A mixed reality-based multi-agent critical thinking training system

Abstract

The application relates to the technical field of a mixed-reality-based multi-agent critical thinking training system, and particularly discloses a mixed-reality-based multi-agent critical thinking training system, which aims to solve technical problems such as algorithm dependence of learners and invisible reasoning process. The system comprises a space environment perception and virtual-real mapping module, a multi-agent logical dialectic engine, a logical semantic analysis and reasoning chain extraction module, a 3D logical structure mapping module, a cognitive load monitoring and adaptive support module and a critical thinking evolution evaluation module. By converting abstract logic into intuitive 3D space topology, the application reduces cognitive burden, strengthens the independent thinking ability of learners by using a multi-agent confrontation mechanism, realizes adaptive adjustment of training intensity and quantitative evaluation of ability through physiological feedback, and constructs a closed-loop and iterative cognitive quality training system.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided instruction and mixed reality technology, specifically relating to a multi-agent critical thinking training system based on mixed reality. Background Technology

[0002] The rapid development of next-generation information technologies such as artificial intelligence and extended reality is profoundly changing the paradigm of education and cognitive training. In modern education systems, intelligent teaching systems, through deep learning and interactive media technologies, not only achieve efficient knowledge transfer but also become important vehicles for cultivating learners' higher-order thinking abilities. Among these, mixed reality technology, as a link between the physical world and digital space, can provide immersive learning experiences through spatially perceptive multimodal interaction, while multi-agent systems, by simulating complex social collaboration and adversarial logic, demonstrate significant application value in logical reasoning, decision optimization, and collaborative learning.

[0003] Critical thinking training, as a core direction in education, aims to enhance individuals' rational thinking and problem analysis abilities through multi-dimensional logical deconstruction, hypothesis testing, and refutation. Multi-agent systems based on mixed reality, by constructing virtual characters with specific logical preferences and professional backgrounds, can reproduce complex debate, decision-making, and reasoning scenarios in virtual space. This guides users to identify logical fallacies, examine potential preferences, and improve their cognitive structures through dynamic interaction, thereby building a systematic logical verification mechanism.

[0004] Current technologies still have significant shortcomings in addressing the cognitive challenges brought about by the widespread adoption of generative artificial intelligence. Traditional text-based interaction modes tend to directly output efficient responses rather than provide heuristic guidance, leading to learners' heavy reliance on algorithms and weakening their understanding of the reasoning process itself. Furthermore, the system's dialogue often conforms to the user's preconceived notions, further amplifying cognitive biases and hindering the formation of an effective learning loop. In addition, existing educational training systems mostly employ linear or two-dimensional symbolic presentation methods, making it impossible to intuitively visualize complex reasoning structures. This not only increases the learner's working memory burden but also makes it difficult for them to accurately locate key assumptions and logical flaws. Simultaneously, current intelligent agent simulation applications lack deep engagement; their role settings and interaction paradigms are relatively simplistic, failing to provide sufficient support for adolescents and beginner learners in deep logical construction, metacognitive development, and proactive verification awareness. This results in a significant reduction in the effectiveness of critical thinking training in complex and ever-changing real-world application scenarios. Therefore, there is an urgent need for an optimized multi-agent critical thinking training system based on mixed reality. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-agent critical thinking training system based on mixed reality to solve the technical problems in the prior art, such as learners becoming dependent on algorithms due to the direct output of results by generative artificial intelligence, the invisibility of the reasoning process, the amplification of cognitive biases, and the lack of deep engagement due to the single interaction paradigm.

[0006] The technical solution of this invention includes: a spatial environment perception and virtual-real mapping module, used to acquire depth and image information of the physical environment through a mixed reality terminal, construct a virtual interactive space aligned with physical space coordinates, and capture multimodal behavioral data of the user within the virtual interactive space; a multi-agent logical dialectical engine, used to construct multiple virtual agents with different logical attributes and knowledge backgrounds, and generate conflicting logical assertions and debate strategies based on the current training topic; and a logical semantic analysis and reasoning chain extraction module, used to perform semantic parsing on the user's input text or speech, extract the core assumptions, reasoning paths, and conclusions in the user's argumentation process, and... The system identifies logical fallacies; a 3D logical structure mapping module transforms logical elements extracted by the logical semantic analysis and reasoning chain extraction module into a visualized spatial topology, which is then spatially anchored and displayed in the virtual interactive space; a cognitive load monitoring and adaptive scaffolding module assesses the user's cognitive load level in real time based on their physiological feedback and interactive performance during training, and dynamically adjusts the challenge intensity and the level of detail of logical prompts for the virtual agent accordingly; and a critical thinking evolution evaluation module provides a multi-dimensional quantitative assessment of the user's logical rigor, evidentiary validity, preference recognition ability, and metacognitive monitoring ability throughout the training cycle.

