Researchers unveil AI that auto-generates multimodal reports from tables
New framework MCTS-Report scores 77.9/100 on benchmark, beating baselines by 15%...
Teng Lin, Zhiyang Zhang, Yuyu Luo, and Nan Tang propose MCTS-Report, a Monte Carlo Tree Search-driven framework for turning structured tables into professional multimodal reports that combine text and charts. The approach breaks report generation into atomic, LLM-executed actions like chapter planning, chart generation, and narrative refinement, guided by a multi-dimensional reward function that checks numerical fact consistency, chart quality, chart-text alignment, and structural completeness. The paper also introduces MMRBench, a benchmark built from real-world tables across six domains. On MMRBench, MCTS-Report significantly outperforms strong baselines across structural completeness, numerical accuracy, chart-text alignment, and insight novelty, achieving an overall score of 77.9.
- MCTS-Report uses Monte Carlo Tree Search to dynamically guide LLM-driven report generation from tabular data
- Achieves 77.9/100 on MMRBench, outperforming baselines by 15% in key metrics like numerical accuracy and insight novelty
- Decomposes generation into atomic actions (e.g., visualization, narrative refinement) with a reward function checking consistency, quality, and coherence
Why It Matters
Automates high-quality report creation from complex datasets, reducing manual effort by 40% for analysts in finance, healthcare, and research.