Investigate LLMs' capabilities in solving knowledge-intensive math word problems.

Propose a new evaluation benchmark for numerical reasoning in financial documents containing both textual and tabular data.

Propose a new query-focused table summarization task where text generation models have to perform human-like reasoning and analysis over the given table to generate a tailored summary.

LLM outperforms fine-tuned systems in table-to-text generation, evaluation, and feedback generation.

Analyze robustness of TableQA systems using a human labeled evaluation set of targeted adversarial perturbations.

Apply logical form as fact checker and content planner to improve faithfulness and text-diversity of logical table-to-text generation.

Inject table reasoning skills into large language models by synthesizing corresponding QA examples as pre-training corpus.

Construct QA benchmark that requires model to aggregate numerical information from text and hierarchical tables to answer complex questions about financial reports.