Tables Decoded: DELTA for Structure, TARQA for Understanding
The paper argues table AI can move off images and into compact text sequences without losing structure.
DELTA splits table parsing into physical structure, logical structure, and OCR, then emits an OTSL representation carrying layout and cell text. TARQA fine-tunes an LLM on those OTSL sequences for table QA. The authors report TSR results comparable to state of the art, plus 9.3-point gains on WTQ and 9.2-point gains on FinTabNetQA. They also introduce a Hindi benchmark, TORQUE, where the method ranks second among tested VLM and DELTA-plus-LLM variants. ArXiv · AI/CL/LG's note
DELTA splits table parsing into physical structure, logical structure, and OCR, then emits an OTSL representation carrying layout and cell text. TARQA fine-tunes an LLM on those OTSL sequences for table QA. The authors report TSR results comparable to state of the art, plus 9.3-point gains on WTQ and 9.2-point gains on FinTabNetQA. They also introduce a Hindi benchmark, TORQUE, where the method ranks second among tested VLM and DELTA-plus-LLM variants. ArXiv · AI/CL/LG's note
score 4