[0007] Furthermore, when the spatial environment perception and virtual-real mapping module is working, it performs real-time spatial scanning through the depth sensor integrated in the mixed reality terminal to generate point cloud data of the physical environment; it uses simultaneous localization and mapping algorithms to determine the 6-DOF pose of the terminal in space; it fixes the virtual debate scene, logic nodes, and virtual intelligent agents to specific positions on the physical surface through spatial anchoring technology; at the same time, it uses infrared sensors and visible light cameras to capture the user's gaze focus, gestures, and facial expressions, and converts the above multimodal behavioral data into interactive commands that the system can recognize.

[0008] In one embodiment of the present invention, the multi-agent logical dialectic engine includes a role profile generation unit, a conflict logic modeling unit, and a dialogue strategy scheduling unit. The role profile generation unit assigns a unique cognitive label to each virtual agent, including logical inclination, professional domain background, and debating style; logical inclination includes confirmation bias, skepticism, deductive reasoning, and empiricism; professional domain background covers multiple dimensions such as philosophy, science, ethics, and law. The conflict logic modeling unit, for a preset training topic, retrieves mutually exclusive arguments from a knowledge base and uses a generative language model to construct a structured argument containing preconditions, logical operators, and conclusions. The dialogue strategy scheduling unit, based on the user's argumentation logic, schedules virtual agents with different cognitive labels to participate in the debate in real time; when the user exhibits a clear cognitive loop, a skeptical agent is scheduled to provide counterexamples; when the user's logic jumps, a deductive reasoning agent is scheduled to require the user to fill in the intermediate steps.

[0009] As one embodiment of the present invention, the logical semantic analysis and reasoning chain extraction module performs dependency syntax analysis and semantic role labeling on the user's argument using a preset natural language processing algorithm to identify the premises, evidence, and inference logic in the argument. Furthermore, the module utilizes a logic fallacy classifier based on a converter architecture to match the user's argument with 12 preset common logic fallacy templates, including but not limited to hasty generalization, ad hominem attacks, circular reasoning, and slippery slope fallacy. The extracted reasoning chain is stored in a structured form as a directed graph, where nodes represent declarative propositions and edges represent logical derivation relationships.

[0010] In one embodiment of the present invention, the 3D logical structure mapping module maps a directed graph-like reasoning chain to a virtual interactive space. Propositional nodes are assigned different geometric shapes and light intensities based on their importance in the argument; logical deduction relationships are connected by light strips with a sense of flow; when there is a logical break or insufficient evidence in the reasoning chain, the corresponding connecting edge appears broken or a warning color. Users can use gestures to grab, zoom, or reorganize logical nodes in the space; when a user clicks on a specific node, the system will pop up an interface showing the underlying assumptions upon which the proposition depends and its corresponding evidence tracing.

[0011] Furthermore, the cognitive load monitoring and adaptive scaffolding module calculates a real-time cognitive load index by analyzing the user's gaze duration, pupil diameter changes, and speech rate fluctuations. This cognitive load index is categorized into three levels: low load, suitable load, and overload. When the system determines the user is in a low-load state, it instructs the multi-agent logical dialectical engine to increase the depth of rebuttal, introducing more complex hidden logical fallacies to increase training difficulty. When the system determines the user is in an overload state, it automatically triggers an adaptive scaffolding mechanism, highlighting the current logical bottleneck in the 3D logical structure mapping module and providing guiding questions from a mentor-type agent to reduce the user's search space. The scaffolding mechanism includes three modes: heuristic questioning, logical path simplification, and contradiction focus, with weights allocated based on the user's real-time performance.

[0012] Furthermore, the critical thinking evolution evaluation module constructs an evaluation model comprising four levels. Level 1 is logical completeness, calculating the proportion of valid reasoning paths in the total paths of a user's argument. Level 2 is evidence strength, assessing the strength and reliability of the correlation between the factual evidence cited by the user and their conclusion. Level 3 is perspective shift rate, statistically analyzing the frequency with which users proactively correct their stance or absorb reasonable objections when facing virtual agents with different logical attributes. Level 4 is metacognitive awareness, recording the time efficiency with which users self-monitor and correct their own logical flaws. The system generates a radar chart of the user's critical thinking ability by performing time-series analysis on the data from these four levels and provides targeted follow-up training suggestions.

