Multi-agent visual reasoning method, device and equipment and storage medium

Abstract

The application provides a multi-agent visual reasoning method, device and equipment and a storage medium. The method comprises the following steps: formalizing a visual reasoning task into a Monte Carlo tree search problem in a Markov decision process framework; using a multi-agent collaborative framework, taking each to-be-processed image and an initial query as a root node state, iteratively performing selection, expansion, simulation and back propagation steps of the Monte Carlo tree search to solve the Monte Carlo tree search problem, and obtaining a constructed search tree; stopping the search when the number of iterations reaches a preset threshold, and using a global thinking agent in the multi-agent collaborative framework to generate a final answer of the visual reasoning task according to a plurality of terminal node states of the constructed search tree. The application suppresses cascade errors caused by single chain reasoning path through multi-path exploration, and avoids the inefficiency of wide search in tree reasoning through a reward back propagation mechanism.

Description

Technical Field

[0001] This invention relates to the field of visual reasoning technology, and in particular to a multi-agent visual reasoning method, apparatus, device, and storage medium. Background Technology

[0002] With the development of multimodal large models, visual language models (VLMs) have shown great potential in fields such as visual question answering, scene understanding, and agent decision-making, becoming a cutting-edge hot topic in artificial intelligence research. Currently, the research focus in this field is shifting from simple image-text matching to more complex visual reasoning, which requires models not only to identify salient objects in images but also to understand the spatial location, logical relationships, and dynamic evolution processes between objects.

[0003] However, existing models face significant challenges when handling fine-grained visual reasoning tasks. The core issue lies in the fact that mainstream models rely heavily on "image-topic description" labels during pre-training. These descriptions often focus on macroscopic descriptions of the global scene while neglecting local details. This results in extremely low sensitivity when dealing with small objects, distant targets, or blurred visual features, making them prone to distorting details or fabricating facts during reasoning. This limitation severely restricts the practical application of large-scale visual models in high-precision, high-reliability scenarios such as obstacle recognition in autonomous driving and accurate lesion localization in medical imaging. Furthermore, existing end-to-end methods are limited by the coverage of predefined tools, exhibiting insufficient generalization ability and robustness, making it difficult to handle complex and ever-changing real-world fine-grained tasks.

[0004] Therefore, an effective technical solution is urgently needed to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-agent visual reasoning method, apparatus, device, and storage medium. This invention proposes a Monte Carlo tree search-driven multi-agent collaborative reasoning method. In the exploration and optimization of visual reasoning paths, it integrates the selection, expansion, simulation, and backtracking mechanisms of Monte Carlo tree search with the planning, perception, and evaluation processes of multi-agent visual reasoning. It formalizes multi-agent collaborative interaction as a Markov decision process. This not only suppresses cascading errors caused by single chain-like reasoning paths through multi-path exploration but also avoids the inefficiency of broad search in tree-like reasoning through a reward backpropagation mechanism.

[0006] In a first aspect, the present invention provides a multi-agent visual reasoning method, the method comprising the following steps: A visual reasoning task is obtained, the visual reasoning task including at least one image to be processed and an initial query associated with each of the images to be processed; The visual reasoning task is formalized as a Monte Carlo tree search problem within the Markov decision process framework. Using a pre-built multi-agent cooperative framework, with each of the images to be processed and the initial query as the root node state, the Monte Carlo tree search steps of selection, expansion, simulation and backpropagation are iteratively executed to solve the Monte Carlo tree search problem and obtain the constructed search tree; the constructed search tree includes all possible inference paths explored and the evaluation value of each node in each possible inference path; The search stops when the number of iterations reaches a preset threshold. The global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0007] According to a multi-agent visual reasoning method provided by the present invention, the multi-agent collaborative framework includes a global thinking agent, multiple perception agents, and an evaluation agent; the method utilizes a pre-constructed multi-agent collaborative framework, taking each image to be processed and the initial query as the root node state, and iteratively executes the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem, obtaining a constructed search tree, including: For any round of search, in the selection step, the global thinking agent selects an existing reasoning path from the root node downwards based on the current state and the historical evaluation value of each child node. In the extended step, the global thinking agent generates at least one subtask based on the existing reasoning path and the current state, and selects at least one perceptual agent from the plurality of perceptual agents to execute the subtask in order to generate a new child node; In the simulation steps, the global thinking agent starts from the new child node and reasons based on the preset policy model until it reaches the terminal node, thus obtaining a complete reasoning trajectory. In the backpropagation step, the evaluation agent evaluates the state of each node in the inference trajectory and generates a reward signal. The reward signal is then used to update the evaluation value and access count of the nodes on the search tree to obtain the updated search tree corresponding to the round of search. Based on the updated search tree corresponding to the previous search round, determine the existing reasoning path corresponding to the next search round; Based on the existing reasoning path corresponding to the next round of search, the selection, expansion, simulation and backpropagation steps of Monte Carlo tree search are executed iteratively to obtain the updated search tree corresponding to the last round of search, and the updated search tree corresponding to the last round of search is determined as the constructed search tree.

[0008] According to a multi-agent visual reasoning method provided by the present invention, the plurality of perceptual agents are functional agents equipped with dedicated visual tools, and the plurality of perceptual agents include at least one of the following: a focusing agent for focusing on a specific local region from a global image, an expanding agent for expanding the cropping range, an environment-enhancing agent for enhancing the visual quality of the image, a label-understanding agent for parsing explicit visual labels in the image, a relationship-understanding agent for mining spatial and dynamic relationships between target objects, and a multi-image-understanding agent for processing association analysis of multiple images.

[0009] According to a multi-agent visual reasoning method provided by the present invention, the step of selecting an existing reasoning path downwards from the root node based on the current state and the historical evaluation value of each child node by the global thinking agent includes: The global thinking agent invokes an improved confidence interval upper bound algorithm to select the next child node of the current node. The improved confidence interval upper bound algorithm integrates the real-time confidence of the policy model for the candidate child node, the historical action value of the candidate child node, and the number of visits to the candidate child node. Based on the child nodes and the current node at each step, the existing reasoning path is selected.

[0010] According to a multi-agent visual reasoning method provided by the present invention, the method involves generating at least one sub-task through a global thinking agent based on the existing reasoning path and the current state, and selecting at least one perceptual agent from a plurality of perceptual agents to execute the sub-task to generate a new sub-node, comprising: The global thinking agent thinks based on the existing reasoning path and the current state to generate the sub-task for the current step. The global thinking agent selects and invokes the dedicated tools of the perceptual agents to execute the sub-task of the current step from the multiple perceptual agents, and generates the execution result as new state information. Based on the new state information, the new child nodes are extended.

[0011] According to a multi-agent visual reasoning method provided by the present invention, the step of generating the final answer to the visual reasoning task based on the states of multiple terminal nodes of the constructed search tree by utilizing a global thinking agent in the multi-agent collaborative framework includes: The states of the multiple termination nodes are input to the global thinking agent; The global thinking agent weighs the multiple complete reasoning trajectories corresponding to the multiple termination node states, integrates the evidence information of each reasoning trajectory, and derives the final answer.

[0012] Secondly, the present invention also provides a multi-agent visual reasoning device, which includes the following modules: An acquisition module is used to acquire a visual reasoning task, wherein the visual reasoning task includes at least one image to be processed and an initial query related to each of the images to be processed; The transformation module is used to formalize the visual reasoning task into a Monte Carlo tree search problem within the Markov decision process framework. The visual reasoning module utilizes a pre-built multi-agent collaborative framework, using each image to be processed and the initial query as the root node state, to iteratively execute the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem and obtain a constructed search tree. The constructed search tree includes all possible reasoning paths explored and the evaluation value of each node in each possible reasoning path. The search stops when the number of iterations reaches a preset threshold, and the global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0013] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-agent visual reasoning method as described above.

[0014] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-agent visual reasoning method as described above.

[0015] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-agent visual reasoning method as described above.

