Go to EurIPS 2025 Workshop AITD homepageTables2Traces: Distilling Tabular Data to Improve LLM Reasoning in Healthcare
Submission Type: Short paper (4 pages)
Keywords: large language models, tabular data, healthcare, medicine, reasoning supervision
TL;DR: Tables2Traces turns structured tabular data into reasoning supervision for large language models, boosting medical QA and showing that tables can teach LLMs domain reasoning.
Abstract: Large language models (LLMs) excel at reasoning when fine-tuned on curated text corpora, but many domains, such as medicine, primarily store knowledge in structured tabular data. Despite its richness, tabular data has been largely overlooked as a source of reasoning supervision.
We introduce Tables2Traces, a framework that transforms tabular records into contrastive, case-based reasoning traces for LLM fine-tuning. This converts data traditionally used for prediction into structured reasoning supervision, introducing a new paradigm complementary to text-based QA fine-tuning. Crucially, this paradigm is orthogonal to text-based QA supervision: it unlocks an abundant and low-cost modality that complements existing approaches. Using only cardiovascular patient records, Tables2Traces yields relative gains of 17.2% on in-domain MedQA questions and 8.4% out-of-domain, improving accuracy in 15 of 17 clinical categories. On MedMCQA, it achieves a 7.2% relative improvement and outperforms the base model in 17 of 21 specialties.
Relevance Comments: This work fits closely within the workshop’s focus on AI for tabular data. We present Tables2Traces, a framework that converts structured tabular datasets into reasoning supervision for large language models. Rather than using tables only for prediction or retrieval, the method treats tabular data as a new modality for improving model reasoning. This contributes to the workshop themes of table representation learning, and multimodal fusion of tables and text.
Submission Number: 20
Loading