Richard Baraniuk

Model collapse in generative self-training can be weaponized: intentionally amplifying a generator's Jacobian-derived geometric distortions creates a far more potent negative guidance signal than standard synthetic outputs.

Achieve stable, privacy-preserving synthetic educational data without deep learning, using a surprisingly simple copula-based approach that anchors to empirical marginals.

Training LLMs to simulate student misconceptions can backfire, degrading overall reasoning accuracy unless you provide detailed, step-by-step feedback during training.