AI-powered coding assistants are increasingly turning informal user requests into working software, but these requests often contain ambiguities that repeat in user-specific ways across tasks and sessions. Current disambiguation methods typically handle each ambiguous request in isolation, often by asking for more clarification. However, whether a user's resolved session history can serve as memory for resolving similar ambiguities in new sessions has been largely unexplored.
To address this gap, researchers from Hugging Face introduce CAPA (Cross-session Personalized Ambiguity Adaptation), a new benchmark that formalizes the task of personalized ambiguity adaptation. Given a user's previously resolved coding sessions and a new ambiguous request, an assistant should identify the recurring ambiguity pattern, produce the intended executable solution, and minimize clarification. CAPA characterizes personalized coding ambiguity through six mechanisms and injects these into unambiguous executable tasks using a controlled three-stage generation pipeline.
The benchmark comprises 600 coding sessions across 60 balanced user–ambiguity cells, including 300 held-out evaluation sessions. The researchers evaluated 12 recent LLMs under both no-history and same-user-history conditions, measuring executable success, first-turn success, and turns-to-completion. They also analyzed task difficulty, user identity, and memory-based history use, and proposed a lightweight inference-time method called same-user history gating.
This work lays a foundation for developing long-term coding assistants that better align generated code with user intent while reducing repeated clarification. The paper is available on Hugging Face and has sparked discussion about the practical challenges of building durable memory layers and handling preference drift over time.