A Computer Vision Foundation / CVPR 2026 Open Access paper introduces MACT, a four-agent framework for visual document understanding and reasoning. The authors report average gains of 9.9% to 11.5% over base models on document-understanding benchmarks, while the associated GitHub repository says code will be release...
The Computer Vision Foundation’s CVPR 2026 Open Access page has published a paper introducing MACT, a multi-agent framework for visual document understanding and reasoning.
The paper, titled “Visual Document Understanding and Reasoning: A Multi-Agent Collaboration Framework with Agent-Wise Adaptive Test-Time Scaling,” presents MACT as a system designed for reasoning over visual documents. According to the Computer Vision Foundation / CVPR 2026 Open Access listing, MACT uses four cooperating agents: a planning agent, an execution agent, a judgment agent, and an answer agent.
The arXiv abstract page for the same work identifies MACT as a multi-agent collaboration framework with agent-wise adaptive test-time scaling for visual document understanding and reasoning. The arXiv record says the paper was submitted on August 5, 2025, and revised on November 14, 2025.
Visual document understanding typically involves extracting and reasoning over information from materials such as forms, reports, tables, charts, scanned pages, or document images. The supplied source material does not enumerate the exact benchmark names or document types used in the experiments, so those details should be taken from the full paper rather than inferred from the listing.
The Computer Vision Foundation / CVPR 2026 Open Access page says the MACT paper reports 9.9% to 11.5% average gains over base models on document-understanding benchmarks. The source excerpt does not specify whether those gains are absolute or relative, nor does it list the base models in the provided material.
That distinction matters for readers evaluating the result. Average improvements can vary substantially depending on benchmark composition, baseline strength, evaluation metric, and whether the same model family is used across comparisons. Based on the available source text, the most accurate description is that the authors report average gains in that range over base models, without further qualifying the evaluation setup beyond “document-understanding benchmarks.”
MACT’s central design choice, as described by the CVPR Open Access listing, is to separate document reasoning into four roles. A planning agent is associated with planning the reasoning process, an execution agent with carrying out steps, a judgment agent with evaluating intermediate work, and an answer agent with producing the final response.
The arXiv page further frames the approach around agent-wise adaptive test-time scaling. In general terms, test-time scaling refers to allocating additional computation during inference rather than only during model training. The source excerpt does not provide implementation details for how MACT adapts computation across agents, so the mechanism should not be described more specifically without consulting the full paper.
A linked GitHub repository under YU-deep/MACT exists for the project. According to the repository excerpt, the code “will be released after acceptance.” That means, based on the provided source material, public code was not yet available from that repository at the time described by the source excerpt.
For researchers and practitioners, the repository status is an important caveat. The reported gains are documented in the paper listing, but independent reproduction will depend on access to the implementation, model settings, benchmark configuration, and evaluation scripts.
The MACT paper adds to growing research interest in structured, multi-component reasoning systems for document AI. The Computer Vision Foundation / CVPR 2026 Open Access listing credits the method with 9.9% to 11.5% average gains over base models on document-understanding benchmarks, while the arXiv page records the work as a 2025 submission and revision.
The next practical milestone is code availability. Until the YU-deep/MACT repository releases implementation details, the paper’s claims are best read as reported research results rather than independently verified performance guidance.
The Computer Vision Foundation’s CVPR 2026 Open Access page has published a paper introducing MACT, a multi agent framework for visual document understanding and reasoning.
According to the Computer Vision Foundation / CVPR 2026 Open Access listing, MACT uses four cooperating agents: a planning agent, an execution agent, a judgment agent, and an answer agent.
The arXiv abstract page for the same work identifies MACT as a multi agent collaboration framework with agent wise adaptive test time scaling for visual document understanding and reasoning.
Continue reading
