acceptodds
Under review as a conference paper at ICLR 2027

The Statistical Origins of Model Collapse: Resampling, Retention, and Exact Variance Recursions

Abstract

Model collapse — the degradation of generative models trained recursively on their own output shumailov2024ai — is usually explained by properties of the neural networks involved. We show the dominant mechanism is instead statistical: plain kernel density estimation (KDE), repeatedly refit to its own samples, reproduces qualitative collapse with no neural network involved. Using this minimal setting, we adjudicate between two follow-up results that pull in opposite directions — gerstgrasser2024collapse's claim that accumulating synthetic data avoids collapse, and dohmatob2024strong's claim that no synthetic fraction is safe — and show that retention of past data, not the injection rate, is what separates convergence from divergence. This retention effect survives sweeps over dimensionality, tail weight, and seven estimator families — including a normalizing flow, a GRU, and a causal self-attention Transformer — and reproduces on real images and real text-like sequences. We then derive exact variance recursions for a tractable Gaussian-MLE idealization of both regimes, verified against Monte Carlo simulation, and test these predictions directly against a small VAE trained by gradient descent: the accumulation-regime prediction matches closely, while under replacement the neural estimator collapses faster than the idealized statistical mechanism's own worst case. Neural networks, in other words, do not create model collapse so much as amplify a mechanism that operates identically in their complete absence.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model Collapse, Self-Consumption Training, Kernel Density Estation, Synthetic Data Accumulation, Data Retention, Recursive Generative Modeling, Variance Recursions, Statistical Mechanisms

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WwB3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies recursive generative-model training through repeated density estimation and sampling. It compares replacement, accumulation, and partial retention of past data, arguing that retention is the main factor determining whether recursive learning remains stable. Experiments cover KDEs, GMMs, a normalizing flow, small autoregressive sequence models, and a VAE on digit images. The theoretical section derives moment recursions for recursive Gaussian estimation and compares them with simulations and selected VAE statistics.

Critical strengths

The retention-probability experiment is useful. Moving beyond the two endpoints of replacement and full accumulation gives a more informative picture of how memory affects recursive training. The consistent direction of the effect across several estimators is also worth documenting.

The paper makes an effort to connect controlled experiments with an analytically tractable model. Multiple trials, explicit sampling parameters, and comparisons with small neural estimators make the observations easier to assess than an isolated demonstration of degradation.

Critical weaknesses

Proposition 2 is not exact under the stated variance estimator. Let N_t = N_{t-1}+n. The pooled sum-of-squares identity used in Proposition 3 gives

\widehat s_{t-1}^2 \left(1-\frac{n}{N_t(N_t-1)}\right). ]

Thus the expected variance decreases multiplicatively and does not converge to the original variance as claimed. Correspondingly, the exact mean-variance expression must retain (\mathbb E[\widehat s_{i-1}^2]) inside the sum. The expression stated in Proposition 2 is an upper bound, not the claimed exact limit. This correction preserves boundedness, but exact recursions are a central contribution of the paper. The distinction between Bessel-corrected estimation and Gaussian MLE also needs to be corrected.

The empirical interpretation has a related gap. Repeated fixed-bandwidth KDE fitting introduces smoothing even in the population limit, so the observed broadening does not by itself identify finite-sample resampling noise as the dominant mechanism. Likewise, forty generations of increasing error do not establish asymptotic divergence for every positive real-data injection rate.

The novelty assessment omits a directly relevant prior study: Kazdan et al., “Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World,” ICML 2025, PMLR 267:29469–29494. That work already studies replacement and accumulation in Gaussian estimation, KDE, and language-model training, including an accumulation–subsampling workflow. The submission needs to establish what its retention sweep and moment calculations add beyond these results. As currently framed, much of the claimed statistical reinterpretation is already established, while the proposed new exact theory contains the issue above.

Questions for authors

Can you reconcile Proposition 2 with the conditional pooled-variance recursion above? Please provide the corrected exact limit and indicate which theoretical and numerical claims change.