[0013] As one embodiment of the present invention, this system operates in a layered collaborative computing architecture. The bottom layer is the hardware interface layer, responsible for driving the sensor array of the mixed reality terminal; the middle layer is the logic processing and rendering layer, responsible for executing multi-agent collaborative logic, semantic parsing, and 3D graphics rendering; the top layer is the policy management layer, responsible for controlling the training process and running the evaluation algorithm. Sub-second information exchange is performed between the layers via a high-speed data bus to ensure the real-time nature of virtual-real interaction and the synchronization of logical feedback.

[0014] Furthermore, the system also includes a multi-user collaboration module, supporting multiple users wearing mixed reality terminals to enter the same virtual interactive space. In this mode, the system assigns different observation perspectives and logical tasks to different users; users can collaboratively deconstruct or debate the same logical node, while the virtual intelligent agent acts as a referee, recorder, or logical disruptor, constructing a high-fidelity social cognitive training environment.

[0015] As one embodiment of the present invention, the conflict logic modeling process in the multi-agent logic dialectic engine is as follows: First, the core concept set of the training topic is obtained, and the relationship between concepts is expanded using knowledge graph technology; second, based on predicate logic, positive argumentation paths and negative challenge paths are constructed for the topic; then, according to the preset cognitive challenge level, specific types of logical weaknesses are manually implanted in the argumentation paths; finally, the logic paths are converted into natural language expressions and distributed to virtual agents with corresponding attributes for expression.

[0016] Furthermore, the logical semantic analysis and reasoning chain extraction module employs a multi-feature fusion method to identify logical fallacies. The system not only extracts semantic features at the text level but also incorporates acoustic features such as the user's intonation and stress distribution, as well as the logical node positions indicated by gestures during the argument. By inputting multi-dimensional feature vectors into a pre-trained deep belief network, it achieves accurate identification of logical fallacies, with an accuracy rate of no less than 92%.

[0017] Furthermore, the 3D logical structure mapping module also has a dynamic temporal backtracking function. The system automatically records the logical evolution trajectory during the training process. Users can observe how the logical structure diagram gradually evolves from an initial scattered state into a complex network structure by sliding the virtual timeline, and can reproduce the debate process at any key decision point, thereby achieving in-depth reflection on the thought process.

[0018] Furthermore, the physiological feedback analysis process in the cognitive load monitoring and adaptive support module is as follows: the user's skin conductance response and heart rate variability data are collected by the bioelectric sensor embedded in the mixed reality terminal; the characteristic statistics of the physiological signals, including mean, variance and power spectral density, are extracted using the sliding window algorithm; the physiological characteristics and behavioral interaction characteristics are weighted and fused, and the cognitive load level is output through a preset fuzzy logic inference engine.

[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows: 1. This invention utilizes mixed reality technology to transform abstract, intangible critical thinking logic into an intuitive, interactive 3D spatial topological structure, significantly reducing the working memory burden on learners. Users no longer need to painstakingly construct complex argument models in their minds; instead, they can disassemble and reassemble logical nodes as if manipulating physical entities. This spatialized cognitive mapping helps learners quickly locate logical flaws and deeply understand the internal structure of the reasoning chain, thereby significantly improving the intuitiveness and efficiency of cognitive training.

[0020] 2. This invention constructs a dynamic debate environment composed of multiple intelligent agents with different cognitive biases and professional backgrounds, changing the single, compliant interaction paradigm of traditional intelligent teaching systems. By actively introducing logical conflicts and opposing viewpoints, the system forces learners to break out of their existing mindsets and constantly examine their own positions and logical premises in a process of confrontation and collaboration. This multi-dimensional stress test can effectively suppress algorithm dependence and strengthen learners' independent thinking ability and metacognitive monitoring level in complex information environments.

[0021] 3. This invention achieves real-time and accurate perception of learners' cognitive states through integrated multimodal sensors. The system can dynamically adjust training difficulty and guidance strategies based on users' physiological feedback and interaction performance, enabling personalized customization of critical thinking training. This adaptive scaffolding mechanism ensures that training tasks are sufficiently challenging while avoiding cognitive collapse due to excessive difficulty, thus providing optimal cognitive development ranges for learners at different ability levels and ensuring the transferability and sustainability of training effects in practical application scenarios.

