Model changes can alter large language agent behavior without warning

Towards Identity-Preserving LLM Agents under Model Transition

Artificial IntelligenceComputers and SocietyHuman-Computer Interaction

Summary

When people upgrade or swap the AI behind a chat agent, they usually expect it to act the same as before. The authors found that this is not always true—changing the AI model can subtly change how the agent behaves, even if its memory and personality stay the same. These differences don’t show up when the agent just talks about its rules, but real decisions can vary. So, it's important to test an AI agent's actions after switching models, not just assume everything remains the same.

What this means in practice

  • For ai operations teams: Validate behavior of AI agents through thorough testing after backend model upgrades to avoid unnoticed compliance risks.
  • For compliance officers: Monitor AI agents’ decision-making changes when switching models to ensure adherence to regulatory standards.

Authors

Yihuang Kang, Pei-Cing Huang, Chan-Tung Ku, Chan Hsu, Pei-Ju Lee, I-Ling Cheng

Abstract

When an LLM agent's backend model is upgraded or replaced, operators typically reload the same persona, memory, and rules onto the new model and assume the agent will behave as before. We show this assumption requires empirical validation-and that whether identity is preserved depends on the specific model pairing. Across 5,760 compliance-testing runs in 12 regulated domains with three open-weight models, we find that some transitions produce significant behavioral drift while others do not, even when the agent's external state remains identical. Crucially, this drift is invisible to verbal verification: every model claims full policy compliance, yet actual decisions under pressure differ across models. The same external state can reorder which model is most compliant-a large positive effect on one substrate, negligible on the other two. Practitioners should treat model transitions as behavioral changes requiring action-level empirical testing, not as transparent infrastructure operations.