Monte Carlo Tree Search for Table-to-Multimodal Report Generation

How can AI systems generate professional, multimodal reports—consisting of both text and visual charts—from structured tabular data without falling into the traps of rigid, linear generation pipelines? The authors address the limitations of existing methods, which often suffer from "insight freezing" and poor cross-modal consistency because they process subtasks like table parsing, chart generation, and narrative writing in a fixed, isolated sequence.

The researchers propose MCTS-Report, a framework that reformulates report generation as a structured search problem. Instead of a single-pass pipeline, the system uses Monte Carlo Tree Search (MCTS) to explore a tree of potential report states.

• Search Tree: The process starts with an empty report and uses atomic actions (e.g., chapter planning, chart generation, insight organization) to build the report incrementally.
• Unified Engine: A single LLM acts as the engine for all actions, conditioned on the current state of the report, which helps maintain context across different sections.
• Self-Supervised Reward: To guide the search, the system evaluates completed drafts using a multi-dimensional reward function that checks numerical fact consistency (via SQL), chart quality, structural completeness, and insight novelty. This feedback is backpropagated through the tree to prioritize high-quality generation paths.

Experiments on the newly constructed MMRBench—a dataset of 185 real-world tables across six domains—demonstrate that MCTS-Report significantly outperforms 12 baseline models. By enabling backtracking and global optimization, the framework achieves superior results in structural completeness, numerical accuracy, and chart-text alignment. The ablation studies confirm that the MCTS planning mechanism and the self-supervised reward feedback are critical to the system's performance, effectively reducing the errors common in traditional, non-iterative approaches.