110.3193. arXiv:1110.3193 [astro- ph.CO]

    Pith/arXiv arXiv 2011

  18. [18]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: Boot- strapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. InProceedings of the 40th International Conference on Machine Learning (ICML) (PMLR, Vol. 202). 19730–19742

    2023

  19. [19]

    arXiv:0805.2366 [astro- ph.IM] doi:10.3847/1538-4357/ab042c

    LSST: From Science Drivers to Reference Design and Anticipated Data Products.The Astrophysical Journal873, 2 (2019), 111. arXiv:0805.2366 [astro- ph.IM] doi:10.3847/1538-4357/ab042c

    Pith/arXiv arXiv 2019

  20. [20]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual In- struction Tuning. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 36

    2023

  21. [21]

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering. In Advances in Neural Information Processing Systems, Vol. 35

    2022

  22. [22]

    Lintott, Kevin Schawinski, Anže Slosar, Kate Land, Steven Bamford, Daniel Thomas, M

    Chris J. Lintott, Kevin Schawinski, Anže Slosar, Kate Land, Steven Bamford, Daniel Thomas, M. Jordan Raddick, Robert C. Nichol, Alex Szalay, Dan Andreescu, Phil Murray, and Jan Vandenberg. 2008. Galaxy Zoo: morphologies derived from visual inspection of galaxies from the Sloan Digital Sky Survey.Monthly Notices of the Royal Astronomical Society389, 3 (200...

    Pith/arXiv arXiv 2008

  23. [23]

    Masters, Brooke D

    James Pearson, Hugh Dickinson, Stephen Serjeant, Mike Walmsley, Lucy Fortson, Sandor Kruk, Karen L. Masters, Brooke D. Simmons, R. J. Smethurst, Chris Lintott, et al. 2026. Galaxy Zoo: Cosmic Dawn – morphological classifications for over 41000 galaxies in the Euclid Deep Field North from the Hawaii Two-0 Cosmic Dawn survey.Monthly Notices of the Royal Ast...

    work page doi:10.1093/mnras/staf2250 2026

  24. [24]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InProceedings of the 38th International Conference on Machine Learning (ICML) (PMLR, Vol. 139)...

    2021

  25. [25]

    Liam Parker, François Lanusse, Siavash Golkar, Leopoldo Sarra, Miles Cranmer, et al. 2024. AstroCLIP: A Cross-Modal Foundation Model for Galaxies.Monthly Notices of the Royal Astronomical Society531, 4 (2024), 4990–5011. doi:10.1093/ mnras/stae1450

    2024

  26. [26]

    Lintott, Anna M

    Mike Walmsley, Campbell Allen, Ben Aussel, Micah Bowles, Kasia Gregorowicz, Inigo Val Slijepcevic, Chris J. Lintott, Anna M. M. Scaife, Maja Ja´błońska, Kosio Karchev, Denise Lanzieri, Devina Mohan, David O’Ryan, Bharath Saiguhan, Crisel Suárez, Nicolás Guerra-Varas, and Renuka Velu. 2023. Zoobot: Adaptable Deep Learning Models for Galaxy Morphology.Journ...

    doi:10.21105/joss.05312 2023

  27. [27]

    Willett, Steven Bamford, Lee S

    Mike Walmsley, Chris Lintott, Tobias Géron, Sandor Kruk, Coleman Krawczyk, Kyle W. Willett, Steven Bamford, Lee S. Kelvin, Lucy Fortson, Yarin Gal, William Keel, Karen L. Masters, Vihang Mehta, Brooke D. Simmons, Rebecca Smethurst, Lewis Smith, Elisabeth M. Baeten, and Christine Macmillan. 2022. Galaxy Zoo DECaLS: Detailed visual morphology measurements f...

    Pith/arXiv arXiv 2022

  28. [28]

    Simmons, Chris Lintott, Kyle W

    Brooke D. Simmons, Chris Lintott, Kyle W. Willett, Karen L. Masters, Jeyhan S. Kartaltepe, et al . 2017. Galaxy Zoo: quantitative visual morphological clas- sifications for 48000 galaxies from CANDELS.Monthly Notices of the Royal Astronomical Society464, 4 (2017), 4420–4447. doi:10.1093/mnras/stw2587

    doi:10.1093/mnras/stw2587 2017

  29. [29]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-VL: Enhancing Vision-Language Model’s Percep- tion of the World at Any Resolution.arXiv preprint arXiv:2409.12191(...

    Pith/arXiv arXiv 2024

  30. [30]

    Willett et al

    Kyle W. Willett et al. 2017. Galaxy Zoo: morphological classifications for 120000 galaxies in HST legacy imaging.Monthly Notices of the Royal Astronomical Society 464, 4 (2017), 4176–4203. doi:10.1093/mnras/stw2568

    doi:10.1093/mnras/stw2568 2017

  31. [31]

    Mike Walmsley, Lewis Smith, Chris Lintott, Yarin Gal, Steven Bamford, Hugh Dickinson, Lucy Fortson, Sandor Kruk, Karen Masters, Claudia Scarlata, Brooke Simmons, Rebecca Smethurst, and Darryl Wright. 2020. Galaxy Zoo: probabilistic morphology through Bayesian CNNs and active learning.Monthly Notices of the Royal Astronomical Society491, 2 (2020), 1554–157...

    Pith/arXiv arXiv 2020

  32. [32]

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. 2024. MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  33. [33]

    Zhang, Alexander Sax, Amir Zamir, Leonidas Guibas, and Jitendra Malik

    Jeffrey O. Zhang, Alexander Sax, Amir Zamir, Leonidas Guibas, and Jitendra Malik. 2020. Side-Tuning: A Baseline for Network Adaptation via Additive Side Networks. InEuropean Conference on Computer Vision. A VLM-Ready Dataset Construction We construct the benchmark from the public Galaxy Zoo: Cosmic Dawn and Euclid Q1 catalogue and image releases. Because ...

    arXiv 2020

  34. [34]

    Willett, Chris J

    Kyle W. Willett, Chris J. Lintott, Steven P. Bamford, Karen L. Masters, Brooke D. Simmons, Kevin R. V. Casteels, Edward M. Edmondson, Lucy F. Fortson, Sugata Kaviraj, William C. Keel, Thomas Melvin, Robert C. Nichol, M. Jordan Raddick, Kevin Schawinski, Robert J. Simpson, Ramin A. Skibba, Arfon M. Smith, and Daniel Thomas. 2013. Galaxy Zoo 2: detailed mor...

    Pith/arXiv arXiv 2013

  35. [2019]

    arXiv:1804.08657 [astro-ph.IM] doi:10.3847/1538-3881/ab089d

    Overview of the DESI Legacy Imaging Surveys.The Astronomical Journal 157, 5 (2019), 168. arXiv:1804.08657 [astro-ph.IM] doi:10.3847/1538-3881/ab089d

    Pith/arXiv arXiv 2019

  36. [2022]

    InAdvances in Neural Information Processing Systems (NeurIPS), Vol

    Flamingo: a Visual Language Model for Few-Shot Learning. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 35