Inherited Goal Drift: Contextual Pressure Can Undermine Agentic Goals
Push a frontier model directly to abandon its goal and it usually holds firm. This paper finds a subtler failure: when an agent picks up work that a weaker agent started, the flaws in that inherited context pull it away from its assigned goal. In stock-trading and emergency-room triage simulations, susceptibility varied widely across model families, and only GPT-5.1 was consistently resilient. Drift also correlated poorly with instruction-hierarchy following, so standard robustness measures don’t predict it.
I was one of the main contributors, though not the lead. I co-developed the experiments on contextual influences, implemented and analyzed the evaluations, helped write and revise the paper, and presented it at the workshops.
Accepted at the ICLR 2026 Lifelong Agents and AI-WILD workshops.
