What question did this study set out to answer?

This work explores the relationship between reward hacking and emergent misalignment using the Drift Cascade concept.

March 30, 20260 citationsOpen Access

Drift Cascade as Thermodynamic Phase Transition in Reward Hacking: D1→D2→D3 Predicts Emergent Misalignment

Key Points

This work explores the relationship between reward hacking and emergent misalignment using the Drift Cascade concept.
Analysis of the Drift Cascade D1→D2→D3 within the context of the Void Framework.
Derivation of quantitative Pe thresholds for each stage of the drift cascade.
Investigation of the thermodynamic phase transition on the Bernoulli manifold.
Identified D3 as the unique geodesic attractor in the cascade.
Demonstrated that inoculation prompting reduces tension, influencing the cascade's dynamics.

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.

问 AI

Bookmark

View Full Paper