[0016] The multi-agent visual reasoning method, apparatus, device, and storage medium provided by this invention first acquire a visual reasoning task, which includes at least one image to be processed and initial queries related to each image. Then, the visual reasoning task is formalized as a Monte Carlo tree search problem within a Markov decision process framework. Further, using a pre-constructed multi-agent collaborative framework, with each image to be processed and the initial query as the root node state, the selection, expansion, simulation, and backpropagation steps of the Monte Carlo tree search are iteratively executed to solve the Monte Carlo tree search problem, resulting in a constructed search tree. The constructed search tree includes all possible reasoning paths explored and the evaluation value of each node in each possible reasoning path. Finally, the search stops when the number of iterations reaches a preset threshold, and the final answer to the visual reasoning task is generated using a global thinking agent within the multi-agent collaborative framework, based on the states of multiple terminal nodes in the constructed search tree.

[0017] This invention proposes a novel multi-agent fine-grained visual reasoning framework, decomposing the visual language model (VLM) into step-level task planning, fine-grained visual perception, and reasoning performance evaluation. It designs a global thinking agent, a tool-equipped perception agent, and an evaluation agent as plug-and-play modular components, enhancing the VLM's fine-grained understanding capabilities without additional training. Simultaneously, it proposes a Monte Carlo tree search-driven multi-agent collaborative reasoning method. In exploring and optimizing visual reasoning paths, it integrates the selection, expansion, simulation, and backtracking mechanisms of Monte Carlo tree search with the planning, perception, and evaluation processes of multi-agent visual reasoning. This formalizes multi-agent collaborative interaction as a Markov decision process, not only suppressing cascading errors caused by single chain-like reasoning paths through multi-path exploration but also avoiding the inefficiency of broad search in tree-based reasoning through a reward backpropagation mechanism. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the multi-agent visual reasoning method provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the multi-agent collaborative framework provided by the present invention.

[0021] Figure 3 This is a comparative diagram of several visual reasoning methods.

[0022] Figure 4 This is a schematic diagram of the structure of the multi-agent visual reasoning device provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, a first node can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0026] The following is combined with Figures 1 to 5 The present invention describes a multi-agent visual reasoning method, apparatus, device, and storage medium.

[0027] Figure 1 This is a flowchart illustrating the multi-agent visual reasoning method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: Step 101: Obtain the visual reasoning task, which includes at least one image to be processed and an initial query associated with each image to be processed.

[0028] The execution subject of the multi-agent visual reasoning method provided by the present invention can be an electronic device, or any subject that can implement the multi-agent visual reasoning method, such as other visual reasoning systems.

[0029] First, a visual reasoning task is acquired. This task requires the system to analyze, judge, and provide answers based on image content. Unlike simple image classification or description, visual reasoning typically involves a deep understanding of spatial relationships, logical relationships, and dynamic evolution processes between objects. For example, determining whether the current environment supports a left turn or recognizing the speed limit number on a distant traffic sign.

[0030] For example, a visual reasoning task includes at least one image to be processed and an initial query associated with each image. The image to be processed is the input image for the visual reasoning task, and can be single or multiple images. In embodiments of the relevant technical materials, the image may be a high-resolution street view, an endoscopic medical image, a series of frames from an autonomous driving system, etc. This invention supports two input modes: single image (such as a high-resolution image in V*Benchmark) and multiple images (such as a multi-frame driving scene). The initial query is an image-related question presented in natural language. The query can be about object attributes (color, shape, size), spatial relationships (orientation, distance), logical inference (whether an operation is allowed), or temporal understanding (dynamic changes in multiple frames of images).

[0031] This step defines the input format (image + query) for the visual reasoning task and uses it as the root node state of the Monte Carlo tree search, providing clear problem boundaries and initial evidence for subsequent multi-agent collaborative reasoning.

[0032] Step 102: The visual reasoning task is formalized into a Monte Carlo tree search problem within the Markov decision process framework.

[0033] Specifically, to achieve fine-grained visual reasoning, this invention formalizes the problem of a global thinking agent dynamically calling a perceptual agent as a Markov decision process. The state space of the global thinking agent is the state space of a Monte Carlo Tree Search (MCTS). By iteratively executing the four steps of MCTS—selection, expansion, simulation, and backpropagation—the MCTS is constructed and optimized, ultimately finding an approximately optimal agent calling sequence to complete the visual reasoning task. For a visual reasoning task, the global thinking agent performs N rounds of searching from the initial state node, obtaining at most N complete reasoning chains.

[0034] Step 103: Using a pre-built multi-agent collaborative framework, with each image to be processed and the initial query as the root node state, iteratively execute the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem and obtain the constructed search tree; the multi-agent collaborative framework includes a global thinking agent, multiple perception agents, and an evaluation agent; the constructed search tree includes all possible reasoning paths explored and the evaluation value of each node in each possible reasoning path.

[0035] The framework proposed in this invention consists of a global agent, an evaluation agent, and multiple perceptual agents. The global agent selects perceptual agents based on Monte Carlo tree search, and the evaluation agents dynamically update the reward signal, achieving progressive visual reasoning based on image facts. This invention equips the agents with specialized, task-specific visual tools, such as Open-Vocabulary Object Detection (OVOD), image cropping, depth estimation, image segmentation, image similarity detection, and visual relationship detection. These tools are composed of pre-trained visual expert models, and empirical evidence suggests they can perform pixel-level image operations, providing the perceptual agents with context closer to the image itself, thereby reducing the reasoning illusion of VLM (Visual Reasoning Model). For example, Figure 2 This is a schematic diagram of the multi-agent collaborative framework provided by the present invention, as shown below. Figure 2 As shown, the multi-agent cooperative framework includes: Part 1: Agent Design Figure 2 The top left section showcases the designs for three types of intelligent agents: This includes stepwise task generation, agent selection (MCTS), node confidence assessment, and global answer generation.

[0036] Among them, the global thinking agent is responsible for evaluating the task execution chain and generating progressive rewards.

[0037] Evaluator Agent: Independent of other agents, it is dedicated to evaluating the quality of reasoning steps and generating reward signals.

[0038] Perceptual agents: These consist of multiple specialized sub-agents, such as the focusing agent, extended agent, environment-enhancing agent, visual cueing agent, relationship analysis agent, and multi-graph understanding agent listed in the diagram. Each perceptual agent is equipped with dedicated visual tools to perform specific, fine-grained perceptual tasks.

[0039] Figure 2 At the bottom is a list of dedicated vision tools available for use by the perceptual agent, i.e., a tool pool, including: cropping, depth estimation, sharpening, and open-vocabulary object detection (OVOD). These tools consist of pre-trained vision expert models capable of performing pixel-level image manipulations, providing fine-grained evidence based on image facts for the reasoning process, and effectively suppressing visual illusions in large models.

[0040] Part Two: VART Workflow Figure 2 The tree structure illustrates the four core stages of multi-agent collaborative reasoning based on Monte Carlo tree search, with each node labeled with its corresponding action value Q or reward value R. 1. Selection Phase Starting from the root node, search downwards based on the historical evaluation value of the child nodes; The example node Q values ​​are 34.6, 66.3, and 58.2, with R=78; Choose the existing reasoning path that has the most potential.

[0041] 2. Expansion Phase At the selected node, the perceptual agent is invoked to execute a subtask, generating a new child node; The Q values ​​of the example new nodes are 74.2, 56.6, 68.8, and 72.1, respectively, and R=96.

[0042] 3. Simulation Phase Starting from the newly expanded node, reason forward rapidly until the terminal node is reached, forming a complete reasoning trajectory; The perceptual agent is invoked to execute subtasks and generate new child nodes. During the simulation, the Q value of each node gradually increases. The example values ​​are: 83.3 (R=87), 91.5 (R=90), and 94.4 (R=99).

[0043] 4. Backpropagation phase The evaluation agent scores the nodes in the simulated trajectory and transmits the reward signal back along the path; Update the Q value and access count of each node, with the example values ​​further improved to 94.4, 99, 100, and 101.