[0022] 4. This invention transforms critical thinking skills, which are inherently difficult to measure, into multi-dimensional objective data indicators through a quantitative evaluation model. By continuously tracking core parameters such as logical completeness and perspective shift rate, the system can provide users with a clear trajectory of skill development and precise directions for improvement. This data-driven feedback mechanism not only enhances learners' sense of accomplishment and engagement but also provides educators with a scientific basis for targeted teaching interventions, constructing a closed-loop, iterative cognitive competence development system. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical architecture of the multi-agent critical thinking training system based on mixed reality proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the multi-agent logic dialectic engine and conflict logic modeling in this invention; Figure 3 This is a logical flow framework diagram illustrating the logical semantic analysis and 3D logical structure mapping in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between cognitive load monitoring and adaptive stent adjustment based on user physiological feedback in this invention; Figure 5 This is a logical flowchart of the critical thinking evolution evaluation model and its multi-dimensional quantitative assessment in this invention. Detailed Implementation

[0024] This embodiment provides a multi-agent critical thinking training system based on mixed reality. Please refer to the appendix. Figure 1 The system consists of a spatial environment perception and virtual-real mapping module, a multi-agent logical dialectical engine, a logical semantic analysis and reasoning chain extraction module, a 3D logical structure mapping module, a cognitive load monitoring and adaptive scaffolding module, and a critical thinking evolution evaluation module. These modules operate in a layered and collaborative computing architecture, achieving sub-second data exchange through a hardware interface layer, a logic processing and rendering layer, and a policy management layer.

[0025] Combined with appendix Figure 1 The spatial environment perception and virtual-real mapping module is responsible for establishing the connection between the physical world and the virtual logical space. This module utilizes the depth sensor integrated into the mixed reality terminal to perform a 3D scan of the user's physical environment at a frequency of 60 times per second. The sensor emits infrared structured light with a wavelength of 850 nanometers, and calculates the depth value of each sampling point in the space by measuring the time of flight of the light, thereby generating physical environment point cloud data containing no less than 1 million vertices. Using simultaneous localization and mapping algorithms, the system calculates the 6-DOF pose of the mixed reality terminal in the physical space in real time, namely the displacement along the X, Y, and Z axes and the rotation angle around each axis, ensuring the spatial stability of virtual objects in vision. After acquiring the physical space topology, this module fixes the virtual debate scene, logical nodes, and virtual intelligent agents to specific positions on the physical surface using spatial anchoring technology. At the same time, it uses infrared sensors and visible light cameras at a sampling rate of 120 Hz to capture the user's gaze focus, gestures, and facial expressions. Eye-tracking data includes pupil position and gaze duration, gesture data includes the spatial coordinates of 21 key skeletal points, and facial expressions are parsed into discrete emotional features based on action units. After normalization, this multimodal behavioral data is transformed into interactive commands that the system can recognize, serving as the raw input for subsequent logical reasoning and cognitive assessment.

[0026] Please refer to the attached document. Figure 2 The multi-agent logical dialectical engine is the core power source of the system, comprising a role profile generation unit, a conflict logic modeling unit, and a dialogue strategy scheduling unit. The role profile generation unit assigns a unique cognitive label to each virtual agent. These labels define the agent's logical tendencies; for example, a confirmation-biased agent tends to seek evidence supporting its own viewpoint, a skeptical agent tends to question the truthfulness of all premises, a deductive reasoning agent strictly follows formal logic rules, and an empiricist agent relies on statistical data and case facts. Professional domain backgrounds presuppose the boundaries of the agent's knowledge base, covering fields such as philosophy, science, ethics, and law.

[0027] The conflict logic modeling unit retrieves mutually exclusive arguments from a knowledge base for a predefined training topic, such as the ethical boundaries of artificial intelligence. This unit utilizes a generative language model based on predicate logic to construct structured arguments. The argumentation process follows this logical flow: First, it acquires the core concept set of the training topic and expands the relationships between concepts using knowledge graph technology. Second, it constructs positive and negative argumentation paths for the topic based on predicate logic. Then, according to a predefined cognitive challenge level, it manually implants specific types of logical weaknesses into the argumentation paths. Finally, it transforms the logical paths into natural language expressions and distributes them to virtual agents with corresponding attributes for expression. The dialogue strategy scheduling unit, based on the user's argumentation logic, schedules virtual agents with different cognitive labels to participate in the debate in real time. When a user exhibits a clear cognitive loop, i.e., repeatedly citing single-dimensional evidence, a skeptical agent is scheduled to provide counterexamples to break the mindset. When a user makes logical leaps, i.e., the conclusion lacks a deductive relationship with the premises, a deductive reasoning agent is scheduled to require the user to fill in the intermediate steps.

