TELLER: Dual-Path Iterative Preference Optimization for Table Entity Linking

Name: TELLER: Dual-Path Iterative Preference Optimization for Table Entity Linking
Uploaded: 2026-10-03T13:29:01.391739+00:00
Duration: 365 s
Description: TELLER is a dual-path framework for table entity linking that learns iteratively from model errors and reasoning to improve accuracy. It uses a direct-answer path with iterative DPO and a reasoning path with CoT-SFT followed by L-RPO. This approach boosts accuracy on table entity linking tasks, achieving high performance metrics on both direct prediction and detailed reasoning.

TELLER is a dual-path framework for table entity linking that learns iteratively from model errors and reasoning to improve accuracy. It uses a direct-answer path with iterative DPO and a reasoning path with CoT-SFT followed by L-RPO. This approach boosts accuracy on table entity linking tasks, achieving high performance metrics on both direct prediction and detailed reasoning.

Open the full episode