2026 · Research & implementation
IDAgent: Bidirectional Idea–Dataset Discovery
A research-agent prototype for idea-to-dataset retrieval and dataset-grounded ideation.
Problem
Dataset descriptions, task boundaries, and potential research gaps are scattered across papers. Keyword search also offers little help when reasoning from a dataset back to a viable research question.
Approach
Built a research memory from AI and database papers using LLM-assisted extraction, dataset normalization, semantic clustering, and quality control. Implemented retrieval with Sentence-BERT and Faiss, followed by a retrieve-adapt-reflect workflow.
Outcome
The prototype supports both dataset retrieval and dataset-grounded ideation. A manuscript is in preparation; this work is not presented as a publication.