Aixel

PROJECTS

Our Research

SIGMOD 2026To Appear

AixelNet

A pre-trained model for structured (tabular) data that adapts to new tables through table-level meta features and modular predictor composition. AixelNet summarizes table characteristics, selects and composes base predictors with a hypernetwork, and updates only small parts of the model when adapting.

  • Pre-trained Model for Structured Data Prediction
  • Table-level meta features for adaptation
  • Parameter-efficient adaptation via sparse update
SIGMOD 2026To Appear

AixelAsk

A framework for question answering over large tables that decomposes a question into a directed plan, retrieves evidence from key rows and columns, and performs step-by-step reasoning on the retrieved sub-tables to generate answers.

  • DAG-based question decomposition
  • Row/column retrieval to form sub-tables
  • Stepwise reasoning to produce answers