acceptodds
Withdrawn from ICLR 2027

Multi-Step Reasoning for Deep-Field Galaxy Detection with Uncertainty-Aware Human Feedback

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Source detection in deep-field astronomical imaging poses a formidable challenge characterized by extreme dynamic ranges in scale and flux, low signal-to-noise ratios, ambiguous morphological boundaries, and severe crowding. A fundamental yet overlooked bottleneck is the entanglement of distinct uncertainty sources: aleatoric uncertainty, arising from stochastic background noise, sensor limits, and intrinsic photon shot noise; and epistemic uncertainty, stemming from model underconfidence and out-of-distribution morphology. Disentangling these two uncertainty modes is essential to diagnose whether detection failures are physically bounded by observational limits or algorithmically constrained by representation capacity. To resolve this, we propose a multi-step reasoning framework that couples fine-grained uncertainty disentanglement with a human-in-the-loop calibration loop for robust galaxy detection. Following background stabilization and faint-structure enhancement, a Continuous Thought Machine (CTM) dynamically navigates heterogeneous signal-strength regimes across iterative reasoning steps, progressively integrating multi-stage spatial representations. Crucially, the framework quantifies and decouples aleatoric from epistemic uncertainty maps, isolating ambiguous physical blending from algorithmic failure modes. These spatial uncertainty representations are mapped to an interactive interface that aligns machine uncertainty with expert perceptual intuition, enabling targeted human feedback to iteratively refine model boundaries. Evaluated on simulated deep-field benchmarks, our framework achieves a recall of 88.99% (a 6.30 percentage point improvement over the standard SExtractor baseline of 82.69%). Finally, we demonstrate the real-world utility and cross-instrument generalizability of our framework on ultra-deep observations from the James Webb Space Telescope (JWST) and the Euclid space mission.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
galaxy detection, uncertainty disentanglement, continuous thought, human-in-the-loop learning, deep-field observations

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_8zHJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors present a supervised, uncertainty-aware instance segmentation model for deep field galaxy detection. Their approach is distinguished by a latent reasoning ansatz which iteratively refines the detection. Based on an uncertainty quantification layer, they propose a human-in-the-loop inspection interface for the model's prediction. The detection performance in terms of precision and recall of the proposed segmentation model is evaluated on two simulated datasets (CSST-Galaxy and CSST-Cluster) as well as the real JWST dataset which is equipped with model-generated pseudo labels (derived from the SExtractor-baseline).

Critical strengths

The approach of integrating a latent reasoning model into a detection framework for deep field detection is interesting and the interactive UQ loop seems like an interesting application for practitioning researchers.

Critical weaknesses
  • The paper lacks critical basic components like a related work section positioning the presented approach in the contemporary research landscape. Further, the method description is almost void. See questions Q1, Q2.
  • The evaluation protocol does not support the contribution claims: uncertainty quantification is not validated and the interactive loop just reported; see Q3, Q4
Questions for authors
  • Q1: How exactly does the model look like, in particular: What are fθf_\theta , how is the initial embedding H0H_0 constructed and how many reasoning steps are made? How is the model optimized/trained?
  • Q2: Uncertainty quantification is a central property of the presented method but never sufficiently explained: What are UaU_a and UeU_e in equation (2)? Why do they capture aleatoric or epistemic uncertainty respectively? Which one is used in the illustrations?
  • Q3: If UaU_a and UeU_e are predictive uncertainty estimates, they should correlate with model errors or should react to out-of-distribution inputs (in the case of epistemic). How is this addressed by the current evaluation?
  • Q4: Why the interactive loop is certainty a valuable application idea, how is the claim (ll 245) that it reduces the need for exhaustive manual review, validated?
Reviewer_QBn5OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper "proposes" a multi-step reasoning process for galaxy detection for deep-field astronomical imaging. The method is integrated as part of a web interface for human-in-the-loop galaxy annotation, and compared against a commonly used baseline (SExtractor). The model also claims to estimate uncertainty maps.

Critical strengths

None

Critical weaknesses
  1. There is no description whatsoever of the details of the model architecture used for any of the experiments, and to my knowledge, no supplement or code are provided. The paper results as they stand would therefore be impossible to reproduce.
  2. The only quantitative validation of this method contains in a single table corresponding to an experiment whose details are wildly underspecified. The baseline method, SExtractor, is a highly tunable tool with a comprehensive set of parameters to be selected based on survey specifics. How this tuning was done for experiments is not mentioned at all. Even if these details were to be contained, the reported results are precision/recall for what I assume to be one sensitivity setting, which only represents one point on a PR curve. The baseline method, which we are to believe is inferior, is then used as ground truth labeling for a JWST survey experiment. The main claimed contribution of the paper, which is the estimation and separation of uncertainty into two estimated maps, is completely unvalidated.
Questions for authors

None

Reviewer_9oQEOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This work addresses source detection in astronomical imaging. This is a particularly exciting domain with increasing interest in the machine learning community. The authors propose a "Continuous Thinking Machine" to detect galaxies and augment it with an uncertainty module that generates uncertainty maps on the predictions disentangling aleatoric and epistemic uncertainty.

Critical strengths

The application is very interesting and the authors seem to identify the problems of the domain well. The authors seem to have developed a useful product that may help practitioners of this field.

Critical weaknesses

The paper presents a solution to an interesting problem but does not offer any detail on its methodology beyond high-level arbitrary descriptions like Continuous Though Machines and Uncertainty modules. Disentangling the source of uncertainty is definetly not trivial; yet the authors describe it like their method is common knowledge. The same applies for the "Continuous Thought Machine" module which has been defined as an abbreviation 3 times, without clearly citing its source nor explaining its underlying algorithm. The figures are of particularly low quality. All figures must be converted to a vector format. The main figure (Fig.2) provides a high level description of the method but offers no detail. Overall I think this paper is better suited for a different venue.

As minor comments: There are many formatting issues in the paper, particularly with missing spaces between dots and the following character as well as with citations being too tight to the preceding word.

Questions for authors

By reading this paper I initally have fundamental questions like:

What exactly is your methodology?

What is and how do you use Continous Thought Machines?

What is your Uncertainty modules?

Based on these questions I obviously will need more clarifications. But at this stage the information this paper provides about its methodology is too little.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.