Mitigating Structural Hallucinations in Complex Table Recognition via Post-Training
Abstract
Vision-language models often struggle with complex tables containing merged cells, frequently hallucinating incorrect row and column spans and thereby producing structurally inconsistent table representations. In this work, this work investigates post-training strategies for improving span-level structural fidelity in table recognition. Starting from PaddleOCR-VL 1.6, this paper progressively applies supervised fine-tuning (SFT), direct preference optimization (DPO), and GSPO to reduce rowspan and colspan errors. Experiments on the datasets show consistent improvements in table structure recognition, with particularly strong gains on tables involving complex cell spanning patterns. The results demonstrate that post-training provides an effective way to mitigate span-related structural hallucinations in complex table recognition.
Then back it, or bet against it.
Related papers
Open the market on this paper to see 7 more related papers.