Undergraduate student wins excellent paper award at International Conference on Intelligent Computing
Jang Ye-rin, a Sejong University student majoring in artificial intelligence data science, presents her paper at the 21st International Conference on Intelligent Computing held in Prague, the Czech Republic, from July 3-6. Courtesy of Sejong University
Research team proposes retrieval-based in-context learning framework
A Sejong University student won an excellent paper award at the 21st International Conference on Intelligent Computing (INTCOM 2026) held in Prague, the Czech Republic, from July 3-6.
The university said Friday that Jang Ye-rin, a student majoring in artificial intelligence data science, earned the award for her paper, “AmbICL: Retrieval-Augmented In-Context Learning for Ambiguous Robotic Command Resolution.”
INTCOM is an international academic conference in the field of information and communication technology, organized by the Korean Society for Internet Information.
Jang was a member of a research team led by Gu Yeong-hyeon, a professor in the Department of Artificial Intelligence Data Science at Sejong University.
She was the paper’s first author, with Gu and doctoral student Piao Xianghua serving as co-authors.
“It is especially meaningful to achieve such a positive result at an international academic conference as an undergraduate student,” said Jang.
She added, “I will use this award as motivation to continue conducting in-depth research that can make a meaningful and practical contribution to the field of artificial intelligence.”
The research team proposed a retrieval-based in-context learning (ICL) framework called AmbICL, which enables robots to interpret ambiguous or context-dependent natural-language commands and convert them into executable commands without requiring additional model retraining.
AmbICL uses maximal marginal relevance (MMR), a retrieval technique that searches a database for examples that are both semantically similar to the current command and diverse from one another.
The retrieved examples, along with images of the environment captured by the robot, are then fed into a vision-language model, enabling the system to translate ambiguous natural-language expressions into commands that the robot can execute.
Experimental results showed that the proposed method achieved approximately 2.1 times the performance of conventional approaches that interpret commands without using additional examples.
It also achieved a text similarity score in the 90 percent range, demonstrating strong semantic similarity between the generated commands and the reference sentences.
The university said the MMR-based retrieval approach, which considers not only semantic similarity but also the diversity of retrieved results, outperformed a simple similarity-based retrieval method.
The results experimentally demonstrate the importance of designing effective retrieval strategies for interpreting robotic commands.
Conventional approaches have been limited by the need to collect new data and retrain models whenever robots are deployed in new environments.
In contrast, AmbICL can leverage retrieved examples and information from the robot’s surroundings without additional model training, making it well suited for rapid deployment across a variety of service robot environments.