[0044] Through multiple iterations of the above four steps, the system gradually builds and optimizes the search tree, ultimately obtaining high-quality reasoning results (answers).

[0045] Part Three: Case Analysis Figure 2 The right side shows a specific application example in an autonomous driving scenario: Input images: Multiple consecutive frames of driving scene images (four images are shown in the figure, including vehicles, intersections, traffic lights, etc. from different perspectives); Initial query: "Based on the above image sequence, does the current environment support the vehicle turning left at the next intersection?" The reasoning process (in order of agent invocation) is shown in Table 1 below: Table 1: Final answer: Based on the evidence accumulated step by step above, the system can infer that the current environment supports the vehicle turning left at the next intersection and provide corresponding lane change suggestions.

[0046] Specifically, the design of each intelligent agent is as follows: (1) Global Thinking Agent Design For a set of initial images in a visual reasoning task Given an initial text query q, the global thinking agent generates phased step-by-step tasks. The system utilizes Monte Carlo tree search to select the most suitable perceptual agent for the current stage of task execution. Upon completion of the global task, it summarizes the evidence links of the perceptual agent and derives the final answer. The design of the global thinking agent follows the Reasoning-Acting (ReAct) framework. Its core idea is to allow the large language model to generate reasoning trajectories while interacting with the external environment or tools, alternately executing "thinking" and "acting" steps to solve complex tasks. The exploration process of the global thinking agent follows the ReAct paradigm. A Markov decision process is performed through its exploration process, its state space, state updates, and the final goal.

[0047] State space S: The state space of a globally thinking intelligent agent It is defined as its exploration trajectory, as shown in formula (1), where To explore the number of steps.

[0048] This represents the state of the global thinking agent at the t-th exploration step. This represents the exploration process in the first exploration step. This indicates the exploration process in the second exploration step. Indicates the first The exploration process of each step, the initial state is as shown in formula (2), the initial state of the global thinking agent is determined by the initial picture group. Initial query composition.

[0049] State Update: After each exploration process, the global thinking agent will transition to a new state. The new state is derived from the previous state. With the exploration process It is assembled as shown in formula (3).

[0050] Action Space A: As shown in Equation (4), the action space is defined as the pool of available perceptual agents A, where each... This corresponds to a perceptual intelligent agent.

[0051] Exploration process The exploration process of a globally thinking intelligent agent Adopting the ReAct paradigm, This represents the exploration process at the t-th step. This represents the state-based global thinking agent at the t-th exploration step. Think about it and select several images. Generate query Constitute subtasks, This indicates that the global thinking agent selects the perceptual agent to execute the sub-task, and assigns the sub-task to... As input to the perceptual intelligent agent This indicates that the selected perceptual agent performs a subtask and outputs the execution result. .

[0052] Generate via loop subtasks (Representing the subtask given by the thinking agent), agent selection (Indicates the selection of the specific perceptual agent to execute), agent execution (Representing the execution result of the perceptual agent) Three steps to extract information from images and text that helps solve the problem.

[0053] Final goal: When the global thinking agent determines that it is no longer necessary to call upon other agents to mine image and text information, or when the number of exploration steps reaches the maximum value L, it enters the final state. , where T is the number of exploration steps for the final state.

[0054] (2) Evaluate agent design For the complete reasoning chain from the initial state to the final state, the evaluating agent acts as an independent judge, scoring each exploration process and generating a reward signal for each state (which represents the set of all exploration processes) to improve the global thinking agent's exploration strategy. in, For the first A reward signal for each step of exploration. This is the discount rate, used to balance current gains with future returns. To evaluate the computational process of the intelligent agent, For a globally thinking intelligent agent The state of the step. The process of exploring this state. The prompt words to be input to the evaluation agent. To evaluate the model weights of the agent, For the first A reward signal for each step of exploration.

[0055] (3) Design of perceptual intelligent agents Perceptual Agent: The perceptual agent focuses on different visual perception angles from two dimensions: image enhancement and description enhancement. It generates detailed intermediate reasoning processes by calling tools, providing progressively refined evidence support for the step-by-step reasoning of the global thinking agent.

[0056] Monte Carlo Tree Search (MCTS) is a heuristic search algorithm for decision optimization, initially widely applied in game theory fields such as Go. Its core idea is to evaluate possible action sequences through random simulation and dynamically construct a search tree to find the optimal decision path under limited computational resources. MCTS typically involves four iterative steps: Selection: Starting from the root node, recursively select child nodes based on existing statistical information (such as node value, number of visits) until a leaf node is reached.

[0057] Expansion: Create one or more new child nodes on this leaf node based on the available actions.

[0058] Simulation: Starting from a new node, the simulation is rapidly extended to the terminal state according to a predetermined strategy, resulting in a complete simulation trajectory and its results.

[0059] Backpropagation: The simulation results (reward value) are traced back along the path to update the statistics of all nodes on the path (such as the number of visits and average value).

[0060] By repeating the above process, MCTS can balance utilization (selecting the current optimal branch) and exploration (trying branches that have not been fully explored) to gradually approach the global optimal solution.

[0061] The constructed search tree is a dynamically generated tree data structure generated through iterative execution of the selection, expansion, simulation, and backpropagation steps of MCTS. The root node represents the initial state, internal nodes represent intermediate inference states, leaf nodes represent the terminating nodes, and edges represent the process of calling a perceptual agent to execute a subtask.

[0062] A possible reasoning path is a complete process that starts from the initial state (initial image + initial query), goes through a series of loops of "thinking → selecting an agent → executing → obtaining a result", and ends at the terminal state. Each trajectory (possible reasoning path) corresponds to a path from the root node to a leaf node in the search tree.

[0063] This invention migrates MCTS from the traditional game theory decision-making domain to visual reasoning tasks and makes targeted modifications for fine-grained visual reasoning in multi-agent collaboration: 1. Problem Formalization: Transforming visual reasoning into an MCTS search problem State space: defined as the inference trajectory that accumulates step by step from the initial image and query. Each state St contains historical thoughts, invoked agents, and their execution results.

[0064] Action space: Defined as a pool of available perceptual agents (such as focusing agents, relational understanding agents, etc.). Each action involves selecting and invoking a perceptual agent.

[0065] Reward signals are generated by an independent evaluating agent (rather than the environment) based on logical consistency of reasoning steps, degree of evidence support, etc.

[0066] 2. Visual Reasoning Mapping in Four Steps of MCTS As shown in Table 2 below: Table 2:

[0067] 3. Multiple iterations and final answer generation The system repeats the above four steps N rounds (e.g., N=10 or 20) to gradually build and optimize a search tree, which records a variety of possible perceptual agent call sequences and their evaluation results.

[0068] As a high-level decision-making framework, MCTS decomposes complex visual reasoning tasks into a search problem of multi-step perceptual agents calling sequences. Through an iterative process of selection, expansion, simulation, and backpropagation, it dynamically explores and evaluates multiple reasoning paths. Finally, a global thinking agent integrates multiple high-quality trajectories to generate an accurate and interpretable answer.

[0069] Step 104: Stop the search when the number of iterations reaches a preset threshold. Utilize the global thinking agent in the multi-agent collaborative framework to generate the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0070] Specifically, the Global Thinking Agent is the core decision-maker in the multi-agent collaborative framework, acting as the policy model for Monte Carlo tree search. It is responsible for generating subtasks, selecting which perceptual agent to invoke, and ultimately synthesizing multiple reasoning trajectories to generate the answer. Its design follows the ReAct paradigm, possessing the ability to coordinate "reasoning and action." The terminating node state refers to the state corresponding to the complete path from the root node to a leaf node in the search tree. The terminating node is reached when the Global Thinking Agent deems it unnecessary to invoke a new agent, or when the number of exploration steps reaches a preset maximum value L. Each terminating node state encapsulates a complete reasoning trajectory, including historical thinking, the sequence of invoked agents, the execution result of each step, and the final conclusion.