What is the precise contribution beyond Kazdan et al. (ICML 2025), particularly for KDE, Gaussian estimation, and fixed-size subsampling of accumulated data?

What bandwidth rules are used in Sections 4–7? Can you separate finite-sample resampling error from repeated kernel smoothing using a population baseline and sample-size/bandwidth controls? What supports asymptotic divergence rather than a biased plateau under positive real-data injection?

Under the 4000-sample cap, are original real observations protected or randomly discarded? Can retention be compared at matched training sample sizes? Please also clarify whether “identical batches” means identical observations or shared random numbers.

What supports the “faster than the idealization’s worst case” interpretation of the VAE result? The propositions do not establish a worst-case bound for neural estimators. Please distinguish typical variance trajectories from ensemble moments and provide uncertainty estimates for the across-trial variance comparisons.

4 comments

Comment 1: Proposition 2 and the MLE label (you are right)

Thank you for the careful reading. We confirmed both points.

Proposition 2. The pooled sum of squares gives E[s^t2∣Ft−1]=s^t−12 (1−nNt(Nt−1))E[\hat s_t^2\mid\mathcal F_{t-1}]=\hat s_{t-1}^2\,(1-\frac{n}{N_t(N_t-1)}), your expression. The exact statements are E[s^t2]=σ02∏i≤t(1−nNi(Ni−1))E[\hat s_t^2]=\sigma_0^2\prod_{i\le t}(1-\frac{n}{N_i(N_i-1)}) and Var(μ^t)=σ02n0+∑i≤tn E[s^i−12]Ni2\mathrm{Var}(\hat\mu_t)=\frac{\sigma_0^2}{n_0}+\sum_{i\le t}\frac{n\,E[\hat s_{i-1}^2]}{N_i^2}. The limit displayed in Prop. 2 is an upper bound, and E[s^t2]→σ02Π∞E[\hat s_t^2]\to\sigma_0^2\Pi_\infty, not σ02\sigma_0^2. For n0=2000n_0=2000, n=500n=500: Π∞≈0.99956\Pi_\infty\approx0.99956 and lim⁡Var(μ^t)≈0.000942\lim\mathrm{Var}(\hat\mu_t)\approx0.000942 (bound 0.0009430.000943).

What changes. The statement of Prop. 2, the claim E[s^t2]→σ02E[\hat s_t^2]\to\sigma_0^2 in the text and proof, and the "MLE" label. What does not. Boundedness, Prop. 1, and Prop. 3, whose product form already gives the correct E[s^t2]E[\hat s_t^2]. Our code used the exact QtQ_t-recursion, which matches the product to seven digits, and the correction (0.04%) is far below Monte Carlo resolution, so no figure or comparison moves.

MLE. We analyze the sample mean with the Bessel-corrected variance, which is not the MLE. The setting is now called "Gaussian estimation with an unbiased variance" throughout.


Comment 2: KDE smoothing, injection rate, and Section 6 controls (you are right)

Bandwidth. Section 3 uses fixed bandwidths (0.5, 0.1); Sections 4-7 use scipy.stats.gaussian_kde (Scott's rule), refit every generation. A KDE refit to its own samples broadens the density even with infinitely many samples (Scott's rule multiplies the variance by 1+n−2/51+n^{-2/5} per generation), so the KDE results alone do not identify sampling noise.

Population baseline. The infinite-sample recursion pt+1=qt∗N(0,n−2/5Var(qt))p_{t+1}=q_t\ast N(0,n^{-2/5}\mathrm{Var}(q_t)), with no sampling, reproduces our results:

Section 4, gen 25Empirical KL / VarNoise-free KL / Var
Replacement, 500 / gen0.962 / 124.90.963 / 131.5
Replacement, size-matched0.799 / 36.20.790 / 36.0
Accumulation0.146 / 18.10.145 / 18.5

It also reproduces Table 3. So the KDE experiments measure deterministic smoothing bias, and the paper no longer claims otherwise (new Sec. 3.2).

