Researchers have developed TELLER, a novel approach to table entity linking that iteratively optimizes based on model errors and reasoning. This method addresses limitations in static training data by refreshing preferences with residual errors from the evolving model. TELLER includes a direct-answer path that refines predictions and a reasoning path that uses filtered chain-of-thought rationales for supervised fine-tuning. The system demonstrated improvements in accuracy on benchmarks like TableInstruct and MammoTab V2, enhancing both concise entity prediction and explicit reasoning. AI
IMPACT This research introduces a more adaptive training methodology for entity linking models, potentially improving accuracy and robustness in structured data interpretation.
RANK_REASON Academic paper detailing a new method for table entity linking. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