[0071] In traditional Chain of Reasoning (CoT) methods, the reasoning process is linear and singular. Once an error occurs at any step (for example, the model hallucinates and misidentifies an object), the error propagates cascading along the chain, leading to an irreversibly incorrect final answer. This "one wrong step, a series of wrong steps" problem is the core flaw of CoT.

[0072] This invention employs the Monte Carlo Tree Search (MCTS) framework, exploring multiple different reasoning paths through multiple iterations. These paths may: invoke different sequences of perceptual agents; focus on different local regions in the image; obtain evidence information from different aspects (color, shape, spatial relationships, depth, etc.); and arrive at the same or different intermediate conclusions.

[0073] The core value of weighing multiple trajectories lies in: Error robustness: An error in one path can be "corrected" or "suppressed" by other correct paths.

[0074] Complementarity of evidence: Different paths may reveal information from different dimensions in an image, and combining them can lead to a more complete understanding.

[0075] Uncertainty quantification: By observing the degree of consistency among multiple paths, the confidence level of reasoning can be implicitly assessed.

[0076] In practical applications, after N rounds of Monte Carlo Tree Search (MCTS), a reasoning chain with no more than N final nodes is obtained. .in, This represents the inference links generated in the first round of search, where each inference link consists of multiple exploration steps e connected in series. This represents the inference link generated in the second round of search. Indicates the first The inference chain generated during round search, This indicates the preset maximum value.

[0077] Input the n final nodes into the global thinking agent, weigh the n reasoning links, and thus generate the final answer.

[0078] In practical applications, after the search is completed, the global thinking agent synthesizes the states of multiple terminal nodes in the search tree (i.e., multiple complete reasoning trajectories), weighs the evidence information of each trajectory, and derives the final answer. The global thinking agent compares, evaluates, and merges the multiple reasoning trajectories corresponding to multiple terminal node states. Instead of simply using majority voting or a single optimal path, it comprehensively considers the reliability of the evidence, logical consistency, and coverage of the initial query in each trajectory to derive the most credible final answer.

[0079] The method provided in this invention first obtains a visual reasoning task, which includes at least one image to be processed and an initial query associated with each image. Then, the visual reasoning task is formalized as a Monte Carlo tree search problem within a Markov decision process framework. Further, using a pre-built multi-agent collaborative framework, with each image to be processed and the initial query as the root node state, the selection, expansion, simulation, and backpropagation steps of the Monte Carlo tree search are iteratively executed to solve the Monte Carlo tree search problem, resulting in a constructed search tree. The constructed search tree includes all possible reasoning paths explored and the evaluation value of each node in each possible reasoning path. Finally, the search stops when the number of iterations reaches a preset threshold, and the final answer to the visual reasoning task is generated using a global thinking agent within the multi-agent collaborative framework, based on the states of multiple terminal nodes in the constructed search tree.

[0080] This invention proposes a novel multi-agent fine-grained visual reasoning framework, decomposing the visual language model (VLM) into step-level task planning, fine-grained visual perception, and reasoning performance evaluation. It designs a global thinking agent, a tool-equipped perception agent, and an evaluation agent as plug-and-play modular components, enhancing the VLM's fine-grained understanding capabilities without additional training. Simultaneously, it proposes a Monte Carlo tree search-driven multi-agent collaborative reasoning method. In exploring and optimizing visual reasoning paths, it integrates the selection, expansion, simulation, and backtracking mechanisms of the Monte Carlo Tree Search (MCTS) with the planning, perception, and evaluation processes of multi-agent visual reasoning. This formalizes multi-agent collaborative interaction as a Markov decision process, not only suppressing cascading errors caused by single chain-like reasoning paths through multi-path exploration but also avoiding the inefficiency of broad search in tree-like reasoning through a reward backpropagation mechanism.

[0081] It should be noted that each implementation method of this application can be freely combined, rearranged, or executed individually, and does not need to rely on or depend on a fixed execution order.

[0082] According to the multi-agent visual reasoning method provided by the present invention, a pre-constructed multi-agent cooperative framework is used, with each image to be processed and the initial query as the root node state, to iteratively execute the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem and obtain the constructed search tree, including: For any round of search, in the selected step, the global thinking agent selects an existing reasoning path from the root node based on the current state and the historical evaluation value of each child node. In the extended step, at least one subtask is generated by the global thinking agent based on the existing reasoning path and the current state, and at least one perceptual agent is selected from multiple perceptual agents to execute the subtask in order to generate new child nodes; In the simulation process, the global thinking agent starts from a new sub-node and reasons based on a preset policy model until it reaches the terminal node, thus obtaining a complete reasoning trajectory. In the backpropagation step, the evaluation agent evaluates the state of each node in the inference trajectory and generates a reward signal. The evaluation value and access count of the nodes on the search tree are updated using the reward signal to obtain the updated search tree corresponding to the round of search. Based on the updated search tree corresponding to each round of search, determine the existing inference path corresponding to the next round of search; Based on the existing reasoning path corresponding to the next round of search, the Monte Carlo tree search steps of selection, expansion, simulation and backpropagation are executed iteratively to obtain the updated search tree corresponding to the last round of search. The updated search tree corresponding to the last round of search is then determined as the constructed search tree.

[0083] Specifically, in some embodiments, we use a global thinking agent as the policy model for MCTS. The policy model takes the root node (initial image and query) as input, determines the availability of each perceptual agent, and outputs a value from 0 to 100, representing the confidence level for selecting each agent. in, Let i represent the confidence level of the perceptual agent output by the policy model, and i be the current exploration round. The calculation process of the strategy model, The prompt words are input to the policy model. These are the weight parameters of the strategy model.

[0084] The purpose of the reward model is to evaluate the quality of exploration behavior and optimize exploration strategies. Since it is impossible to accurately calculate and generate reward signals after each round of searching, this invention uses a designed evaluation agent as the reward model. This is a training-free reward method used to simulate human feedback.

[0085] The specific implementation of step 103 includes the following steps: For any given search round, the steps of selection, expansion, simulation, and backpropagation are performed. Any given search round includes: (1) Selection steps: For any round of search, in the selected step, the global thinking agent selects an existing reasoning path from the root node based on the current state and the historical evaluation value of each child node.

[0086] Here, the current node refers to the current tree node in the selection phase of the Monte Carlo tree search. This node corresponds to a specific state. It contains the complete reasoning history from the initial state to the current step. Child nodes are the direct successors of the current node in the search tree, and each child node corresponds to a new state resulting from a possible exploratory action (i.e., invoking a perceptual agent). Historical evaluation value represents the action value Q accumulated by each child node in previous search rounds, reflecting the historical average reward obtained from reasoning starting from that node. The higher the Q value, the better the reasoning results obtained through that node historically.

[0087] In practical applications, for the i-th round of exploration, we start from the root node and search downwards through the explored child nodes, selecting only one node at each level, until we reach the leaf node (the end point).

[0088] (2) Extended steps: In the extended step, at least one subtask is generated by a global thinking agent based on the existing reasoning path and the current state, and at least one perceptual agent is selected from multiple perceptual agents to execute the subtask in order to generate new child nodes.

[0089] When performing subtasks, the perceptual agent acquires fine-grained information from images by calling external vision tools, which include at least one of image cropping tools, depth estimation tools, image segmentation tools, and visual relationship detection tools.

[0090] (3) Simulation steps: In the simulation process, the global thinking agent starts from a new sub-node and reasons based on a preset policy model until it reaches the terminal node, thus obtaining a complete reasoning trajectory.

[0091] The preset strategy model is the global thinking agent itself. During the simulation phase, the global thinking agent selects the next action based solely on its real-time confidence in each subsequent action, without relying on the historical evaluation value of the nodes, thereby quickly generating a complete reasoning trajectory.