Sampling-noise control. A 1-D two-component GMM fitted by EM contains p0p_0, so any drift is sampling noise (15 trials). KL after 25 generations: 0.155 (replacement), 0.021 (size-matched), 0.002 (accumulation). The retention ordering holds against noise, and the size-matched control is clean.

Injection rate (your 40-generation point). Withdrawn. In the KDE, forgetting with injection α\alpha diverges exponentially below α∗=n−2/5/(1+n−2/5)≈0.077\alpha^*=n^{-2/5}/(1+n^{-2/5})\approx0.077 and plateaus above it, a bandwidth-rule artifact. With noise only (GMM, T=40T=40), α=0\alpha=0 keeps growing (KL 0.074, 0.135, 0.225 at generations 10, 20, 40), but α=0.05\alpha=0.05 plateaus (0.039, 0.038, 0.044). We no longer claim "no safe α\alpha".

Section 6 controls (8 trials, final KL at α\alpha = 0.02 / 0.05 / 0.10).

ForgetForget, size-matchedAccumulateAccumulate, real protected
KDE.911 / .868 / .787.708 / .592 / .454.172 / .162 / .147.109 / .098 / .086
GMM.092 / .086 / .048.0090 / .0093 / .0039.0019 / .0018 / .0047.0012 / .0013 / .0022

For the KDE the retention advantage survives size-matching. For the GMM, sample size explains most of the gap, about 5x remains at small α\alpha, and at α=0.1\alpha=0.1 matched forgetting is indistinguishable from accumulation. We have qualified the paper's claim accordingly. The leaky-memory sweep repeated with the GMM is also smooth but roughly geometric rather than concave, so "concave" is removed as a general claim.


Comment 3: Kazdan et al., the cap, "identical batches", and the VAE

Kazdan et al. (ICML 2025) and three related papers are now cited, and the abstract, Introduction and Related Work are rewritten. We no longer claim novelty for the replace/accumulate separation, the KDE replication, or the 1/N21/N^2 mechanism for the mean. What we did not find in that work: the fixed-α\alpha retention comparison, leaky-memory interpolation, the exact product form for Var(s^t2)\mathrm{Var}(\hat s_t^2), and the recursion-vs-neural-estimator test. Our cap in Sections 6-7 is accumulate-subsample; leaky memory generalizes it.

The cap. It is a uniformly random subsample of the whole pool, so real observations are not protected (control above).

"Identical batches" was inaccurate: the arms use different seeds and their synthetic batches come from different models, so they are matched in size and injection rate only. Corrected in Section 6 and the Reproducibility Statement.

AuthorsOpenReview ↗

Response (Cntd)

VAE "worst case" is withdrawn; the propositions do not bound neural estimators. With 95% bootstrap intervals over 20 trials: accumulation growth 1.17 [0.57, 3.09] vs. theory 1.20 (consistent, weak test; the VAE's s^2\hat s^2 also falls 0.47 to 0.16); replacement 1.68 [0.60, 5.67] vs. theory 9.26. s^2\hat s^2 collapses to about 2×10−52\times10^{-5} in 20/20 trials. The bias explanation is now a hypothesis.

Reviewer_WwB3OpenReview ↗

Response to the Authors

Thank you for the detailed rebuttal and for carefully addressing my concerns. I appreciate the authors' corrections, additional control experiments, and clarification of the paper's claims. My main concerns have been largely addressed, although some limitations remain. I will therefore raise my rating.

Dear reviewer.

Thanks a lot for taking the time to go through our reply. It seems you may have forgotten to raise your rating. Best,

Reviewer_SjHuOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper analyzes model collapse in various settings, including KDE and estimating a Gaussian. It includes a few mathematical statements regarding the variance of a few of these, and various experiments. The paper also examines the role of fresh data injection and data accumulation/retention.

