Contribution
Earlier model-collapse studies mostly replaced each generation’s training data with the previous model’s samples. This paper keeps the real data and adds each generation’s samples to it. In that setting the test loss of language models, a molecular diffusion model and (more weakly) a VAE stops growing with generations, and in the linear-regression framework of Dohmatob et al. the test error has a closed form that is bounded in the number of rounds.
Key insight
When data accumulate, round i’s fresh sampling noise is one share in i of the training set, so its effect on the estimate is scaled by 1/i and its effect on squared error by 1/i². The squares are summable, so the error converges. Replacing data gives each round’s noise full weight, and the errors add.
Assumptions
For the theorems (§3, App. B):
- Linear-Gaussian data. x ~ N(0, Σ), y = x·w* + ε, ε ~ N(0, σ²).
- Fixed design. The same T × d matrix X of real covariates is reused every round; only the labels are regenerated, Ŷ_{n+1} = X·ŵ_n + E_{n+1}, with fresh noise E_{n+1} ~ N(0, σ²I_T) of the same variance as the real label noise.
- Under-parameterised, ridgeless OLS. T ≥ d (T ≥ d + 2 for Theorem 2), X full column rank.
- Isotropic features for Theorem 2, Σ = I_d.
- “Model collapse” means test error diverging with rounds (footnote 3); bounded error counts as avoiding it.
For the experiments (§2): one run per configuration as far as reported, five rounds for the language models (ten for replace), eight for the diffusion model and VAE; each round trains a fresh model.
Key results
- Theorem 1. Under accumulation, for all n ≥ 1, ŵ_n = w* + (XᵀX)⁻¹Xᵀ·Σ_{i=1}^{n} E_i / i. Proved by induction (App. B).
- Theorem 2. With Σ = I_d and T ≥ d + 2, E_test^Accum(ŵ_n) = σ²d/(T−d−1) · Σ_{i=1}^{n} 1/i² ≤ σ²d/(T−d−1) · π²/6. Uses E[(XᵀX)⁻¹] = I/(T−d−1) (Lemma 3, from Dohmatob et al.).
- Replace (Dohmatob et al., eq. 4). E_test^Replace(ŵ_n) = σ²d/(T−d−1) · n.
- Replace-Multiple (App. E, eq. 11). Fitting round i to T·i fresh samples from the previous model alone gives σ²d/(T−d−1) · Σ 1/k ≈ σ²d/(T−d−1) · log n: slower, still unbounded.
- Numerics (Fig. 7, Fig. 16). Simulation matches the formulas for d = 10 and 32, T ∈ {100, 316, 1000}, σ² from 0.01 to 100, up to 200 rounds.
- Language models (§2.1, Table 2). GPT-2 9M and Llama-2 12M/42M/126M on TinyStories, one epoch per round. Test cross-entropy at round 4, against round 1: accumulate −0.06 to −0.28; replace +0.53 to +0.66 at temperature 1.0 and +4.00 at temperature 0.3 (GPT-2); replace with data grown to the accumulated size, +0.35 to +0.36.
- Diffusion (§2.2, Fig. 4). GeoDiff on 40,000 GEOM-Drugs conformations: under replace, loss jumps at the first synthetic round and then stays roughly flat; under accumulate, roughly constant.
- VAE (§2.3, Figs. 5–6, 15). CelebA 64×64: under replace, samples collapse to one mode; under accumulate, test error “does increase with the number of iterations (albeit much more slowly)”, samples lose minor features (glasses, hats) while reconstructions keep them.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | In the fixed-design linear model, accumulation gives test error σ²d/(T−d−1)·Σ1/i², bounded by π²/6 times the first round’s | strong | Theorem 2, proved |
| C2 | In the linear model, under accumulation the error still rises with every round and never falls below the first round’s | strong | Theorem 2’s sum is increasing in n (not stated by the authors as such); the language models’ loss instead fell slightly (Table 2) |
| C3 | Matching the amount of data does not remove collapse: what bounds it is keeping the earlier (real) data | moderate | Replace-Multiple, eq. 11 and Table 2’s (*) column |
| C4 | Small language models trained on accumulated data keep their test loss over five rounds | moderate | Table 2, Fig. 2; single runs |
| C5 | Accumulation avoids collapse across model families | weak | holds for the language and diffusion models; the VAE’s error still rises |
Concepts
- Replace / accumulate / replace-multiple. Train round n on only the newest synthetic set; on the real data plus every synthetic set so far; on a synthetic set from the newest model as large as the accumulated one.