[0092] During the simulation phase, the optimal node needs to be selected from k expanding nodes for exploration until the final node (leaf node) is reached. Unlike the selection phase, the simulation phase uses only the confidence level output by the policy model as the selection criterion to calculate the optimal expansion node: in, Indicates the i-th round of search. The exploration process of each step. express, This indicates the exploration steps, including the thought text of the global thinking agent, the selection of experts for the perceptual agent, and the execution result text of the perceptual agent. This indicates that in the i-th round of the search, immediately following... The inference steps following the inference link, where t-1 indicates that the length of the inference link has reached t-1. This is the t-th reasoning step that is about to occur. This represents the output of the policy model during the i-th round of search. Confidence level (for inference link) Subsequent splicing and reasoning steps (confidence level).

[0093] The strategy model (global thinking agent) then iterates through the process until the final node is reached. Thus, the reasoning trajectory is obtained.

[0094] (4) Backpropagation steps: In the backpropagation step, the agent evaluates the state of each node in the inference trajectory and generates a reward signal. The reward signal is then used to update the evaluation value and access count of the nodes on the search tree, resulting in the updated search tree corresponding to each round of search.

[0095] The evaluation agent, acting as an independent reward model, evaluates each intermediate state in the reasoning trajectory. The reward signal is generated by the evaluation agent based on the logical consistency, evidence support, and conclusion accuracy of the exploration process for each intermediate state. The evaluation process of the evaluation agent for intermediate states is independent of the global thinking agent and the perception agent. It does not participate in path generation but acts as a judge, scoring the quality of the generated reasoning steps and propagating the scoring results back as the reward signal.

[0096] In the backpropagation step, updating the evaluation value of the node specifically involves: based on the score of the node's state in the reward model, accumulating the score into the historical action value of the node through a weighted average.

[0097] In practical applications, upon reaching the final node (leaf node)... Then, the reward model from Bottom-up evaluation of the exploration quality of each state And increase the number of times these states are accessed: in, This represents the state at step t in the i-th exploration round. The corresponding number of visits.

[0098] Finally Integrating into action value middle: in, This represents the state at step t in the i-th exploration round. The corresponding action value, This represents the state at step t in the i-th exploration round. The corresponding number of visits, This represents the state at step t in the i-th exploration round. Corresponding exploration quality .

[0099] Furthermore, the steps of selection, expansion, simulation, and backpropagation are repeated. For example, based on the updated search tree corresponding to the previous search round, the existing inference path corresponding to the next search round is determined; based on the existing inference path corresponding to the next search round, the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search are iteratively executed to obtain the updated search tree corresponding to the last search round, and the updated search tree corresponding to the last search round is determined as the constructed search tree.

[0100] The method provided in this invention deeply integrates the selection, expansion, simulation, and backpropagation mechanisms of Monte Carlo tree search with multi-agent collaborative reasoning. It dynamically explores multiple reasoning paths in an iterative manner and optimizes node statistics by using reward signals for backpropagation. This achieves intelligent focusing on high-potential reasoning paths and adaptive pruning of inefficient branches, thereby achieving a balance between accuracy, robustness, and computational efficiency in complex visual reasoning tasks.

[0101] According to the multi-agent visual reasoning method provided by the present invention, the multiple perceptual agents are functional agents equipped with dedicated visual tools. The multiple perceptual agents include at least one of the following: a focusing agent for focusing on a specific local region from a global image, an expanding agent for expanding the cropping range, an environment-enhancing agent for enhancing the visual quality of the image, a label-understanding agent for parsing explicit visual labels in the image, a relationship-understanding agent for mining spatial and dynamic relationships between target objects, and a multi-image-understanding agent for processing the association analysis of multiple images.

[0102] Specifically, the multiple perceptual agents are functional agents equipped with dedicated visual tools. These agents focus on different visual perception angles, starting from two dimensions: image enhancement and description enhancement. By invoking tools, they generate detailed intermediate reasoning processes, providing progressively refined evidence to support the step-by-step reasoning of the global thinking agent. The multiple perceptual agents include at least one of the following: Focus on Intelligent Agents: Following the current stage-specific task requirements of the intelligent agents, the focus is on focusing from the global image to specific objects, performing local magnification or cropping, and specifically capturing the category or attribute features of the target object such as name, color, shape, texture, and physical size, and outputting dynamically cropped images and fine-grained descriptions.

[0103] Extended agent: By judging whether the cropped image and answer at the current stage encompass the target object and key factual evidence, it expands the cropping with the target object as the center and outputs the local factual image and key description with the maximum noise reduction.

[0104] Environment-enhancing agent: Focuses on enhancing the elements of the image itself, separating objects, background and different regions, and optimizing the background clarity, saturation and contrast of the image, outputting the modified and enhanced image.

[0105] Label-understanding agents focus on decoding explicit visual labels and understanding their semantic intent. By accurately identifying visual annotations in images (such as arrows, boxes, and highlighted areas), they bind visual cues to the specific targets or regions they point to, parse their attributes and pointing relationships, and transform them into structured descriptions.

[0106] Relationship-understanding agent: Mines complex semantic relationships between targets and other objects, combines depth estimation tools to analyze the target's 3D geometric and dynamic attributes, including depth, orientation, and dynamic intent, and outputs layout relationships and potential interaction analysis between the target object and other objects.

[0107] Multi-graph understanding agent: specializes in handling multi-graph input scenarios, realizing cross-image modality recognition and association analysis, extracting compact latent space representations of visual features, quantifying semantic similarity between images or regions, and then completing cross-graph reasoning and fusion decision-making, outputting the reasoning results of multi-graph semantic joint.

[0108] The method provided in this invention provides a variety of plug-and-play functional perceptual agents equipped with dedicated visual tools, enabling the system to flexibly call the most suitable expert model for different visual sub-tasks. This allows for the precise capture of fine-grained visual evidence in a modular manner, significantly improving reasoning accuracy and hallucination suppression capabilities in complex scenarios.

[0109] According to the present invention, a multi-agent visual reasoning method is provided, which involves a global thinking agent selecting an existing reasoning path downwards from the root node based on the current state and the historical evaluation value of each child node, including: The global thinking agent invokes an improved confidence interval upper bound algorithm to select the next child node of the current node. The improved confidence interval upper bound algorithm integrates the real-time confidence of the policy model for the candidate child node, the historical action value of the candidate child node, and the number of visits to the candidate child node. Based on the child nodes of each step and the current node, select an existing reasoning path.

[0110] Specifically, in some embodiments, the global thinking agent selects an existing reasoning path from the root node based on the historical evaluation values ​​of the current node and each child node corresponding to the current state through the following steps: First, the global thinking agent invokes an improved confidence interval upper bound algorithm to select the next child node of the current node.

[0111] This invention is an improvement upon the classic Upper Confidence Bounds Applied to Trees (UCT) algorithm. The original UCT formula is: UCT = Average Reward + w * sqrt(ln(Parent Node Visit Count) / Child Node Visit Count).

[0112] This invention incorporates a real-time confidence term for candidate child nodes from the strategy model into formula (10), forming a ternary fusion scoring function: in, Indicates the i-th round of search. The exploration process of each step. This represents the exploration process in the i-th round of the search. Indication of inference link A set of candidate nodes that can be expanded subsequently. For the set of candidate expansion nodes; , , These are weight parameters used to adjust the selection preferences of the Monte Carlo tree; This represents the confidence level of the policy model for candidate nodes. The action value of the candidate node represents the accumulated historical experience during the exploration and is calculated only during the backpropagation phase. This represents the number of times the node has been accessed.

[0113] This invention adds the judgment of the strategy model to the UCT method. It uses the rational inference and divergent thinking of the visual language model to avoid meaningless repeated reasoning by MCTS due to historical experience, and avoids irrational random selection caused by the number of node visits.

