PolicyBank: Evolving Policy Understanding for Evolving Agents

This paper aims to advance the field of Machine Learning, specifically focusing on the alignment and reliability of autonomous agents. By enabling agents to evolve their understanding of policies through interaction, our framework aims to bridge the gap between often ambiguous, under/misspecified natural language instructions and true system requirements, potentially reducing systematic failures in production systems (e.g., customer service, regulatory compliance). However, we acknowledge that systems capable of dynamically refining their own constraint logic introduce a risk of policy drift, where an agent might internalize incorrect behaviors from noisy or malicious feedback. To mitigate this, PolicyBank utilizes a structured, human-readable memory format, ensuring that any evolution in the agent's policy understanding remains transparent and auditable by human overseers.