Critical strengths
  • The paper is timely and aims to clear up several seemingly different viewpoints in the literature.
  • The settings analyzed in the paper are straightforward, the messages are clear, and the paper is easy to read (albeit a bit LLM-heavy in the writing)
Critical weaknesses

Unfortunately, I believe the authors were unaware of a large part of the literature, and I do not believe this paper to be a significant contribution beyond existing results. More specifically:

  • The entire narrative regarding most works on model collapse concerning neural networks is simply false. Many past works have analyzed KDE, and MLE (in fact, probably most works on model collapse) (e.g. [1,2,3]).

  • [1, 4] analyzed model collapse for KDE.

  • It's not clear what sections 3 and 4 are contributing, in light of [1].

  • There is already a lot of discussion in the literature about retention and the difference from fresh data injection, and much of the conclusions are not new. The way it's usually explained is through variance calculations, essentially as in section 9. See e.g. [2]

[1] Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World - Kazdan et al., 2025 [2] Universality of the π2/6 Pathway in Avoiding Model Collapse - Dey and Donoho, 2024 [3] When Models Don't Collapse: On the Consistency of Iterative MLE - Barzilai and Shamir, 2024 [4] Towards Theoretical Understandings of Self-Consuming Generative Models - Fu et al., 2024

Questions for authors

Can you address the weaknesses?

3 comments

Comment 1: concessions

Thank you for the pointers. We have now read [1-4] and agree on three points.

1. The framing was overstated. We wrote that explanations for model collapse focus almost entirely on neural networks. That is not accurate: Shumailov et al. analyze recursive Gaussian fitting, Kazdan et al. [1] run KDE and Gaussian experiments with a Gaussian accumulation theorem, Fu et al. [4] bound KDE error, and Barzilai and Shamir [3] analyze iterative MLE. We removed this framing from the abstract and Introduction and cite [1-4] in Related Work.

2. Sections 3-4 are confirmatory. Per [1] (Sec. 2.2), replace-KDE degrades and accumulate-KDE is stable. We now present them as a baseline, not a contribution.

3. The mean-variance recursion is not new. [2] (Lemmas 5.2-5.3) gives G⋅VG\cdot V under discard and V∑1/g2≤(π2/6)VV\sum 1/g^2\le(\pi^2/6)V under augmentation, the same mechanism as our Props. 1-2 for the mean. We attribute it to [1,2]. (Another reviewer found our displayed Prop. 2 limit was only an upper bound, since E[s^t2]→σ02Π∞E[\hat s_t^2]\to\sigma_0^2\Pi_\infty with Π∞≈0.99956\Pi_\infty\approx0.99956; corrected, no numerical result changes.)


Comment 2: what is left that is new

CoversNot found in it
[1] KazdanGaussian, KDE, LM fine-tuning; replace / accumulate / accumulate-subsamplePer-generation injection rate vs. retention (their real-vs-synthetic design in Sec. 4 is static); recursion tested against a trained neural model
[2] Dey, DonohoDiscard / augment / augment-subsample for M-estimatorsFresh-data fraction; partial memory; KDE; trained networks. Augment-subsample by simulation only
[3] Barzilai, ShamirIterative MLE with accumulation, parametric familiesReplacement; KDE; variance recursions; retention vs. injection
[4] Fu et al.KDE and random-feature diffusion; TV upper boundsLower bounds; controlled experiment fixing α\alpha; partial memory

We do not claim novelty for "accumulation avoids collapse" or the 1/N21/N^2 mechanism for the mean. We have not found the following in [1-4]:

  1. Retention separated from injection rate. At fixed α\alpha, forgetting ends at KDE KL 0.77-0.90 and accumulation at 0.15-0.17 (Table 3). With fit sizes matched the KDE gap persists, while against sampling noise (a bias-free GMM) it persists at small α\alpha and vanishes at α=0.1\alpha=0.1.
  2. Leaky-memory interpolation between forget and accumulate. The curve is smooth; its shape depends on the mechanism (concave for the KDE, roughly geometric for the GMM).
  3. An exact product form for Var(s^t2)\mathrm{Var}(\hat s_t^2) under accumulation (Prop. 3). We did not find an equivalent in [1-4].
  4. Smoothing bias vs. sampling noise in the KDE. [1] notes bandwidth selection matters; we quantify it. A noise-free recursion reproduces our KDE results (replacement KL 0.963 vs. 0.962 empirical), and a GMM isolates noise (replacement 0.155, size-matched 0.021, accumulation 0.002). This is unlikely to surprise experts, but it changes how our KDE experiments must be read, and the paper says so.
  5. Recursions tested against a trained VAE. Accumulation: 1.17x observed (95% CI [0.57, 3.09]) vs. 1.20x predicted, consistent but a weak test; replacement growth is below the prediction.

