Upstream Coherence Measurement System (UCMS) – Phase 1 Pilot: Quantifying Model Collapse via Coherence Half-Life in Synthetic Data Loops
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Generative AI models that train recursively on their own outputs degrade structurally before downstream metrics react. This preprint introduces the Upstream Coherence Measurement System (UCMS) and its primary KPI — the Coherence Half-Life (τ½): the number of recursive fine-tuning generations before a structural coherence score C(g) falls by half. The framework operationalizes model collapse as a measurable engineering constraint rather than a vague systemic risk.The UCMS Standard Coherence Fidelity Layer (SCFL) probes internal representation geometry using four operator-mapped proxies: manifold geometry preservation (ΔG), relational consistency (φ), tail/curvature stability (κ), and cumulative drift load (δ). Combined with the Coherence–Entropy Transform (T), SCFL feeds a composite upstream coherence score C(g) that decays exponentially under synthetic data pressure, enabling precise half-life estimation via τ½ = (ln 2) / λ.The Phase 1 Pilot protocol targets Llama-3 8B across three synthetic regimes (0%, 50%, 100%) for 5–6 recursive generations, logging SCFL, T, C(g), entropy, and perplexity per generation. The core claim is narrow and testable: SCFL + T detect structural decay earlier than perplexity. If T diverges before C(g) crosses 0.5 while perplexity remains flat — the instrument is validated.This preprint establishes the measurement framework, operator definitions, canonical SCFL formula, computation pipeline, and fully reproducible pseudocode protocol. Empirical results are forthcoming. Computational collaborators with GPU access and HuggingFace fine-tuning experience are invited to contact the author regarding co-publication of the empirical companion paper. Estimated compute: 20–40 A100 hours. Estimated implementation time: one to two weeks.
Fields of Study & Subjects
Topic Modeling
Adversarial Robustness in Machine Learning
Ethics and Social Impacts of AI
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Preprint