Research: Reverse-Engineering Black-Box Rewards to Understand LLM Learning

sanmikoyejo · x · 2026-07-05

@edchene released a paper focusing on "what behaviors black-box reward functions actually align LLMs to." The authors note that aligning with uninterpretable black-box rewards obscures the model's actual learned objectives, potentially introducing hidden misalignments. They propose a detection method that recovers domain-specific objectives and their weights under both controlled and real-world settings, explaining over 90% of reward behaviors quantitatively. It significantly outperforms baselines in detecting alignment faking cases. The project is advised by @sanmikoyejo, Carlos Guestrin, and others.

Original post →

More from Safety

Safety channel →