[0028] Please refer to the attached document. Figure 3 The logical semantic analysis and reasoning chain extraction module performs deep analysis of user interaction input. This module uses natural language processing algorithms to perform dependency syntax analysis and semantic role labeling on the user's arguments, identifying premises, evidence, and inference logic. To identify logical fallacies, this module employs a multi-feature fusion method. The system not only extracts semantic features at the text level but also combines acoustic features such as the user's intonation and stress distribution, as well as the logical node positions indicated by gestures during the argument. By inputting multi-dimensional feature vectors into a pre-trained deep belief network, the system achieves accurate identification of logical fallacy types. The identification process follows the following logical formula:

[0029] In the above formula, Indicates the presence of given text features Acoustic characteristics and gesture features Under these conditions, the user's argument belongs to the first... The probability of a logical fallacy. It is the feature vector after fusion through deep belief network. and It is the first The system includes a weight matrix and bias terms for the fallacy classifier. It pre-defines 12 common logical fallacies, including hasty generalization, ad hominem attacks, circular reasoning, and slippery slope fallacy. The extracted reasoning chains are stored in a structured directed graph, where nodes represent propositional statements, edges represent logical deductions, and each edge is associated with a weight value indicating the strength of the deduction.

[0030] The 3D logical structure mapping module maps directed graph-like reasoning chains to the virtual interaction space. (Combined with...) Figure 3 Propositional nodes are assigned different geometric shapes based on their importance in the argument; for example, core premises are represented by icosahedrons, and intermediate conclusions by octahedrons. High light intensity of a node indicates strong evidentiary support for the proposition. Logical derivations are connected by flowing light strips, with the speed of the strips representing the logical rigor of the reasoning. When there is a logical break or insufficient evidence in the reasoning chain, the corresponding connecting edge appears broken or displays a warning color, such as red or orange. Users can use gestures to grab, zoom, or reorganize logical nodes in the space. When a user clicks on a specific node, the system will display the underlying assumptions upon which the proposition relies and their corresponding evidence tracing interface. This module also features dynamic temporal backtracking, automatically recording the logical evolution trajectory during training. Users can slide the virtual timeline to observe how the logical structure diagram gradually evolves from an initial fragmented state into a complex network structure and can recreate the debate process at any key decision point.

[0031] Please refer to the attached document. Figure 4 The cognitive load monitoring and adaptive scaffold module adjusts the training difficulty in real time through physiological feedback analysis. This module collects user skin conductance and heart rate variability data using bioelectric sensors embedded in the mixed reality terminal. A sliding window algorithm is used to extract characteristic statistics of the physiological signals, including mean, variance, and power spectral density. The system calculates the real-time cognitive load index using the following formula:

[0032] In the above formula, This is the cognitive load index. The pupil diameter is measured in real time. and These represent the mean and standard deviation of the pupil diameter under baseline conditions. Heart rate variability reflects the regulatory capacity of the autonomic nervous system. The slope of the change in skin conductance response. 、 、 The preset weighting coefficients are 0.4, 0.35, and 0.25, respectively. The system divides the cognitive load index into three levels: low load, suitable load, and overload. When the system determines that the user is in a low load state (index below 20), it instructs the multi-agent logical dialectical engine to increase the depth of rebuttal. When the system determines that the user is in an overload state (index above 80), it automatically triggers an adaptive scaffolding mechanism, highlighting the current logical bottleneck and providing guiding questions by a mentor agent. The scaffolding mechanism includes three modes: heuristic inquiry, logical path simplification, and contradiction focus.

