New Multi-Agent Automaton Sets State-of-the-Art in Visual Reasoning

Global: New Multi-Agent Automaton Sets State-of-the-Art in Visual Reasoning

A team of computer scientists has introduced a multi‑agent system designed to enhance the interpretability of vision‑language models while reducing hallucinations on complex queries. The work, posted on arXiv in January 2026, proposes a hierarchical finite‑state automaton named MATA (Multi‑Agent hierarchical Trainable Automaton) that coordinates several specialized agents to solve visual reasoning tasks.

Background and Motivation

Existing compositional approaches often rely on a single reasoning agent or on hand‑crafted pipelines, which limits their ability to dynamically select complementary agents or to manage overlapping capabilities. Researchers argue that such constraints hinder transparent decision‑making and scalability in multimodal AI systems.

MATA Architecture

The proposed framework structures reasoning as a hierarchy: a trainable hyper‑agent governs top‑level state transitions, while each subordinate agent corresponds to a specific state and operates a compact rule‑based sub‑automaton for micro‑control. All agents share a common memory space, enabling a transparent execution history that can be inspected after inference.

Training Approach

To teach the hyper‑agent how to select appropriate transitions, the authors constructed transition‑trajectory trees and converted them into memory‑to‑next‑state pairs, creating the MATA‑SFT‑90K dataset for supervised fine‑tuning. A large language model (LLM) was then fine‑tuned on this dataset to serve as the transition policy, allowing it to assess both the query and the capabilities of each agent before making a selection.

Benchmark Performance

Across several visual reasoning benchmarks, MATA achieved results that surpass both monolithic models and prior compositional baselines, according to the authors’ evaluation. The reported improvements suggest that dynamic agent selection can contribute to higher accuracy and more reliable reasoning outcomes.

Open Resources

The codebase and the MATA‑SFT‑90K dataset have been released publicly on GitHub, providing the research community with the tools needed to replicate and extend the experiments.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.