Mixture-of-Minds: Multi-Agent Reinforcement Learning for Table Understanding

TL;DR

Mixture-of-Minds tackles table understanding by splitting reasoning into planning, coding, and answering across three specialized agents, enabling robust inference with precise table manipulation. It introduces a self-improvement training framework that uses Monte Carlo Tree Search (MCTS) rollouts to generate pseudo-gold intermediate supervision and optimizes agents via Group Relative Policy Optimization (GRPO). Experiments on TableBench and FinQA show substantial gains, including surpassing OpenAI o4-mini-high with smaller LLMs and achieving a top performance of $62.13\%$ under test-time scaling on TableBench. Overall, the work demonstrates that structured multi-agent workflows combined with RL can significantly enhance reliability, interpretability, and scalability in table reasoning.

Abstract

Understanding and reasoning over tables is a critical capability for many real-world applications. Large language models (LLMs) have shown promise on this task, but current approaches remain limited. Fine-tuning based methods strengthen language reasoning; yet they are prone to arithmetic errors and hallucination. In contrast, tool-based methods enable precise table manipulation but rely on rigid schemas and lack semantic understanding. These complementary drawbacks highlight the need for approaches that integrate robust reasoning with reliable table processing. In this work, we propose Mixture-of-Minds, a multi-agent framework that decomposes table reasoning into three specialized roles: planning, coding, and answering. This design enables each agent to focus on a specific aspect of the task while leveraging code execution for precise table manipulation. Building on this workflow, we introduce a self-improvement training framework that employs Monte Carlo Tree Search (MCTS) rollouts to generate pseudo-gold trajectories and optimize agents with reinforcement learning (RL). Extensive experiments show that Mixture-of-Minds delivers substantial gains, reaching 62.13% on TableBench and surpassing OpenAI-o4-mini-high. These results demonstrate the promise of combining structured multi-agent workflows with RL to advance table understanding.

Figures (4)

• Figure 1: Overview of the Mixture-of-Minds workflow. The process is decomposed into three agents: (1) the planning agent, which outlines a structured plan; (2) the coding agent, which generates and executes code; and (3) the answering agent, which synthesizes the plan, table, and code outputs to produce the final answer.
• Figure 2: Illustration of the Mixture-of-Minds workflow with MCTS-style rollouts. Each agent produces multiple candidate inferences and we evaluate the final outputs to identify the correct answer. In this example, Answer $\gamma$ is judged as the correct answer, and we trace back to Code 2 and Plan 1—the steps that produced it—which are then marked as pseudo-gold intermediate supervision for training.
• Figure 3: Ablation studies across different models. Each subplot shows performance on Fact Checking (FC), Numerical Reasoning (NR), and Data Analysis (DA) under sequential training stages: Mixture-of-Minds w/o training, +Plan agent training, +Plan+Code agent training, and +Plan+Code+Answer agent training.
• Figure 4: Reward trajectories of the planning, coding, and answering agents on Qwen3-8B. Each plot illustrates how our designed reward functions guide learning over training.