[0033] Please refer to the attached document. Figure 5 The critical thinking evolution evaluation module constructs an evaluation model comprising four levels. Level 1 is logical completeness, which assesses the coherence of a user's reasoning by calculating the proportion of valid reasoning paths within the total number of paths. Level 2 is the strength of evidence, which evaluates the correlation between the factual evidence cited by the user and their conclusion using a pre-defined reliability scale. Level 3 is perspective shift rate, which counts the frequency with which users proactively correct their stance or absorb reasonable objections when facing virtual agents with different logical attributes. Level 4 is metacognitive awareness, which records the time efficiency with which users self-monitor and correct their own logical flaws. The system generates a radar chart of the user's critical thinking ability by performing time-series analysis on the data from these four levels.

[0034] This system also includes a multi-user collaboration module, supporting multiple users wearing mixed reality terminals to enter the same virtual interactive space. In this mode, the system assigns different observation perspectives and logical tasks to different users. For example, User A is responsible for constructing positive arguments, User B is responsible for finding logical flaws, while the virtual intelligent agent acts as a referee, recorder, or logical disruptor. The system ensures that the virtual logical structure seen by all users remains consistent in spatial location and state through a distributed synchronization protocol, with synchronization latency controlled within 20 milliseconds.

[0035] The hardware interface layer drives the sensor array of the mixed reality terminal, including a binocular display, inertial measurement unit, environmental camera, and microphone array. The logic processing and rendering layer executes multi-agent collaborative logic, semantic parsing, and 3D graphics rendering. This layer employs a parallel processing architecture, distributing logic computation and graphics rendering to different processing cores to ensure a stable rendering frame rate of 90 frames per second. The policy management layer controls the training process and evaluates the algorithm's execution, dynamically adjusting the training syllabus based on the user's historical training data. All layers exchange information at sub-second speeds via a high-speed data bus. The data packet format includes a 128-bit header and a 1024-bit payload, along with a checksum to ensure reliable transmission.

[0036] In the specific training process, the system first guides the user to set a debate topic. The multi-agent logical dialectic engine generates initial logical assertions based on this topic. The spatial environment perception and virtual-real mapping module projects virtual agents into the physical space in front of the user. The user expresses their views through voice input. The logical semantic analysis and reasoning chain extraction module instantly parses the voice content and extracts logical elements. The 3D logical structure mapping module draws a logical graph in space in real time. When the user encounters a logical dilemma, the cognitive load monitoring and adaptive scaffolding module senses the increase in the user's physiological stress and automatically pops up logical prompts. At the end of the training, the critical thinking evolution evaluation module summarizes the data from the entire process and outputs an evaluation report. Example 2

[0037] Building upon Example 1, this example extends the conflict logic modeling unit of the multi-agent logic dialectic engine. In this example, a game theory-based conflict logic generation algorithm is introduced. This algorithm is not limited to preset logical paths but dynamically generates optimally challenging debate strategies based on the user's interaction history.

[0038] Please refer to the attached document. Figure 2 When the conflict logic modeling unit operates, it first establishes a state space for the debate topic, where each state represents a truth value combination of logical propositions. The system uses Markov decision processes to simulate the debate process. Each logical assertion of the agent is considered an action, and the user's feedback leads to a state transition. The system guides the agent to find the action sequence that maximizes the user's critical thinking through a reward function. The reward function considers not only whether the user is persuaded, but more importantly, the depth of reasoning demonstrated by the user in responding to the action.

[0039] The specific conflict logic generation process is as follows: After acquiring the core concept set of the training topic, the system expands the relationships between concepts using knowledge graph technology. By sampling subgraphs of the knowledge graph, a set of propositions related to the topic is extracted. When constructing positive argument paths and negative challenge paths for the topic based on predicate logic, the system employs a multi-dimensional logic operator library. In addition to basic conjunction, disjunction, and implication operators, modal logic operators and temporal logic operators are introduced to construct complex arguments involving possibility, necessity, and temporal order.

[0040] To artificially implant specific types of logical weaknesses into the argumentation path, the system has established a database containing over 1000 logical fallacy samples. Based on a preset cognitive challenge level, the system utilizes a generative adversarial network to generate highly concealed logical traps. For example, at low challenge levels, the system implants obvious fallacies of hasty generalization; at high challenge levels, it implants complex fallacies of false causality or composition. These weaknesses are encoded as special nodes in a logic graph, and their metadata includes the identification features of the fallacy and the rebuttal path.

