Model Collapse in Recursive Synthetic Data Training: Mechanisms, Evaluation, and Mitigation Strategies

Abstract

Generative models increasingly produce content that can enter later training corpora, creating recursive synthetic-data feedback loops. This review examines when such loops cause model collapse and when synthetic data remain useful. It synthesizes theoreti cal, empirical, and methodological evidence across language, diffusion, evaluation, mitigation, and governance, organized into four dimensions: training regime, degradation mechanism, outcome, and intervention point. Collapse is most consistently associate d with repeated replacement of independent observations with unverified outputs, resulting in tail loss, declining diversity, cumulative error, and bias amplification; recursive diffusion studies also report a shift toward memorization. However, stable or beneficial outcomes occur when independent data or accumulated history are retained and generated samples are corrected, filtered, reweighted, or verified. Findings vary across model families, sampling policies, synthetic ratios, recursive depths, and metr ics, limiting direct comparison and large-scale generalization. Priorities include standardized multigeneration benchmarks, multidimensional longitudinal evaluation, recovery research, realistic multimodel ecosystems, and provenance-aware governance. Overall, model collapse is conditional on the workflow rather than an inevitable consequence of synthetic data.