# Model Collapse

Generative models retrained on their own output narrow their distribution and lose the tails first, and training pipelines have lacked a way to correct this while they run. HOMEOSTASIS fuses signed, hash-chained provenance records with three model vital signs, predictive entropy, calibration error and representational drift, into a Lyapunov-style collapse-risk score, and a PID controller adjusts the human-data mixing ratio and quarantine decisions. On a five-dimensional Gaussian recursion over 30 seeds and 20 generations, closed-loop control lowered final collapse risk from 2.67 to 0.48 and held tail coverage at 97.4% against 66.2% without control, while provenance-gated quarantine showed no detectable benefit beyond control alone.

Generative models retrained on their own output narrow their distribution and lose the tails first, and training pipelines have lacked a way to correct this while they run. HOMEOSTASIS fuses signed, hash-chained provenance records with three model vital signs, predictive entropy, calibration error and representational drift, into a Lyapunov-style collapse-risk score, and a PID controller adjusts the human-data mixing ratio and quarantine decisions. On a five-dimensional Gaussian recursion over 30 seeds and 20 generations, closed-loop control lowered final collapse risk from 2.67 to 0.48 and held tail coverage at 97.4% against 66.2% without control, while provenance-gated quarantine showed no detectable benefit beyond control alone.

This work models iterative generative-model training as a closed-loop stochastic process on the probability simplex and analyzes its dynamics under the Fisher-Rao metric instead of the Euclidean metric, deriving contraction and invariance bounds that remain meaningful as dimension grows and giving a human-to-synthetic data ratio threshold that guarantees convergence to a Fisher-Rao ball, concluding that the effective required human-data ratio is higher than previously implied.

This work models iterative generative-model training as a closed-loop stochastic process on the probability simplex and analyzes its dynamics under the Fisher-Rao metric instead of the Euclidean metric, deriving contraction and invariance bounds that remain meaningful as dimension grows and giving a human-to-synthetic data ratio threshold that guarantees convergence to a Fisher-Rao ball, concluding that the effective required human-data ratio is higher than previously implied.

Starting from Llama 3.1 8B, the study first fine-tunes a reviewer on official ICLR reviews from 2018–2023, then trains four successor models on ICLR 2024 data with synthetic reviews generated by that reviewer mixed at 0%, 33%, 66%, and 100%, finding that higher synthetic exposure compresses rating distributions and monotonically reduces same-paper and corpus-level semantic diversity (about 11% and 5%), a pattern the authors call scientific-judgment collapse, and introduces TrustReviewer, which mitigates this tendency through training-time corpus curation and test-time paired activation steering.

Starting from Llama 3.1 8B, the study first fine-tunes a reviewer on official ICLR reviews from 2018–2023, then trains four successor models on ICLR 2024 data with synthetic reviews generated by that reviewer mixed at 0%, 33%, 66%, and 100%, finding that higher synthetic exposure compresses rating distributions and monotonically reduces same-paper and corpus-level semantic diversity (about 11% and 5%), a pattern the authors call scientific-judgment collapse, and introduces TrustReviewer, which mitigates this tendency through training-time corpus curation and test-time paired activation steering.