Can Hyperbolic Fine-Tuning Close the Gap Between What LLMs Say and Do? A Null Result

Caltech WAVE Fellows Program · 2026

After alignment training, large language models (LLMs) give consistent, human-like answers when asked to rate their own personality, but those self-reports do not predict how the models behave. Self-CTRL, a recent training method, tries to close this gap by rewarding self-reports that match behavior. I asked whether the geometry of the space in which a model is fine-tuned limits how well this can work.