acceptodds
Under review as a conference paper at ICLR 2027

Model Collapse Needs Only One Generation of Self-Training to Become Measurable

Abstract

Recursive self-training on synthetic data generated by antecedent models introduces severe distribution shifts and cumulative degradation, commonly referred to as model collapse. While prior literature debates the asymptotic limits of multi-generation feedback loops, the precise timing and measurable emergence of collapse remain underspecified. We present a controlled empirical and theoretical analysis demonstrating that model collapse is not an asymptotic artifact; it becomes strictly measurable after exactly one generation of self-training. Using discrete, countable state spaces and controlled MNIST image distributions, we track the total-variation distance between original and self-trained data generators. Our findings reveal that the initial collapse transition is highly pronounced yet sublinear across subsequent iterations, and downstream classification accuracy directly tracks this distributional degradation rather than masking it. Furthermore, we evaluate mitigation strategies, showing that data mixing ratios must scale logarithmically with generation depth to preserve support coverage. We conclude that model collapse constitutes an immediate operational bottleneck rather than a distant theoretical failure mode, necessitating strict provenance tracking in synthetic data pipelines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
model collapse, self-training, distribution shift, data degradation, generative models, recursive training

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_VFTEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper should receive a desk reject, since a large proportion of the text is AI-generated and does not make any sense. For example, section 5 and section 6.

Critical strengths

N.A.

Critical weaknesses

The discussion section "A partially supported claim is the most common and least well reported outcome in empirical work, because it resists compression into a headline. We therefore give the predicate-level split rather than an overall verdict: the supported predicates are usable, the unsupported ones are not, and collapsing the two into a single sentence would misstate the evidence in whichever direction the summary leaned." is 100% AI generated, on GPTzero detection.

Questions for authors

N.A, since this paper is not human-written.

Reviewer_Pp9WOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

N/A

Critical strengths

N/A

Critical weaknesses

The paper should be rejected for the following reasons:

  1. Significant change of abstract: The abstract in the main paper is substantially different from the version for the abstract deadline used for reviewer matching. The new version is poorly written, with fragmented sentences which do not explain the scope of the work.
  2. Absence of references: The submitted PDF file does not contain any bibliographic entry, with a few supposed citations in Sections 1 and 2 leading to no actual references. As a consequence, the motivation, contributions, and related work are barely explained and contextualized.
  3. Formatting issues: At the bottom of Page 7, a table (supposedly Table 7) exceeds the margins and goes beyond the page boundary.
  4. Absence of AI use statement: The submitted PDF file does not contain the mandatory AI disclosure section. Based on the formatting flaws, low-quality content and incoherent writing, there is good reason to suspect that a substantial portion of this submission was generated using AI without proper disclosure.
Questions for authors

N/A

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

The paper is on "model collapse," by which I think the authors, or the AI that wrote the submission, meant mode collapse. I cannot provide a summary for this paper because I don't even know what I read.

Critical strengths

There are no strengths in this paper.

Critical weaknesses
  1. No references: All references are question marks.
  2. The paper talks about "model collapse," by which I think the authors meant "mode collapse." The very first line in the abstract is not valid. There are ways to measure mode collapse apart from manually inspecting the samples. In fact, what the authors described in the second line of the abstract is how it is already done.
  3. The results are all junk: The same y-ticklabels for different entries, flat plot lines, etc.
Questions for authors

No questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.