Researchers propose MCTS-Report, a Monte Carlo Tree Search framework for generating multimodal reports from tabular data

Researchers from an unnamed institution propose MCTS-Report, a Monte Carlo Tree Search (MCTS)-driven framework designed to generate professional multimodal reports—combining textual analysis and visual charts—from structured tabular data. The core innovation is to model report generation as a progressive construction process over a structured search space, decomposing the task into atomic actions such as chapter planning, visualization task identification, chart generation, insight organization, and narrative refinement.

Each action is executed by a large language model (LLM) based on dynamic reasoning conditioned on the current report state. During MCTS, the LLM generates step-by-step reasoning and actions, storing the reasoning trajectory in each node to ensure context-aware and coherent report construction.

To guide the search process, the authors design a multi-dimensional reward function that jointly evaluates numerical fact consistency (using SQL), chart quality, chart-text alignment, and structural completeness. The reward function also incorporates a diversity penalty to suppress repeated charts and a precondition check to prune invalid actions.

The team introduces MMRBench, a benchmark dataset comprising real-world tables from six domains, each paired with expert-refined reference report structures and verifiable key insights. This benchmark is intended to standardize evaluation for multimodal report generation tasks.

In experiments 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. The authors attribute this performance to the framework’s ability to jointly optimize multiple dimensions of report quality through guided search and dynamic reasoning.