EUVIMEDEUVIMED entdecken
Crossref

Selective Retroactive Answer Updating for RAG-Based Assistants in Evolving Information Environments

Tiziana Hernandez, Ian Rudy, Tomek Strzalkowski

Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment · 08.11.2026 · 22 (1) · 269-278

DOI: 10.1609/aiide.v22i1.43585 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, online erschienen am 08.11.2026. Verfasst von Tiziana Hernandez und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Most task-oriented dialogue systems and conversational AI function primarily as question answering tools, that is, the user poses one question at a time, and the system generates a response based on its own pretrained model or an external data repository. In both cases, the response is considered “final”, reflecting the information available at the query time. However, the answer may need to be subsequently revised when the underlying data changes in a meaningful way that makes the previous answer incomplete or invalid. This is particularly important in sustained human-AI interactions when a series of interlocking questions are part of a larger task. In this paper, we present the Ahil AI conversational agent specifically designed to assist analysts in completing complex information tasks, such as a situational assessment, news production, or judgment formation, while the available information is both fragmented and continuously revised and updated. Ahil AI supports the analyst by answering their questions while also tracking changes in the underlying data. When new information arrives, Ahil evaluates whether any earlier answers need to be revised and, if so, alerts the analyst. We evaluate the efficacy of this response revision process on a fictional geopolitical intelligence dataset with gradually released documents. The evaluation contains 126 annotated cases in which evidence from the newly released document entered the RAG context. Conditional on this successful retrieval, Ahil achieved 0.837 recall, 0.632 precision, and 0.720 F1 for identifying update-worthy prior answers. These results therefore measure Ahil’s ability to select meaningful updates after retrieval, rather than the end-to-end performance of the complete retrieval and updating pipeline. The results suggest that selective answer updating is a distinct challenge for RAG-based assistants: Ahil usually detects meaningful updates, but sometimes over-alerts when new evidence is relevant but not important enough to notify the analyst.

Zitieren

APA

Hernandez, T., Rudy, I., & Strzalkowski, T. (2026). Selective Retroactive Answer Updating for RAG-Based Assistants in Evolving Information Environments. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 22(1), 269-278. https://doi.org/10.1609/aiide.v22i1.43585

Vancouver

Hernandez T, Rudy I, Strzalkowski T. Selective Retroactive Answer Updating for RAG-Based Assistants in Evolving Information Environments. Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment. 2026;22(1):269-278. doi:10.1609/aiide.v22i1.43585

BibTeX
@article{Hernandez2026,
  author = {Hernandez, Tiziana and Rudy, Ian and Strzalkowski, Tomek},
  title = {Selective Retroactive Answer Updating for RAG-Based Assistants in Evolving Information Environments},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment},
  year = {2026},
  volume = {22},
  number = {1},
  pages = {269--278},
  doi = {10.1609/aiide.v22i1.43585}
}
Kostenlos starten →