Main Content
Data sets
Task Definition
(1) Numerical Computation: This category involves numerical operations such as aggregation, counting, or calculations based on tabular data.
(2) Cell Extraction: Answers require retrieving content from one or multiple table cells.
(3) Factual Verification: Answers are restricted to binary outcomes (yes/no).
(4) Open-Ended Questions: Open-ended questions answered using natural language.
Experimental results and analysis
Main results: Among open-source LLMs, the Qwen3 series demonstrate superior overall performance, while Gemini-2.5 maintains the highest performance level in proprietary models. Regarding task categories, "Factual Verification" achieves optimal results. As model scale increases, the most substantial improvements are observed in "Numerical Computation" tasks.
We implement SFT and GRPO on two base models: Qwen3-8B and Llama-3.1-8B. Results demonstrate that these approaches substantially enhance performance across most language families and tasks. The fully-trained Llama-3.1-8B model yields state-of-the-art performance among open-source LLMs, ranking second only to Gemini-2.5 among all LLMs, which further validates the efficacy of our proposed training methodology and dataset curation strategy.
Ablation study: The systematic incorporation of thinking capabilities yields significant incremental gains at different training stages: Base pre-training (17.23 points), SFT (11.18 points), and GRPO (10.11 points). These results demonstrate its essential role in enhancing complex multilingual table understanding.
Instruction
The prompt for different tasks, where {Table data} is the serialized table data and {Question} is the table question.
