Researchers introduce HAMR, a machine-learning technique that fixes a widespread problem in AI training — when datasets have far fewer examples of rare categories, making AI models ignore the minority cases

Researchers at an academic institution published HAMR (Hardness-Aware Meta-Resample), a new training method for natural-language AI systems that solves class imbalance — the problem where datasets have 100 examples of one category but only 5 of another, causing AI to ignore the rare ones entirely.

HAMR uses two strategies: it identifies which training examples are genuinely hard to learn (rather than just upweighting all minority examples blindly), and it groups semantically similar examples together so the model learns patterns from hard cases and their neighbors — the practical effect is that minority-class accuracy improves substantially while not sacrificing performance on common categories.

For any team building NLP systems on imbalanced real-world data — biomedical text classification, disaster-response message routing, or sentiment analysis where negative feedback is rare — this approach means better performance on the edge cases that actually matter (like identifying medical emergencies in a sea of routine notes), without requiring expensive manual data balancing.