Machine learning researchers close a 25-year-old math problem about learning from bad data

Researchers proved that using randomized guesses instead of fixed ones can cut error rates in half when algorithms learn from deliberately corrupted data. In practice, this means AI systems designed to handle adversarial attacks or intentionally poisoned training data could be significantly more accurate than previously possible.

This is a pure math result, not a deployed system. The paper answers three specific theoretical questions that have sat open since the late 1990s and early 2000s, filling gaps that recent work (2025–2026) had highlighted. What's interesting to a non-expert is the basic insight: if you're trying to learn from data you can't trust, occasionally making random decisions instead of deterministic ones actually helps. The practical implications are narrow—this matters most for AI systems explicitly designed to survive poisoned training data or adversarial attacks—but the closure of a 25-year problem in learning theory is the kind of foundational work that sometimes becomes important later when engineers realize they need exactly this property.