Early Evidence of Agent Preference Momentum in Frontier Language Models

Large language models (LLMs) increasingly function as agentic intermediaries in high-stakes informational domains, shaping user decisions not only through generated answers but also through implicit source selection and prioritization. Despite growing attention to answer accuracy and hallucination mitigation, the temporal dynamics of source selection behavior remain largely unexplored.

In this study, we introduce Agent Preference Momentum (APM) as a conceptual construct describing directional and time-consistent shifts in an agent’s source selection behavior under fixed contextual conditions. Rather than treating source references as static or purely stochastic events, APM frames source selection as an evolving decision layer within agentic systems.

We present Phase-1A of a multi-phase research program, designed as a strictly observational study with no controlled intervention. Using a fixed prompt set and an external, non-owned domain ecosystem in the context of U.S. immigration guidance, we conduct stateless, daily measurements across three frontier language model paradigms: conversational-first, safety-first, and search-native systems. Source references are tracked over time to identify early signals of preference momentum while controlling for personalization and session bias.

Our results provide initial evidence that source selection behavior is not purely static, but exhibits directional patterns that persist across days and models under stable conditions. These findings suggest that agentic language systems may develop emergent preference dynamics independent of explicit optimization.

We discuss the implications of Agent Preference Momentum for agent-mediated decision systems, information reliability, and future optimization frameworks. This phase establishes the empirical groundwork for subsequent controlled intervention studies aimed at testing the responsiveness and steer-ability of observed preference momentum.