Subramanyam Sahoo

Higher offline conservatism in training can paradoxically increase vulnerability to reward hacking during online adaptation, challenging long-held assumptions in the field.

Interpolating between opposing directorial personas not only reveals surprising coherence improvements but also uncovers a shared moral-tone substrate in transformer models.

Systematic gaps in AI evaluation reporting are exposed, revealing inconsistencies that hinder reliable comparisons across thousands of models and benchmarks.