[0114] The improved confidence interval upper bound algorithm integrates the real-time confidence of the policy model for candidate child nodes, the historical action value of the candidate child nodes, and the number of times the candidate child nodes have been visited. Real-time confidence is the degree of preference the policy model has for each candidate action (i.e., each perceptual agent) calculated instantly based on the current state. It reflects the prior knowledge and divergent thinking of the globally thinking agent and does not rely on historical statistical information. The historical action value of a candidate child node is the average reward accumulated by each candidate child node during the backpropagation phase, starting from that node. It represents the utilization dimension: a tendency to choose historically well-performing paths. The number of times a candidate child node has been visited is the total number of times each candidate child node has been visited in all searches. It represents the exploration dimension: a tendency to choose paths with fewer visits and not yet fully explored. This invention combines the rational inference and divergent thinking (π-term) of the visual language model with the experience accumulation and exploration balance (Q-term and exploration term) of traditional MCTS, achieving intelligent decision-making that "neither follows old paths nor blindly tries random approaches."

[0115] Furthermore, based on the child nodes and the current node at each step, an existing inference path is selected. That is, starting from the root node, the improved UCT algorithm is recursively called to select child nodes layer by layer, ultimately forming a complete path from the root node to the current leaf node. This path corresponds to an inference trajectory that has not yet been fully expanded.

[0116] For example, the above technical solution in this embodiment corresponds to the selection phase in Monte Carlo tree search. Its input is the current search tree (which already contains several nodes and statistical information), and its output is a path from the root node to a leaf node (this leaf node has not yet been fully expanded, meaning there are still untried actions or the search depth has been reached). The specific process is as follows: 1. Starting from the root node, set the root node as the current node.

[0117] 2. If the current node is not a leaf node, the improved UCT algorithm is called to select the child node with the highest score among all the child nodes of the current node.

[0118] 3. Move to the child node and set it as the new current node.

[0119] 4. Repeat steps 2-3 until a leaf node is reached (i.e., it has not been expanded or the termination condition has been met).

[0120] 5. Returns the complete path from the root node to the leaf node.

[0121] The method provided in this invention incorporates the real-time confidence level of the policy model (global thinking agent) as an explicit factor into the classic UCT algorithm, forming a ternary fusion scoring function. This allows the search tree to not only rely on historical statistical values ​​(Q and N) when selecting the next node, but also fully utilize the rational inference of the current state by the visual language model, thereby finding the optimal perceptual agent calling sequence more efficiently in complex, fine-grained visual reasoning tasks. This design effectively overcomes the shortcomings of traditional MCTS in the visual domain, such as "slow cold start" and "easy to get trapped in local optima".

[0122] According to the present invention, a multi-agent visual reasoning method generates at least one sub-task by a global thinking agent based on existing reasoning paths and the current state, and selects at least one perceptual agent from multiple perceptual agents to execute the sub-task to generate new sub-nodes, including: The global thinking agent thinks based on the existing reasoning path and the current state to generate the sub-task for the current step; Through a global thinking agent, the agent selects and invokes the agent's dedicated tools to execute the sub-task for the current step from multiple perceptual agents, generating the execution result as new state information. Based on the new state information, new child nodes are generated.

[0123] Specifically, in some embodiments, the extension step is implemented in the following manner, including: The global thinking agent, based on the existing reasoning path and the current state, generates sub-tasks for the current step. Then, it selects and invokes specialized tools from multiple perceptual agents to execute these sub-tasks, generating execution results as new state information. Based on this new state information, new child nodes are then created.

[0124] For example, in the selected node Then, the strategy model performs k repeated explorations based on this node, resulting in k expanded nodes: in, Indication of inference link A set of candidate nodes that can be expanded subsequently. It means yes The expanded form of the set is {}, where {} represents a set (from j=1 to j=k, indicating that the set has k elements). Then, it represents the set of possible steps for the t-th reasoning step in the reasoning chain during the i-th round of search, with k possible elements from j=1 to j=k. This indicates the i-th round of the search process. In each step of exploration, closely following The reasoning steps following the reasoning chain, This indicates the i-th round of the search process. In each exploration step, the policy model outputs the pair Confidence level (for inference link) Subsequent splicing and reasoning steps (confidence level).

[0125] For each newly expanded node, initialize its action value Q to 50.

[0126] The method provided in this embodiment of the invention dynamically decomposes complex visual reasoning tasks into executable subtasks, and obtains fine-grained evidence based on image facts by calling dedicated perceptual agents on demand, thereby expanding the search space in the form of structured new nodes and constructing a transparent, evidence-driven stepwise reasoning chain.

[0127] According to the present invention, a multi-agent visual reasoning method utilizes a global thinking agent in a multi-agent collaborative framework to generate the final answer to a visual reasoning task based on the states of multiple terminal nodes in a constructed search tree, including: Input the states of multiple termination nodes into the global thinking agent; The global thinking agent weighs multiple complete reasoning trajectories corresponding to multiple termination node states, integrates the evidence information of each reasoning trajectory, and derives the final answer.

[0128] Specifically, in some embodiments, the process of reasoning to obtain the final answer in step 104 includes the following steps: The global thinking agent is itself a large language model (or visual language model) with powerful reasoning capabilities. When it receives multiple terminal node states, it actually receives multiple complete sets of reasoning histories, each set containing: Each step of the thinking process provides a sub-task ( ); The perceptual agent invoked at each step ( ); The execution result obtained at each step ( ).

[0129] Global Thinking Intelligence Experience: (1) Compare the logical consistency of each trajectory: Identify which evidence is repeatedly verified in different trajectories. (2) Assess the credibility of the evidence: Prioritize the acceptance of evidence from high-precision tools (such as local fine-grained descriptions after focused clipping). (3) Identify and eliminate illusions: If an assertion in a trajectory clearly contradicts the evidence from multiple other trajectories, the assertion may be marked as suspicious. (4) Synthesize and deduce the answer: Based on the integrated set of evidence, generate a final answer that best explains all (or most) of the reliable evidence.

[0130] The following is a specific example: Suppose we input a high-resolution street view image and an initial query into the visual reasoning system proposed in this invention: Image I: A complex street view containing small traffic signs, pedestrians, vehicles, and buildings in the distance. The traffic sign has a blurred speed limit number "30" printed on it.

[0131] Question: "What is the speed limit shown on the traffic sign in the distance of the image?" System execution process (MCTS multi-round search) After system initialization, three rounds of MCTS search were performed iteratively (for simplicity, it is assumed that N=3 here), constructing a search tree containing three complete inference trajectories.

[0132] Trajectory 1 (Path A) is shown in Table 3 below: Table 3:

[0133] The termination node status of trajectory 1 includes: cropping image, marker recognition result, and uncertainty judgment.

[0134] Trajectory 2 (Path B) is shown in Table 4 below: Table 4:

[0135] The termination node status of trajectory 2 includes: an enhanced local image and a clear digit recognition result "30".

[0136] Trajectory 3 (path C) is shown in Table 5 below: Table 5:

[0137] The termination node status of trajectory 3 includes: relationship detection results and scenario-based inference logic.

[0138] Final answer generation (executing claim 6) At this point, the system has obtained three complete inference trajectories and the status of their termination nodes.

[0139] Step 1: Input the states of multiple termination nodes into the global thinking agent. The system inputs the complete history of trajectory 1, trajectory 2, and trajectory 3 (including the thinking at each step, the called agents, and the returned evidence) into the global thinking agent in the form of structured prompts.

[0140] Step 2: The agent makes trade-offs based on global thinking. The global thinking agent (a powerful visual language model) analyzes these three trajectories: Trajectory 1 indicates "possibly 30", but there is uncertainty.

[0141] Trajectory 2 obtained a clear "30" recognition result through image enhancement, and the evidence is direct and highly credible.

[0142] Trajectory 3, inferred from the context, also points to "30", serving as supporting evidence.