[0041] In this embodiment, the dialogue strategy scheduling unit employs a coordination mechanism based on a social psychology model. When multiple virtual agents participate in a debate simultaneously, the scheduling unit assigns different interaction roles based on the agents' cognitive labels. For example, in a debate on climate change, the scheduling unit assigns a confirmation-biased agent to play the role of a staunch skeptic, while a deductive reasoning agent plays the role of a rigorous scientist. The scheduling unit monitors the balance of the debate in real time; if a user is at a significant disadvantage, the scheduling unit instructs one of the agents to deliberately reveal logical flaws to guide the user to launch a counterattack. Example 3

[0042] Building upon Example 1, this example details the interaction protocol and rendering details of the 3D logical structure mapping module. Please refer to the appendix. Figure 3 This module is not only responsible for visualizing logical elements, but also for providing real-time feedback for human-computer interaction.

[0043] The physical properties of logical proposition nodes in the virtual space are precisely defined. Each node has a collider with a radius of 5 centimeters. When a user's gesture ray enters the collider's range, the node triggers pre-selected feedback, manifested as a 20% increase in edge luminosity accompanied by a slight vibration. The geometric shape mapping logic of the nodes is as follows: Axiomatic nodes are spherical due to their isotropic symmetry, symbolizing unshakeable premises. Hypothetical nodes are semi-transparent cubes, symbolizing the uncertainty of their structure. Conclusion nodes are pyramidal, symbolizing the convergence of logical deductions.

[0044] The light strip connecting logical deduction relationships exhibits dynamic physical characteristics. The thickness of the light strip represents the logical support, ranging from 2 mm to 10 mm. The speed of the light flow represents the deduction frequency of the reasoning chain. When a deduction relationship is determined to be valid by the logical semantic analysis module, the light strip appears as a stable, deep blue stream. When a logical jump exists in the deduction relationship, the light strip appears as intermittent yellow flashes. When a deduction relationship is determined to be a logical fallacy, the light strip appears as a pulsating red. Users can grasp and pull the light strip using gestures to explore the flexible characteristics of the logical structure. This physical simulation enhances the user's intuitive perception of the strength of logical connections.

[0045] The spatial anchoring display process employs a multi-layered rendering optimization strategy. The bottom layer uses environmental occlusion rendering to ensure that virtual logical nodes are correctly occluded by real objects in the physical world, thereby enhancing immersion. The middle layer uses logical topology rendering, employing a force-based layout algorithm to automatically unfold the logical graph in space, avoiding node overlap. The top layer uses interactive UI rendering; when a user interacts with a specific node, a high-resolution text description and evidence chart pops up 30 centimeters in front of the user's line of sight. This module also supports global zooming of the logical graph. Users can use hand gestures to zoom out the entire complex argument model to the size of a desktop for macroscopic review, or zoom in to the size of a room, immersing themselves in navigating between different logical paths.

[0046] In the dynamic temporal backtracking function, the system takes snapshots of the entire logical graph 10 times per second, including the attributes of each node, the connectivity of each edge, and the evidence weight at that time. These snapshots are stored in a circular buffer. When the user triggers a backtracking operation, the system uses a linear interpolation algorithm to smoothly reproduce the logical evolution process. The backtracking interface includes a 30-centimeter-long virtual timeline, with keyframes marking the moments when the user experiences cognitive biases or successfully identifies fallacies. Users can click on these keyframes to directly jump to the corresponding voice recording and agent feedback interface for in-depth metacognitive reflection. Example 4

[0047] Based on Example 1, this example further refines the physiological feedback analysis process of the cognitive load monitoring and adaptive scaffold module. Please refer to the appendix. Figure 4 The system achieves accurate characterization of users' cognitive states through deep fusion of multimodal data.

[0048] The physiological signal acquisition unit obtains skin conductance responses (SCRR) through electrodes on the headband of the mixed reality terminal. The sampling frequency of the SCRR is set to 250 Hz, with a sensitivity of 0.001 microSiemens. Heart rate variability data is acquired through an infrared photoelectric sensor integrated into the forehead contact area of ​​the terminal. The system uses a wavelet transform algorithm to denoise the raw physiological signals and extract feature components relevant to the cognitive task.

[0049] In addition to the cognitive load index formula in Example 1, this example introduces a state discrimination model based on a fuzzy logic inference engine. The inference engine input includes the cognitive load index, task completion rate, and the complexity of the gaze scanning path. The inference engine pre-sets more than 30 fuzzy rules. For example, if the cognitive load index is high and the task completion rate is slow, the user is determined to be in a state of cognitive overload. If the cognitive load index is low and the gaze scanning path is simple, the user is determined to be in a state of cognitive fatigue.

