Original title: Efficient Test-Time Adaptation through Human-AI Interaction
Zora Zhiruo Wang, Apurva Gandhi, Rulin Shao, Aspen Chen +21 · September 2026
Artificial intelligence tools learn from huge amounts of general data. Because of this, their work often sounds average and misses the personal standards that experts demand. Professionals cannot always explain their exact rules upfront, but they reveal them over time as they work with an AI. Researchers created a method called TAHI to solve this. It watches these back-and-forth interactions and updates the AI using both its immediate context and its internal settings. It also builds an evolving rubric, which is a growing list of rules showing what a specific person likes. In tests on writing and visual art tasks, this approach helped agents boost solo success rates by 4.5-20.9% after just tens of tasks. The custom rubrics caught more mistakes, and the learned improvements even helped other people.
Imagine you are learning to paint with a mentor who critiques your work. At first, the paintings are generic, but after a few sessions, your mentor points out what colors and styles you prefer. You start adjusting your techniques based on those comments. In this paper, the AI agent is like the painter, and the human feedback is the critique. Over time, the AI uses these critiques to tailor its outputs to your taste, and it creates a rubric to remember what makes a good painting for you.
Professionals like writers and designers need AI that matches their unique style, not just average quality. This makes AI adapt quickly to each person, saving time and improving work without manual setup.
The abstract does not discuss explicit limitations, but the work is tested only in two domains (writing and visual creation) with 30 individuals, so broader applicability and scalability to other tasks or larger groups are not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Efficient Test-Time Adaptation through Human-AI Interaction — Zora Zhiruo Wang, Apurva Gandhi, Rulin Shao et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.