Does verifier filtering actually prevent model collapse long term?

The paper asks whether filtering synthetic training data through a verifier — "whether a human or a better model" — actually prevents model collapse, or only postpones it. Its answer, worked out in a linear regression testbed and then checked on VAEs trained on MNIST and SmolLM2-135M fine-tuned on XSUM, is that verifier-guided retraining "will not cause model collapse" and can produce real near-term improvement, but "ultimately drives the parameter estimate to the verifier's knowledge center in the long run." Unless the verifier is "perfectly reliable," the paper states, "these early gains will plateau and may even reverse."

The mechanism is a "verifier-induced bias-variance trade-off: filtering synthetic data reduces variance but may introduce bias." The verifier is modeled as holding a knowledge ball around some center θc with radius r, giving only binary accept/reject feedback on each synthetic sample (deliberately modeled this way because, the paper notes, real verifiers — LLM raters or human annotators in RLHF-style setups — "may not even know" their own implicit θc or r explicitly). Each round of filtering injects more of that verifier knowledge into the estimator while the contribution of the original real data decays, so the long-run fixed point of the retraining process is the verifier's center, not the truth. This yields three phases: an unbiased verifier (θc = θ⋆) drives continuous convergence to the true parameter; a mildly biased verifier gives short-term improvement that "eventually plateaus or deteriorates as verifier bias accumulates" — called "the most practically relevant" case; a strongly biased verifier causes degradation and collapse even with filtering in place.

This complicates Does training on AI-generated content permanently degrade model quality?, which establishes collapse as the default outcome of unfiltered recursive training. This paper agrees collapse is the default but shows that inserting any external verifier changes the near-term trajectory — filtering is not cosmetic, it measurably reduces variance and can produce genuine improvement for a while. What it denies is that filtering is a durable fix: the long-run attractor simply moves from "the training distribution's vanishing tail" to "the verifier's own knowledge center," so a biased-but-useful verifier still ends in degradation, just on a longer clock. It also stands in contrast to Can models trained on many imperfect experts outperform everyone?'s route out of collapse, which works by denoising through diversity across many training sources rather than through a single discriminating verifier — this paper's framework has no mechanism for canceling verifier bias through multiplicity, since every filtering round draws on the same θc.

The theory is proven only in a well-specified linear-regression setting with a single ground-truth parameter, which the paper itself flags as an idealization: real generative models, especially language models, are approximating an unknown, likely non-parametric data-generating process, and "a singular 'true model' may not even exist" for natural language, so there may be no well-defined θ⋆ for a biased verifier to drift away from. The VAE and LLM experiments are offered as qualitative confirmation only, and the paper explicitly leaves verifier dynamics in LLMs and vision models, and non-block-orthogonal synthetic designs, as open future work. The implication for LLM-as-judge or RLHF-style self-improvement loops is that filtering should buy real but time-limited gains proportional to verifier quality, with the size and bias of that verifier setting a ceiling that more filtering rounds cannot raise — a claim this paper argues for mathematically rather than measures directly in a frontier model.