New survey maps how AI copilots learn and adapt to user preferences

A new comprehensive survey examines how AI copilots—systems designed to assist knowledge workers and developers—can be tuned to match what individual users actually want from them. The research consolidates fragmented work across AI personalization, human-computer interaction, and language model adaptation into a unified framework.

The personalization challenge#

As AI assistants become routine parts of daily work, making them genuinely useful requires more than general capability. Each person works differently, has different priorities, and prefers different interaction styles. The survey identifies preference optimization—a system's capacity to recognize, interpret, and respond to what users actually need—as central to making copilots effective and satisfying to use. Despite decades of work on intelligent assistants and algorithmic optimization, how these approaches combine specifically within AI copilots has remained largely unmapped.

A structured taxonomy#

The researchers introduce a three-phase taxonomy for how preference signals flow through AI copilots: pre-interaction phases where initial setup and context-gathering happen; mid-interaction phases during active use; and post-interaction phases where feedback refines future behavior. Each technique is assessed for its strengths, weaknesses, and practical implications for design.

Building aligned systems#

The framework emphasizes that effective copilots must be both persona-aware and end-to-end adaptive—capable of evolving throughout their deployment rather than remaining static after initial training. The survey consolidates insights needed to build user-aligned AI assistants that genuinely learn from how people actually interact with them.