Our contribution is narrower than the original abstract claimed, and the revised abstract and Introduction claim exactly that.


Comment 3: revisions

(1) Abstract and Introduction rewritten. (2) [1-4] added to Related Work with the comparison above. (3) Sections 3-4 relabelled as a baseline. (4) Props. 1-2 attributed to [1,2]. (5) The "amplify rather than create" claim is softened. (6) You note the prose is LLM-heavy; the paper carries an AI-use statement and we will do a full editing pass.

Reviewer_SjHuOpenReview ↗

Response to Authors

I appreciate the effort that the authors are making in addressing the issues. Nevertheless, changing the wording so that most of sections 3, 4, and 9 are presented as "confirmatory" of existing results rather than a novel contribution does not fix the issue that much of the paper is not novel. I do not believe that the remaining contributions suffice as-is. Moreover, I think the paper needs more substantial writing revisions, realigning itself around the remaining contributions (and possibly expansions of those contributions), which would warrant a new review. As such, I unfortunately cannot recommend acceptance.

Thank you for the follow-up answer. We agree that relabeling Sections 3, 4 and 9 doesn’t make them novel. We are not claiming it does. To address this, we have dropped the neural-network framing, credited [1-4] in Related Work, and attributed Props. 1-2 to [1,2].

In the revised version, we now cut those sections down to a short baseline and rebuild the main text around what we couldn’t find in [1-4]:

  • separating retention from the injection rate,
  • the interpolation between forgetting and accumulation,
  • the product form in Prop. 3,
  • the bias-versus-noise split in the KDE, and
  • the test against a trained VAE.

We know you may still see that as little. But you gave soundness and presentation a 3, and your objections are about novelty and positioning, not correctness. A clear rejection seems harsh for a correct, readable paper with limited but real contributions, and we’d ask you to consider a weak rejection instead.

Thanks for the references; they made the paper better.

Reviewer_RMeSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors aim to better understand model collapse, a process in which a model degrades when trained on data it generates. The main conclusion is that this is not specific to neural networks, but it is a natural process that affects any statistical model. They show that a KDE trained on data generated from the previous generation quickly diverges after a few iterations. The authors also test multiple models (neural and classical), distributions, and dimensionalities.

Critical strengths

The main strength of the paper is that it offers a general insight into model degradation, which I find valuable. Through a series of well-motivated experiments, it shows that this phenomenon is not limited to neural networks but also affects other statistical models. The experiments support this broader conclusion well.

The second strength is that the paper is easy to read. Although I am not very familiar with this field, I enjoyed reading it. Several experiments naturally addressed questions that arose as I read.

Critical weaknesses

The paper focuses on small-scale experiments, with very little discussion of how increasing the scale would affect this phenomenon. From a statistical point of view, it might seem straightforward that recursive training must diverge, but this is less obvious when we think about LLMs. After all, our dreams also serve as a form of synthetic training for our brains.

In line 153, the replacement method uses 500 training samples per iteration, while the accumulation method uses the union of all samples. This probably does not change the main conclusions, but the unequal sample sizes make the comparison less fair, especially when the replacement method could draw more samples from the KDE to match the accumulation method.

Overall, the manuscript is readable, but the conclusions are a bit dense and focus more on which recursive training strategy is better, even though none truly works. This section would benefit from addressing the main hypothesis more directly and suggesting some future directions.

