acceptodds
Under review as a conference paper at ICLR 2027

WHAT DOES A DIFFERENT COUNTERFACTUAL COST? LOSS-CALIBRATED AUDITS OF LEARNED COUPLINGS

Abstract

Fitting outcome distributions does not determine individual counterfactual predictions. We audit that gap in an already trained flow. Gaussian-source rotations carried through the frozen flow change its pairings while preserving the complete conditional density path, and an exact quadratic identity prices finite alternatives under the original flow-matching objective, allowing query-specific search and independent evaluation without retraining. On a Kang stimulation task the released scCBGM-FM editor closes most of the distance between unedited and stimulated cells; an alternative gives up under one percent of that gain, costs a fraction of a percent of the fitted objective on both the global mixture and the edited condition, and still moves a majority of a 233-cell held-out panel by at least a tenth of a control standard deviation. Every cell keeps a positive predicted increment, so an aggregate reporting only the direction of effect records nothing, while the mean effect and the responder fraction both move. A costlier alternative, qualifying only at the secondary budget, reverses a responder call that four independently trained fits agree on. Learned Gaussian experiments check the accounting against a closed-form boundary price, which a strict reversal approaches without attaining, and show that the query determines which alternative is found and that optimizing an estimated price understates it. Good task performance and agreement across refits can thus coexist with sensitivity in individual predictions: audit the prediction the decision rests on, not only the aggregate reported beside it.

Primary area:
causal reasoning
Keywords:
Flow Matching, Counterfactual Inference, Coupling Sensitivity, Generative Models, Model Auditing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8SoAOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper builds on Nasr-Esfahany & Kiciman 2023 and studies the increase in loss function values after rotating the initial distributions. It then studies the effect of rotating initial noise distribution on single-cell editor's predictions.

Critical strengths
  1. Corrollary 1 is neat, relating the transport map coming from OT and and CFM by inverting the covariance of the initial noise SS. May consider move to body.
  2. Paper did not overstate results.
Critical weaknesses
  1. Dense Writing section 4.1, 4.2 and 5, can't follow what's going on.
  2. I believe the paper is AI written which is fine and the authors acknowledged AI usage. But the sentences are filled with 'it', 'that', 'they', 'both' without out spelling out what each means. One looses track of the meaning easily. Also the abstract is dense and wordy.
  3. I don't see clearly the paper's contribution over Nasr-Esfahany & Kiciman 2023.
Questions for authors
  1. Give a simple example with visualizations and write in simplest words what is the question the paper is addressing. I know sec 4.1 is the proof-of-concept example, but it's hard to read with unclear words such as 'both arms' and 'the family' and so on. Also give clear visualization of the setup of the example.
  2. The intro needs significant rewritting. State what community and what audience would care about the work. How is this work useful for respective fields. For example, why would a single-cell biologist care about section 5. Do you have a pipeline that biologists can adapt to their work?
  3. Per weakness 3.
  4. As far as I understand, most of the results looks neutral and doesn't look like one can draw conclusions from them?
Reviewer_nF54OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper suggests a new method for counterfactual inference based on observational data. For that, it suggests finding such couplings between the probability distributions of potential outcomes that minimize the overall transportation cost. The authors showcase their method for predicting counterfactual expressions based on single-cell data.

Critical strengths

None (the writing was so bad that I failed to see any novelty or significance of the method).

Critical weaknesses

The paper has a multitude of clarity issues and does not meet a baseline for a scientific publication: It looks like poorly generated LLM content, not understandable to a human reader. I got very frustrated with reading it and, thus, I stopped at Sec. 3.2. Here are some main weaknesses that I found:

  • The abstract and the introduction fail to explain the setting, why counterfactual inference matters, and why transport-based counterfactuals are relevant for real-world applications
  • The inconsistency of terminology and notation makes it almost impossible to read. For example, the third sentence of the abstract proceeds as: “Gaussian-source rotations carried through the frozen flow change its pairings while preserving the complete conditional density path, and an exact quadratic identity prices finite alternatives under the original flow-matching objective, allowing query-specific search and independent evaluation without retraining.” Almost none of the terms here are later mentioned in the abstract.
  • 80% of the sentences are not connected to each other content-wise
  • The style of the paper is odd: the introduction provides extremely fine-grained details about the method (and even results); related work only appears at the end (and it is not clear how it relates to the proposed method, e.g., transport-based counterfactuals).
  • The notation is inconsistent and overloaded with details.

Given the issues above, the only thing that is clear to me is a recommendation for a clear rejection.

Questions for authors
  • Can the authors separate the parts of the paper with identification and estimation?
Reviewer_QAWcOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors study the problem of how an individual cell prediction may change for different flow matching models that approximate the overall distribution well. A flow matching model may be able to predict overall response of a treatment but not that on individual cells.

The authors take a trained model and change how it pairs untreated cells with treated cells. They do this by constructing the alternative through an existing flow, preserving its complete conditional path. In essence, this would allow changing individual predictions while retaining overall distribution. Overall, good task performance and agreement across different models does not mean individual predictions will be consistent.

Critical strengths
  • The observation that flow matching models can learn distributions but still may not identify how individual samples correspond is important, particularly in biology where one may like to understand effect of a treatment on an individual cell.
  • Likewise, strong distribution-level performance and agreement across differently trained models does not mean the individual predictions are settled. The authors show this quantitatively.
Critical weaknesses
  • The paper is extremely hard to read and understand. While this may be due to me not being an active researcher in this specific field, but despite that, it is does not even motivate the problem/approach clearly, using unnecessarily complex jargon throughout the paper. For reference, a related paper [1] is much clearer to understand even for a new reader.
  • The authors search for methods where the training loss does not change substantially. But why should a small change in training loss imply that an alternative counterfactual prediction is plausible? This is not clear.
  • The evaluation is limited with synthetic Gaussians and one real dataset (Kang CD4 T cells 1 control-to-stimulated edits).

[1] Single-Cell Concept Bottleneck Generative Models for Interpretable and Controllable Cellular Editing. Alma Andersson, Aya Abdelsalam Ismail, Edward De Brouwer, Doron Haviv, Tommaso Biancalani, Kyunghyun Cho, Gabriele Scalia, Aïcha BenTaieb, Hector Corrada Bravo. ICML 2026.

Questions for authors

NA

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.