A new arXiv paper introduces HABIT-Bench, an evaluation framework for large language model (LLM) agents. The benchmark focuses on habit induction—the process by which agents infer users' recurring preferences and behaviors from long-term interactions.
The authors note that in extended use, users reveal habits through their choices and actions. These habits capture implicit user information that agents must infer across conversations. HABIT-Bench is designed around "longitudinal weak evidence," meaning individual signals may be subtle or ambiguous, but together they reveal consistent patterns.
The work highlights a challenge for agent memory: accumulating and reasoning over scattered evidence over time. While the abstract does not detail specific results, the benchmark's purpose is to measure and improve this capability in LLM agents.