Matteo Benati, Alessandro Londei, Denise Lanzieri, Raffaello Mastromarino, Vittorio Loreto · 2026 · arXiv
DOI: 10.48550/arXiv.2610.30009 · open-access PDF
Wir studieren diese Degeneration spektral in einer kontrollierten Markov-Einstellung. Das Eigenwertspektrum des gelernten Übergangsoperators bietet einen direkten Blick auf den Prozess, da die subdominanten Eigenwerte festlegen, wie schnell die Kette mischt und daher wie viel von dem Zustandsraum sie weiter erkundet. Wir finden, dass die Probenahme bei Temperatur $T_s < 1$ auf den Operator wie wiederholte Verschärfung wirkt, sie in Richtung eines deterministischen Systems drückt: der subdominante Eigenwert
Machine translation (Meta M2M100); the English original follows.
Training generative models on synthetic data causes loss of diversity and a decrease in performance, a phenomenon known as model collapse. We study this degeneration spectrally in a controlled Markov setting. The eigenvalue spectrum of the learned transition operator offers a direct view of the process, since the subdominant eigenvalues set how fast the chain mixes and therefore how much of the state space it continues to explore. We find that sampling at temperature $T_s < 1$ acts on the operator like repeated sharpening, pushing it toward a deterministic system: the subdominant eigenvalue
Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese
GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.30009, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txtSources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange