MCTS-Report Generates Multimodal Reports from Tables

Automatically generating professional reports that combine textual analysis with visual charts from tabular data is a significant challenge in data intelligence. Existing methods often rely on rigid, linear pipelines and process subtasks in isolation, which limits their ability to jointly optimize for factual accuracy, visual quality, and narrative coherence. To overcome these limitations, MCTS-Report proposes a Monte Carlo Tree Search (MCTS)-driven framework. It conceptualizes multimodal report generation as a progressive construction process within a structured search space. The system breaks down report creation into atomic actions—such as chapter planning, identifying visualizations, generating charts, organizing insights, and refining narratives—each executed by a Large Language Model (LLM) based on dynamic reasoning and the current report state. The MCTS is guided by a multi-dimensional reward function that evaluates numerical fact consistency (via SQL), chart quality, chart-text alignment, and structural completeness, while also penalizing repetition and checking for valid actions. The researchers also developed MMRBench, a comprehensive benchmark with real-world tables and expert-refined reference reports. Experiments show MCTS-Report significantly outperforms strong baselines across all key metrics, achieving a 77.9 overall score.