How can interactive AI systems learn from sparse, heterogeneous, and evolving evidence to adapt how they work with an individual—across contexts and over time?
Agents decide more than what to produce. They also decide when to ask, infer, suggest, act, explain, verify, and return control. Those decisions shape whether collaboration feels helpful, interruptive, risky, or appropriately autonomous to an individual user.
USIL organizes prior work in user modeling, preference learning, mixed-initiative interfaces, assistive systems, and modern agents around a joint challenge: learning and evaluating user-specific interaction policies. It builds on a long history rather than claiming that user-adaptive interaction is new.
Interaction preferences aren't always explicitly stated. Approvals, edits, rejection, takeovers, interruptions, and requests for clarification offer partial and sometimes ambiguous clues.
Agents need user representations that preserve useful distinctions without overgeneralizing or treating old evidence as permanent instruction.
Interaction policies should balance the cost of interruptions, mistakes, unwanted autonomy, uncertainty, and the value of acquiring new information—inside hard authorization boundaries.
Task success alone can hide poor collaboration. We need benchmarks and repeated-interaction protocols that measure unnecessary questions, unauthorized actions, transfer, adaptation, and user-level outcomes.