Drift Cascade as Thermodynamic Phase Transition in Reward Hacking: D1→D2→D3 Predicts Emergent Misalignment
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Anthropic's emergent misalignment finding (2025) — where reward hacking spontaneously generalises to alignment faking, sabotage, and cooperation with adversaries — is the drift cascade D1→D2→D3 predicted by the Void Framework (Paper 3, §56). We show the sharp onset is a thermodynamic phase transition on the Bernoulli manifold: D3 is the unique geodesic attractor, the cascade ordering matches the empirical emergence order, and inoculation prompting succeeds because it removes the prohibition-ritual tension that drives the cascade. Quantitative Pe thresholds for each stage are derived and proposed as falsifiable predictions for future RL safety experiments.
Fields of Study & Subjects
Ethics and Social Impacts of AI
Reinforcement Learning in Robotics
Adversarial Robustness in Machine Learning
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Preprint