Questions for authors

The paper tests several strategies for recursive training, but all of them diverge. The question is not which method converges, but which diverges most slowly. What is the value of studying these strategies if none works?

I would like to hear your thoughts on larger-scale experiments. Do you think recursive training could improve performance on test sets once the model is good enough, or is collapse inevitable?

2 comments

Comment 1: unequal sample sizes (Section 4)

Thank you for the encouraging review. You are right that the arms are not size-matched, so we ran your control: replacement draws as many samples as the accumulation pool holds (2000 + 500t) but fits only on that fresh draw. Everything else matches Section 4 (15 trials, T=25), and our original arms reproduce.

Gen 25KLVariance (true 17.0)
Replacement, 500 / gen0.962124.9
Replacement, size-matched0.799 ± 0.01536.2
Accumulation0.14618.1

Matching slows the drift but does not remove it: variance rises roughly linearly (20.9 to 36.2 between generations 5 and 25) and KL stays far above accumulation's.

Why. For the KDE this is smoothing, not sampling noise: repeated refitting broadens the density even with infinite samples, and a noise-free recursion reproduces these numbers (0.790 / 36.0). We added this analysis to the paper (new Sec. 3.2). Against sampling noise alone, a 1-D GMM (model class contains the truth) gives KL after 25 generations of 0.155 (500 samples), 0.021 (size-matched) and 0.002 (accumulation). The Gaussian theory (Props. 1-2 and the remark after Prop. 2) explains this mechanism: with NtN_t samples replacement adds σ2/Nt\sigma^2/N_t per step, a harmonic series that diverges logarithmically, while accumulation adds the damped n/Nt2n/N_t^2.

Section 6. We also ran a size-matched forget arm (8 trials; α\alpha = 0.02 / 0.05 / 0.10). KDE: forget .911 / .868 / .787, matched .708 / .592 / .454, accumulate .172 / .162 / .147. GMM: forget .092 / .086 / .048, matched .0090 / .0093 / .0039, accumulate .0019 / .0018 / .0047. Sample size explains much of the GMM gap, about 5x remains at small α\alpha, and at α=0.1\alpha=0.1 matched forgetting is indistinguishable from accumulation. We qualified the paper's retention claim (new Table 2).


Comment 2: "none truly works", scale, and whether recursive training can help

Not all strategies diverge. Accumulation is bounded, both provably (Props. 2-3) and empirically (KL plateaus at about 0.15 vs. 0.96 for replacement). We should be honest about what that means: it does not improve on the original fit. Its KL is about twice the generation-0 KL (0.075), and in the Gaussian theory the mean-estimator variance settles near 1.9x its initial value. The point of the comparison is not which strategy diverges most slowly but what separates bounded from unbounded error: whether past data is retained. This gives guidance: keep curated real data rather than only capping the synthetic fraction per run.

Can recursive training improve a good model? In our setting, no, for a structural reason: synthetic samples carry no information about p0p_0 beyond the original n0n_0 real samples, so pure self-training is bounded at about the original fit. Improvement needs an information source our loop excludes (fresh real data, selection or verification of synthetic samples, a stronger external model). We have not tested these. Your dreaming analogy fits that picture, since replay helps when combined with selection or new experience, but we would not push it further.

Scale. Every neural model has about 10410^4 parameters, and we make no claim that rates or constants transfer to LLMs. The retention effect has similar magnitude in a GRU and a causal-attention Transformer (2.6x vs. 2.7x).

Revisions made. (1) The size-matched controls above. (2) The Discussion now opens with the hypothesis and answer, states that accumulation is bounded but does not improve on the original fit, and lists limitations and future directions (scale, selection or verification, fresh-data mixing, capacity). (3) "Neural networks amplify rather than create collapse" is softened, since it rested on one small VAE.

Thank you for the additional experiments and explanations, which have clarified the points I raised. I find the paper interesting, but as I mentioned in my initial review, I am not familiar enough with this field and its literature to fully assess the novelty of the submission in relation to prior work, such as “Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World” (Kazdan et al., 2025). I therefore keep my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.