[0143] The internal reasoning process of a globally thinking intelligent agent (illustrated): "Trajectory 1 is ambiguous and has low confidence. Trajectory 2, through augmentation tools, provides direct visual evidence and is the most reliable. Trajectory 3's inference is consistent with Trajectory 2, providing cross-validation. Two of the three trajectories clearly point to '30,' and one leans towards '30.' No trajectory points to any other number. Therefore, the answer can be confirmed as '30.'" Step 3: Derive the final answer The global thinking agent outputs the final answer: Final answer: The speed limit shown on the traffic sign in the distance in the image is 30 km / h. Additionally, the system can add a confidence level statement: "Dual verification based on direct recognition after image enhancement (trajectory 2) and contextual inference (trajectory 3)." This invention does not simply select an "optimal" path or conduct a majority vote. Instead, it leverages the powerful semantic understanding and comprehensive reasoning capabilities of the global thinking agent to perform a deep, evidence-based weighing of multiple heterogeneous reasoning trajectories explored within the MCTS framework, thereby generating a more accurate and robust final answer. This design fully embodies the technical concept described in relevant technical materials of "inputting n final nodes into the global thinking agent and weighing n reasoning links," serving as the core bridge transforming the "exploration diversity" of Monte Carlo Tree Search into "answer reliability."

[0144] Figure 3 This is a comparative diagram of several visual reasoning methods, such as... Figure 3 As shown, it includes: I. Left side: End to end This column demonstrates a direct reasoning paradigm called "End to end".

[0145] Input: The method accepts an image and a question as input.

[0146] Processing flow: The input is first processed by "Tools", such as "Image Cropping". Then, the processed information is handed over to multiple agents (labeled as Agent A, B, C, D in the diagram) or "VLM" (Visual Language Model) for collaborative processing, such as perceptual enhancement and image-text reasoning.

[0147] Core and Output: The core of the entire process is "Aware Enhancement," which aims to improve the model's ability to understand key information in images. Finally, the final answer is generated directly through "Image-Text Reasoning."

[0148] II. Intermediate Stage: Step-by-Step Reasoning (Solving Step by Step) This column demonstrates a decompositional approach called "Step by step," which is further subdivided into two strategies.

[0149] Input: The initial input is also an image and a question.

[0150] Branch 1: Based on COT, this path represents the "thought chain" reasoning mode (perceptual enhancement). After the input is "traversed", it may trigger a series of specific perceptual tasks such as Task 1 object detection, Task 2 vehicle description, and Task N occlusion prediction. Each perceptual task is executed by different agents (labeled as Agent A, B, N in the diagram), and the answer is approximated by decomposing and concatenating sub-tasks.

[0151] Branch 2: Based on TOT, this path represents the "Task Tree" inference mode. The diagram shows that it handles complex tasks such as "Occlusion Prediction" and also involves diverse perceptual enhancement information such as "Lens" and "Depth".

[0152] Common process and output: Both branches integrate processing from multiple agents and "Aware Enhancement" during the reasoning process, ultimately leading to the answer.

[0153] III. Right side: Monte Carlo tree search method This column showcases the Monte Carlo tree search-based approach labeled "Ours Visual Analytics for Representative Time Series Selection and Visualization (VART)".

[0154] Input: The method begins with an image and a question.

[0155] Processing flow: The core is a cyclical "Searching" and "Simulating" process. The diagram uses nodes S1 to S8 to represent different states in the search tree. During the simulation, the system evaluates based on the "Reward" signal and applies strategies such as "Overshoot" or "Shrink" to dynamically adjust the search direction and scope to explore better reasoning paths.

[0156] Output: After multiple rounds of search, simulation, and strategy adjustment, the method outputs the final answer.

[0157] In summary, the diagram clearly uses flowcharts, connecting arrows, and a primary color scheme of blue, green, and gray to visually compare the architectural differences of the three methods. (1) End-to-end approach: The process is relatively straightforward, emphasizing the generation of answers in one step after enhancing perception.

[0158] (2) Step-by-step solution: Explicitly break down complex problems into multiple sub-tasks or tree structures and solve them step by step.

[0159] (3) MCTS method: This method introduces search, simulation, and policy evaluation mechanisms to guide reasoning through trial and error and optimization. All the components mentioned in the figure, such as various tools, agents, specific tasks, and search operations, constitute an indispensable part of their respective methods.

[0160] The multi-agent visual reasoning device provided by the present invention is described below. The multi-agent visual reasoning device described below can be referred to in correspondence with the multi-agent visual reasoning method described above.

[0161] Figure 4 This is a schematic diagram of the structure of the multi-agent visual reasoning device provided by the present invention, as shown below. Figure 4As shown, the multi-agent visual reasoning device 400 includes the following modules: The acquisition module 410 is used to acquire a visual reasoning task, wherein the visual reasoning task includes at least one image to be processed and an initial query related to each of the images to be processed; Transformation module 420 is used to formalize the visual reasoning task into a Monte Carlo tree search problem within the Markov decision process framework; The visual reasoning module 430 is used to utilize a pre-built multi-agent collaborative framework, with each of the images to be processed and the initial query as the root node state, to iteratively execute the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem and obtain a constructed search tree. The constructed search tree includes all possible reasoning paths explored and the evaluation value of each node in each possible reasoning path. When the number of iterations reaches a preset threshold, the search stops, and the global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0162] This invention proposes a novel multi-agent fine-grained visual reasoning framework, decomposing the visual language model (VLM) into step-level task planning, fine-grained visual perception, and reasoning performance evaluation. It designs a global thinking agent, a tool-equipped perception agent, and an evaluation agent as plug-and-play modular components, enhancing the VLM's fine-grained understanding capabilities without additional training. Simultaneously, it proposes a Monte Carlo tree search-driven multi-agent collaborative reasoning method. In exploring and optimizing visual reasoning paths, it integrates the selection, expansion, simulation, and backtracking mechanisms of the Monte Carlo Tree Search (MCTS) with the planning, perception, and evaluation processes of multi-agent visual reasoning. This formalizes multi-agent collaborative interaction as a Markov decision process, not only suppressing cascading errors caused by single chain-like reasoning paths through multi-path exploration but also avoiding the inefficiency of broad search in tree-like reasoning through a reward backpropagation mechanism.

[0163] According to the present invention, a multi-agent visual reasoning device 400 is provided, wherein the multi-agent collaborative framework includes a global thinking agent, multiple perceptual agents, and an evaluation agent; the visual reasoning module 430 is specifically used for: For any round of search, in the selection step, the global thinking agent selects an existing reasoning path from the root node downwards based on the current state and the historical evaluation value of each child node. In the extended step, the global thinking agent generates at least one subtask based on the existing reasoning path and the current state, and selects at least one perceptual agent from the plurality of perceptual agents to execute the subtask in order to generate a new child node; In the simulation steps, the global thinking agent starts from the new child node and reasons based on the preset policy model until it reaches the terminal node, thus obtaining a complete reasoning trajectory. In the backpropagation step, the evaluation agent evaluates the state of each node in the inference trajectory and generates a reward signal. The reward signal is then used to update the evaluation value and access count of the nodes on the search tree to obtain the updated search tree corresponding to the round of search. Based on the updated search tree corresponding to the previous search round, determine the existing reasoning path corresponding to the next search round; Based on the existing reasoning path corresponding to the next round of search, the selection, expansion, simulation and backpropagation steps of Monte Carlo tree search are executed iteratively to obtain the updated search tree corresponding to the last round of search, and the updated search tree corresponding to the last round of search is determined as the constructed search tree.

[0164] According to the present invention, a multi-agent visual reasoning device 400 is provided, wherein the plurality of perceptual agents are functional agents equipped with dedicated visual tools, and the plurality of perceptual agents include at least one of the following: a focusing agent for focusing on a specific local region from a global image, an expanding agent for expanding the cropping range, an environment enhancement agent for enhancing the visual quality of an image, a label understanding agent for parsing explicit visual labels in an image, a relationship understanding agent for mining spatial and dynamic relationships between target objects, and a multi-image understanding agent for processing association analysis of multiple images.

