Selected from more than 17,000 submissions, the study lets AI resume its reasoning from the point where an expert makes a correction
A research team led by Professor Dongwoo Kim of the Department of Computer Science and Engineering and the Graduate School of Artificial Intelligence at POSTECH received the Outstanding Paper Award at EMNLP 2026 (Conference on Empirical Methods in Natural Language Processing). The international conference on natural language processing took place in Budapest, Hungary, and the award was presented on October 27 (local time). This is the first time a team made up entirely of researchers from Korean institutions has received the award.
Researchers, developers and other experts now often solve problems by working back and forth with large language models (LLMs). This collaboration has clear limits. When a user points out an error in the AI’s answer, the model may ignore the correction, change parts it had already solved correctly, or make new mistakes. The cost also rises as the conversation goes on, because the model has to process the growing conversation history again with every turn.
The team approached the problem by treating feedback as a direct edit to the answer rather than as a new instruction. In the method they propose, called “in-place feedback,” the user corrects the wrong part of the AI’s response directly, and the AI picks up its reasoning from that corrected point. It works much like a teacher marking an incorrect line of a solution in red pen: the student continues from that line instead of starting over.
The team tested the approach on five benchmarks that require reasoning in mathematics, science, coding and logic. In-place feedback performed better than conventional conversational feedback while using fewer tokens. In a user study with 15 participants who had experience writing academic papers, participants were more satisfied with the final results and reported less fatigue from using the AI. The study shows that building trust between AI and human experts takes more than better model performance. It also requires interaction designs that let the AI correctly carry forward a person’s expertise and the intent behind each correction.
“This award recognizes that it matters how AI interacts with users, not only how well the model performs. Interaction should accurately reflect users’ expertise and the intent behind their corrections,” said Professor Kim. “I am grateful to the students and all the members of our team for their hard work. We will continue to look for ways for humans and AI to work together with greater trust.”
EMNLP and ACL are regarded as the two most prestigious conferences in natural language processing. This year, 17,661 papers were submitted to EMNLP, and only 22 (about 0.1%) received the Outstanding Paper Award.