- Model collapse (as used here). Test error growing without bound in the number of rounds. The authors list three other uses in the literature: collapse to a few modes, to uniformity, and amplification of artifacts (§4).
Connections
It extends Dohmatob, Feng and Kempe (2024a), “Model collapse demystified”, from replace to accumulate, and reuses their Lemma 3. Marchi et al. (2024) obtained the same square-summable variance independently. Alemohammad et al. (2023) ran one accumulating experiment and concluded that real data “only delays” degradation; this paper reads their plot as a plateau. Martínez et al. (2023a) found fast deterioration even with accumulation on a small dataset; the paper leaves the discrepancy open. Mobahi et al. (2020) study the noiseless case (self-distillation), where predictions shrink to zero.
Bearing on the record
- QUESTION-016 (when does iterated
reconstruction preserve or destroy task-relative sufficiency). The LIT
registered the paper as “a case of retained side information stopping
the loss of sufficiency”. The paper supports something narrower.
- The loss is bounded, not stopped. Σ_{i≤n} 1/i² increases with n, so every round adds error; the limit is π²/6 ≈ 1.64 times the first model’s.
- What is retained is the source, not side information. The real sample stays in every round’s training pool. That is not an interpreter’s side information in QUESTION-016’s sense (A86 §3).
- It is still a pure chain from the real sample. By Theorem 1 every ŵ_n is the first-round estimate plus independent noise. So each model is a garbling of the real data, and nothing in the paper contradicts the question’s point that sufficiency can only decline along such a chain. The paper shows how slowly it can decline, for one task.
- The task is a single one. Squared-error prediction under the original distribution, not a family Q of decision problems (CLAIM-050).
- CLAIM-267 (and CLAIM-056): with e_(i+1) ≤ L_i·e_i + ε_i, end-to-end error is bounded by the discrepancies weighted by the later Lipschitz constants. Accumulation is an exact instance in which the per-round discrepancy is scaled by 1/i, so the weighted sum converges; replacement is the case in which it is not, and the errors add linearly. It is the record’s first closed-form example of the “accumulated but bounded” regime.
- CLAIM-081 (retaining the ancestor and drifting are different profiles). The VAE under accumulation is a measured case: reconstructions keep the minor features while samples drift away from them (App. D, Fig. 15). Shown in pictures only.
- THEORY-155 (Bayesian iterated learning converges to the prior). A contrasting chain. There the data are replaced and the chain forgets its start; here the start is kept in the pool and the chain stays near it.
- CLAIM-199 (a garbling can raise a receiver’s usable information but never above the source). Consistent: no round here beats the first.
- A new THEORY is filed for C1–C3: THEORY-289.
- ML practice. The paper’s message for training on web data that contains model outputs (accumulate, do not replace) is advice for practice and is not filed here.
Limitations
- The theory reuses the real covariates X every round and regenerates only the labels, with noise of exactly the real variance. It is silent on covariate drift, on mode loss, and on any mechanism other than added label noise.
- The experiments are short (five rounds for language models), appear to be single runs, and report no error bars.
- For the VAE, accumulation slowed collapse but did not stop it, which the abstract does not say.
- “Avoids model collapse” is defined as bounded test error. Under other definitions the authors themselves list (mode collapse, artifact amplification), the theorem says nothing.
Open questions
- Does the bound survive covariate regeneration (sampling x from the model as well), or a model class whose per-round error is not additive? A nonlinear or over-parameterised version of Theorem 2 would settle it.
- Do the VAE’s rising error and the language models’ flat error reflect the model class or the task? The same accumulation protocol over more rounds and seeds, across classes, would tell.