Source
Gerstgrasser, Schaeffer, Dey, Rafailov, Pai, Sleight, Hughes, Korbak, Agrawal, Gromov, Roberts, Yang, Donoho and Koyejo (2024), LIT-971 (arXiv 2404.01413v2), read in full in NOTE-815: §3, Appendix B (proofs), Appendix E (Replace-Multiple), §2 and Appendix C (experiments).
What was actually shown
The chain. T real points (x, y) with x ~ N(0, I_d) and y = x·w* + ε, ε ~ N(0, σ²), T ≥ d + 2. Round 1 fits ordinary least squares. Each later round draws new labels for the same covariates X from the previous fit, Ŷ_{n+1} = X·ŵ_n + E_{n+1} with fresh noise of variance σ², and fits again.
The theorems. If each round’s data are appended to all earlier rounds’, the n-th fit is ŵ_n = w* + (XᵀX)⁻¹Xᵀ·Σ_{i≤n} E_i/i (Theorem 1, proved by induction), and its test error is
E_test^Accum(ŵ_n) = σ²d/(T−d−1) · Σ_{i=1}^{n} 1/i² < σ²d/(T−d−1) · π²/6
(Theorem 2). If each round uses only the newest labels, the error is σ²d/(T−d−1)·n (Dohmatob et al.). If each round uses a newest-only set as large as the accumulated one, it is σ²d/(T−d−1)·Σ_{k≤n} 1/k ≈ σ²d/(T−d−1)·log n (App. E). Simulations match all three up to 200 rounds (Figs. 7, 16).
What could have come out otherwise. Accumulated synthetic data could have dragged the estimate as replacement does, only more slowly. The Replace-Multiple case shows that matching the amount of data is not enough: what bounds the error is that earlier rounds, the real one among them, keep their share.
Neural models. Language models of 9M–126M parameters on TinyStories hold or slightly lower their test loss over five accumulating rounds (−0.06 to −0.28 nats at round 4) where replacing raises it (+0.53 to +0.66 at temperature 1.0), and a molecular diffusion model behaves alike. These are single, short runs and are not part of the claim.
What this does not say
- It does not say the loss stops. Σ 1/i² increases with n. Every round adds error, and the n-th model is never better than the first; the limit is about 1.64 times the first fit’s error.
- It does not say synthetic data are harmless or helpful. In this model each synthetic round only adds noise. The bound comes from diluting that noise, not from the synthetic data carrying anything.
- It does not cover other kinds of collapse. The covariates are never regenerated, the noise is Gaussian with the real variance, and “collapse” means divergent test error. Loss of modes, drift of the input distribution and artifact amplification are outside the theorem. The VAE experiment, where accumulated data still lost minor features from samples and test error still rose, is a case it does not reach.
- It does not say retained side information preserves sufficiency. What is retained is the original sample, in the training pool. By Theorem 1 every model is the first estimate plus independent noise, a garbling of the real data. Only one task (squared-error prediction) is measured.
- It is not an instruction. “Accumulate, do not replace” as a policy for training on web data is the paper’s practical message and is not filed here.
Connections
- QUESTION-016. It bears on the question without answering it. It is a chain in which loss for one task stays bounded because the source is kept, not because an interpreter brings side information, and sufficiency for a family of decisions is not measured. It does show that “progressively destroys” admits a bounded form: a pure chain from the source whose loss converges.
- 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. This is an exact instance in which the discrepancy of round i is scaled by 1/i, so the weighted sum converges, and replacement is the case in which the discrepancies add.
- THEORY-155. A contrasting chain: Bayesian iterated learning replaces the data and converges to the prior, forgetting the start. Here the start is kept in every round and the chain stays near it.
- CLAIM-081. The VAE run is a pictured case of its two profiles, reconstructions keeping detail while samples drift. Pictures only.
- CLAIM-050. The theorem is about one decision problem, so it says nothing about Blackwell comparison over a family Q.
Note of 2026-10-19: Theorems 1 and 2, re-derived
The promotion audit (batch 042) re-derived the closed forms. Let X be the T × d covariate matrix, rows N(0, I_d), T ≥ d + 2, and A = (XᵀX)⁻¹Xᵀ, so AX = I. Round 1’s labels are Xw* + E₁; round i’s are Xŵ_{i−1} + E_i, with the E_i independent N(0, σ²I_T).
- Accumulate. Round n fits least squares on n stacked copies of X, so ŵ_n = (1/n)Σ_{i≤n} Ay_i, with Ay₁ = w* + AE₁ and Ay_i = ŵ_{i−1} + AE_i. If ŵ_k = w* + AS_k with S_k = Σ_{i≤k} E_i/i for every k < n, then nŵ_n = nw* + A(Σ_{k<n} S_k + Σ_{i≤n} E_i), and Σ_{k<n} S_k = Σ_{i<n} E_i(n − i)/i, so the sum in brackets is nS_n.
- Error. E‖ŵ_n − w*‖² = σ²Σ_{i≤n}(1/i²)·E tr((XᵀX)⁻¹), and the inverse-Wishart mean E[(XᵀX)⁻¹] = I/(T − d − 1) gives σ²d/(T − d − 1)·Σ_{i≤n} 1/i², below π²/6 times the first fit’s σ²d/(T − d − 1) and larger at every round.
- Replace. ŵ_n = ŵ_{n−1} + AE_n = w* + AΣ_{i≤n} E_i, so the error is σ²d/(T − d − 1)·n.