Model collapse

Definition (Shumailov et al. 2024, as written): “a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation.” Early collapse is losing information about the tails; late collapse is convergence to a distribution with “little resemblance to the original one,” with variance collapsed toward a point. Distinct from catastrophic forgetting and from data poisoning. See shumailov-2024-model-collapse.

The pollution pathway was named three years before the proofs. Section 6.2 of bender-2021-stochastic-parrots warned that LM text re-enters the next generation’s training sets and carries forward what was problematic in the first. Bender et al. supplied argument, not theorem; the 2024 literature supplies theorems under narrow idealizations.

The setup that decides everything

The recursion carries three shares: fresh model output (alpha), previous-generation data (beta), original real data (gamma). Theorems of doom live at beta = gamma = 0 (pure replacement, fixed sample size). Most public summaries of “AI eats itself” drop this premise, which is the decisive qualification in the literature.

Proven, in the idealized settings

• Discrete support with exact approximation: the only absorbing states are Dirac deltas, so absorption is certain. The leap to neural nets rests on floating-point discreteness, an argument, not a theorem.
• Gaussian moment recursion (Shumailov Thm 3.1): variance to zero almost surely, Wasserstein divergence to infinity, with unbiased sample estimators.
• Rates (Suresh et al., arXiv:2412.17646): Gaussian sigma reaches zero at roughly n generations for n samples per generation; time to forget a discrete item is about linear in its original count. “Tails go first,” quantitatively.
• Borji (arXiv:2410.12954) corroborates the statistical inevitability under KDE/sampling idealizations.

Proven or shown on the other side

• Borkar (arXiv:2506.09401): the asymptotics bifurcate “depending on whether an external source, howsoever minor, is also contributing samples.” Two regimes, not one doom. (Abstract-level; theorem statements unread.)
• Zenil v2 Theorem 5 (zenil-2026-self-improvement-limits, full-text read 2026-10-04): the convergence half proven from inside the collapse formalism itself. If the exogenous fraction stays bounded away from zero, the self-training recursion pulls Q toward P in the ideal-capacity limit; the collapse theorems need it to vanish. The doom paper cites Borkar and Gerstgrasser for exactly this bifurcation and states its own version.
• gerstgrasser-2024-accumulating-data: accumulation (keep real data, add synthetic) avoids collapse across sizes and architectures; Dey and Donoho (arXiv:2410.22812) give the augment-workflow name and report the pi-squared-over-6 linear-regression risk bound, proved by Gerstgrasser et al.
• Fu et al. (ICML 2024 arXiv:2402.11778; ICLR 2025 arXiv:2502.18865): first generalization theory of self-consuming loops; “even a constant-sized proportion of real data ensures convergence”; TV bounds are non-monotone in synthetic-set size.
• Khelifa et al. (arXiv:2606.13796): recursively trained diffusion converges geometrically to a unique non-degenerate limit, a low-pass-filtered infinite mixture, driven by reverse-diffusion early stopping (an implementation artifact), and annealed truncation removes the drift. The Dirac-delta attractor is not the only asymptotics on offer.
• Amin et al. (arXiv:2601.15468): with origin-oblivious ERM, failure to converge on some PAC problems is real, but algorithms exist that learn arbitrary VC classes under arbitrary contamination. Collapse is partly an estimator choice.
• dohmatob-2024-strong-model-collapse: a tiny synthetic share still breaks scaling (error non-monotone in dataset size). Compatible with the above; different question. The escape route, verification, has its own catch: Yi et al. (arXiv:2510.16657) prove verifier-guided loops converge to the verifier’s “knowledge center,” not the truth, and imperfect verifiers plateau or reverse. Kanabar and Gastpar (arXiv:2501.19273) add that provenance information keeps unbounded minimax value even when accumulation averts total collapse.

Where it connects

The recursion premise “no new real-world data is added” is explicit in the collapse proofs (for example, arXiv:2412.14872, metadata verified). The societal pessimism rests on a data-exhaustion estimate the public literature has not agreed on. zenil-2026-self-improvement-limits turns the same machinery into a recursive-self-improvement ceiling and quarantines it to alpha_t going to zero. The bridge to AGI rhetoric is agi-impossibility-claims; the human-culture channel is separate, see llm-homogenization.

Sources

Pages above carry per-paper links. Run-level provenance: outputs/llm-degradation-toward-the-mean.provenance.md and the collapse-theory worker note.