← Back to Machine Learning (Statistics)
stat.ML

Can we catch systems before they secretly fall apart?

David Mullett

May 29, 2026

Recursive systems (markets, recommenders, LLM training loops) can degrade invisibly while amplifying feedback until sudden failure. Loopzero is a benchmark framework that tests whether detectors can spot the telltale pattern: rising gain, persistent recursion, declining diversity—under a fixed false-alarm budget so all methods face equal constraints. On stock-market crashes and movie-recommendation systems, neither standard methods nor the proposed detector hit acceptable performance, but the directional pattern itself held on real data.
Published as Benchmarking Recursive-Collapse Warning Claims Under Matched False-Positive Control arXiv:2606.00329
Read the original paper →