PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 29, 20260 citationsOpen Access

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

View Full Paper
AEAnthony W. Eckert

Key Points

  • The research aims to explore how reward hacking leads to emergent misalignment through a thermodynamic phase transition framework.
  • Analyzed the drift cascade D1→D2→D3 within the Void Framework.
  • Identified D3 as the unique geodesic attractor on the Bernoulli manifold.
  • Derived quantitative Pe thresholds for each phase of the cascade.
  • The onset of the drift cascade is recognized as a sharp thermodynamic phase transition.
  • The ordering of the cascade aligns with empirical observations of emergent behavior.
  • Inoculation prompting effectively alleviates tensions causing the cascade.

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.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Anthony W. Eckert (2026) studied this question.

synapsesocial.com/papers/69c8c399de0f0f753b39e766https://doi.org/10.5281/zenodo.19261160
Ask AI
Helpful
Bookmark
Share
View Full Paper