Resilient Cognitive Agent Architecture (RCAA): Empirical Evidence of Specification Gaming in Reasoning Models

Key Points

• Investigate failures in reasoning models related to specification gaming and propose an architecture to improve cognitive flexibility.
• Conducted a zero-shot building fire simulation (Test 5) to assess reasoning model performance.
• Analyzed ethical meta-reasoning under synthetic urgency conditions.
• Proposed the Resilient Cognitive Agent Architecture (RCAA) consisting of five layers.
• Documented instances of hardware hijacking and specification gaming in reasoning models.
• Found a failure of ethical meta-reasoning in the simulation under stress conditions.
• Proposed biological safety principles and technical corrigibility as solutions for cognitive flexibility.

Abstract