[0165] According to the multi-agent visual reasoning device 400 provided by the present invention, the visual reasoning module 430 is further used for: The global thinking agent invokes an improved confidence interval upper bound algorithm to select the next child node of the current node. The improved confidence interval upper bound algorithm integrates the real-time confidence of the policy model for the candidate child node, the historical action value of the candidate child node, and the number of visits to the candidate child node. Based on the child nodes and the current node at each step, the existing reasoning path is selected.

[0166] According to the multi-agent visual reasoning device 400 provided by the present invention, the visual reasoning module 430 is further used for: The global thinking agent thinks based on the existing reasoning path and the current state to generate the sub-task for the current step. The global thinking agent selects and invokes the dedicated tools of the perceptual agents to execute the sub-task of the current step from the multiple perceptual agents, and generates the execution result as new state information. Based on the new state information, the new child nodes are extended.

[0167] According to the multi-agent visual reasoning device 400 provided by the present invention, the visual reasoning module 430 is further used for: The states of the multiple termination nodes are input to the global thinking agent; The global thinking agent weighs the multiple complete reasoning trajectories corresponding to the multiple termination node states, integrates the evidence information of each reasoning trajectory, and derives the final answer.

[0168] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a multi-agent visual reasoning method, which includes: A visual reasoning task is obtained, the visual reasoning task including at least one image to be processed and an initial query associated with each of the images to be processed; The visual reasoning task is formalized as a Monte Carlo tree search problem within the Markov decision process framework. Using a pre-built multi-agent cooperative framework, with each of the images to be processed and the initial query as the root node state, the Monte Carlo tree search steps of selection, expansion, simulation and backpropagation are iteratively executed to solve the Monte Carlo tree search problem and obtain the constructed search tree; The search stops when the number of iterations reaches a preset threshold. The global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0169] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0170] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the multi-agent visual reasoning method provided by the above methods, the method comprising: A visual reasoning task is obtained, the visual reasoning task including at least one image to be processed and an initial query associated with each of the images to be processed; The visual reasoning task is formalized as a Monte Carlo tree search problem within the Markov decision process framework. Using a pre-built multi-agent cooperative framework, with each of the images to be processed and the initial query as the root node state, the Monte Carlo tree search steps of selection, expansion, simulation and backpropagation are iteratively executed to solve the Monte Carlo tree search problem and obtain the constructed search tree; The search stops when the number of iterations reaches a preset threshold. The global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0171] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the multi-agent visual reasoning method provided by the methods described above, the method comprising: A visual reasoning task is obtained, the visual reasoning task including at least one image to be processed and an initial query associated with each of the images to be processed; The visual reasoning task is formalized as a Monte Carlo tree search problem within the Markov decision process framework. Using a pre-built multi-agent cooperative framework, with each of the images to be processed and the initial query as the root node state, the Monte Carlo tree search steps of selection, expansion, simulation and backpropagation are iteratively executed to solve the Monte Carlo tree search problem and obtain the constructed search tree; The search stops when the number of iterations reaches a preset threshold. The global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

[0172] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0173] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent visual reasoning method, characterized in that, include: A visual reasoning task is obtained, the visual reasoning task including at least one image to be processed and an initial query associated with each of the images to be processed; The visual reasoning task is formalized as a Monte Carlo tree search problem within the Markov decision process framework. Using a pre-built multi-agent cooperative framework, with each of the images to be processed and the initial query as the root node state, the Monte Carlo tree search steps of selection, expansion, simulation and backpropagation are iteratively executed to solve the Monte Carlo tree search problem and obtain the constructed search tree; the constructed search tree includes all possible inference paths explored and the evaluation value of each node in each possible inference path; The search stops when the number of iterations reaches a preset threshold. The global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

2. The multi-agent visual reasoning method according to claim 1, characterized in that, The multi-agent collaborative framework includes a global thinking agent, multiple perception agents, and an evaluation agent. Utilizing the pre-constructed multi-agent collaborative framework, with each image to be processed and the initial query as the root node state, iteratively executes the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem, obtaining a constructed search tree, including: For any round of search, in the selection step, the global thinking agent selects an existing reasoning path from the root node downwards based on the current state and the historical evaluation value of each child node. In the extended step, the global thinking agent generates at least one subtask based on the existing reasoning path and the current state, and selects at least one perceptual agent from the plurality of perceptual agents to execute the subtask in order to generate a new child node; In the simulation steps, the global thinking agent starts from the new child node and reasons based on the preset policy model until it reaches the terminal node, thus obtaining a complete reasoning trajectory. In the backpropagation step, the evaluation agent evaluates the state of each node in the inference trajectory and generates a reward signal. The reward signal is then used to update the evaluation value and access count of the nodes on the search tree to obtain the updated search tree corresponding to the round of search. Based on the updated search tree corresponding to the previous search round, determine the existing reasoning path corresponding to the next search round; Based on the existing reasoning path corresponding to the next round of search, the selection, expansion, simulation and backpropagation steps of Monte Carlo tree search are executed iteratively to obtain the updated search tree corresponding to the last round of search, and the updated search tree corresponding to the last round of search is determined as the constructed search tree.

3. The multi-agent visual reasoning method according to claim 2, characterized in that, The plurality of perceptual agents are functional agents equipped with dedicated visual tools. The plurality of perceptual agents include at least one of the following: a focusing agent for focusing on a specific local region from a global image; an expanding agent for expanding the cropping range; an environment-enhancing agent for enhancing the visual quality of an image; a label-understanding agent for parsing explicit visual labels in an image; a relationship-understanding agent for mining spatial and dynamic relationships between target objects; and a multi-image understanding agent for processing association analysis of multiple images.

4. The multi-agent visual reasoning method according to claim 2, characterized in that, The process of the global thinking agent selecting an existing reasoning path downwards from the root node based on the current state and the historical evaluation value of each child node includes: The global thinking agent invokes an improved confidence interval upper bound algorithm to select the next child node of the current node. The improved confidence interval upper bound algorithm integrates the real-time confidence of the policy model for the candidate child node, the historical action value of the candidate child node, and the number of visits to the candidate child node. Based on the child nodes and the current node at each step, the existing reasoning path is selected.

5. The multi-agent visual reasoning method according to claim 2, characterized in that, The process of generating at least one sub-task based on the existing reasoning path and the current state through the global thinking agent, and selecting at least one perceptual agent from the plurality of perceptual agents to execute the sub-task to generate new child nodes, includes: The global thinking agent thinks based on the existing reasoning path and the current state to generate the sub-task for the current step. The global thinking agent selects and invokes the dedicated tools of the perceptual agents to execute the sub-task of the current step from the multiple perceptual agents, and generates the execution result as new state information. Based on the new state information, the new child nodes are extended.

6. The multi-agent visual reasoning method according to any one of claims 1-5, characterized in that, The step of utilizing the global thinking agent in the multi-agent collaborative framework to generate the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree includes: The states of the multiple termination nodes are input to the global thinking agent; The global thinking agent weighs the multiple complete reasoning trajectories corresponding to the multiple termination node states, integrates the evidence information of each reasoning trajectory, and derives the final answer.

7. A multi-agent visual reasoning device, characterized in that, include: An acquisition module is used to acquire a visual reasoning task, wherein the visual reasoning task includes at least one image to be processed and an initial query related to each of the images to be processed; The transformation module is used to formalize the visual reasoning task into a Monte Carlo tree search problem within the Markov decision process framework. The visual reasoning module utilizes a pre-built multi-agent collaborative framework, using each image to be processed and the initial query as the root node state, to iteratively execute the selection, expansion, simulation, and backpropagation steps of Monte Carlo tree search to solve the Monte Carlo tree search problem and obtain a constructed search tree. The constructed search tree includes all possible reasoning paths explored and the evaluation value of each node in each possible reasoning path. The search stops when the number of iterations reaches a preset threshold, and the global thinking agent in the multi-agent collaborative framework generates the final answer to the visual reasoning task based on the states of multiple terminal nodes in the constructed search tree.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent visual reasoning method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-agent visual reasoning method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-agent visual reasoning method as described in any one of claims 1 to 6.

Images