Efficient Test-Time Adaptation through Human-AI Interaction

Efficient Test-Time Adaptation through Human-AI Interaction introduces TAHI, a framework that learns a user’s tacit preferences from ongoing collaboration rather than relying only on pretraining or one-off feedback. Built on the Qwen3.6-35B-A3B agent backbone, with Anthropic’s claude-sonnet-4-6 supporting context induction and rubric evolution, TAHI captures four interaction channels: messages, plan edits, direct artifact edits, and verification-criterion changes. It adapts through either editable memory and procedural skills or Direct Preference Optimization on LoRA adapters, training the agent to produce the user-preferred result in one pass. Across 30 individuals and 600 tasks covering paper abstract writing and HTML data visualization, adaptation improved solo task success by 4.5–20.9% within only 20 task sessions; the largest gain was 20.9% for weight adaptation on visualization. On held-out tasks, the learned preferences generalized across tasks, while weight adaptation reduced input-token requirements by 62.6% for writing and 88.3% for visualization relative to context adaptation. TAHI also evolves evaluation rubrics from interaction traces: these rubrics captured 16.0–22.3% more failures than rubrics produced by language models or humans alone. The results show that agents can absorb both personalized strategies and shared community standards, although implicit visual judgment, especially color and emphasis, remains difficult to transfer.