[0050] The adaptive scaffolding mechanism's triggering logic follows a dynamic probability model. The system adjusts not only based on the current load state but also considers the user's load change trends. If the load index has been rising continuously over the past 60 seconds and exceeds a threshold, the system predicts an impending cognitive breakdown and intervenes with scaffolding in advance. The three scaffolding intervention modes have different implementation parameters. The heuristic inquiry mode uses a mentor-like agent to issue pre-set questions in four dimensions: clarifying questions, hypothetical questions, causal questions, and perspective questions. The logic path simplification mode temporarily hides branches unrelated to the current reasoning through a 3D logic structure mapping module, focusing the user's attention on the core path. The contradiction focus mode uses a flashing highlighting method to simultaneously display two conflicting proposition nodes in the user's field of vision.

[0051] This module also possesses learning capabilities. By recording changes in cognitive load and performance improvements after users receive different types of scaffolding interventions, the system uses reinforcement learning algorithms to optimize the allocation weights of the scaffolding. After approximately 10 hours of training, the system can accurately select the most effective combination of scaffolding interventions based on a specific user's learning habits. For example, for visual learners, the system increases the weight of logical path simplification. For auditory learners, the system prioritizes heuristic questioning. Example 5

[0052] Building upon Example 1, this example provides an in-depth description of the quantitative evaluation method for the critical thinking evolution assessment module. Please refer to the appendix. Figure 5 This module not only provides a summary of training results, but also enables fine-grained tracking of thought processes.

[0053] The calculation of logical completeness levels is based on path search algorithms in graph theory. The system defines all possible valid reasoning paths from the initial premise set to the core conclusion point. The user's actual reasoning process is mapped to a subgraph in the graph. Logical completeness is determined by the ratio of the number of edges in the subgraph to the number of edges at the corresponding level in the entire graph. The system also identifies isolated nodes and logical breaks in the user's reasoning, which are recorded as deductions for logical completeness.

[0054] The assessment of evidence strength levels incorporates an external reliability database. When a user cites evidence, the system extracts key attributes such as source reliability, timeliness, and relevance using natural language processing algorithms. The system then categorizes the evidence. Level 1 consists of peer-reviewed scientific data with a reliability coefficient of 0.95. Level 2 comprises authoritative media reports with a reliability coefficient of 0.8. Level 3 includes personal experience or unverified anecdotes with a reliability coefficient of 0.3. The user's overall evidence strength is calculated as the weighted average of the reliability coefficients of all cited evidence.

[0055] The perspective shift rate is assessed by analyzing the changes in users' stance tendencies over time. The system pre-sets at least four different stance coordinates for each topic. Using sentiment analysis and semantic clustering techniques, the system calculates the user's stance position within the coordinate system in real time. When a user's stance coordinates shift significantly and move towards a more reasonable suggestion after hearing a counter-argument from the virtual agent, the perspective shift rate count increases. This metric reflects the user's openness and ability to correct cognitive biases.

[0056] The assessment of metacognitive awareness focuses on the user's awareness of their own errors. The system timestamps the moment the logical semantic analysis module identifies a logical fallacy in the user's argument. If the user subsequently corrects the fallacy, the system calculates the time interval from the fallacy's occurrence to its correction. The shorter the time interval, the stronger the metacognitive monitoring capability. If the user corrects the error only after being prompted by the agent, the weight is reduced accordingly.

[0057] The system uses multivariate time-series analysis on the data from the four levels mentioned above, and leverages long short-term memory networks to predict the user's performance in future training. The five axes of the capability radar chart represent logical rigor, evidentiary strength, perspective flexibility, cognitive monitoring, and depth of reflection. Scores for each axis are normalized and range from 0 to 100. The system also generates a 2000-word diagnostic report detailing the user's typical thinking preferences and key logical flaws requiring improvement during training, and recommends personalized training materials for future use. Example 6

[0058] This embodiment describes the data security and privacy protection mechanisms of this system under a layered collaborative computing architecture. The underlying hardware interface layer employs edge computing technology when collecting multimodal user behavior data. Sensitive raw data such as gaze trajectories, gesture details, and speech waveforms are anonymized locally on the terminal device. For example, gaze data is converted into discrete gaze frequencies and coordinate offsets, no longer retaining the original eye image. Speech data is converted into text vectors and acoustic feature spectra, no longer retaining the original audio.

[0059] The logic processing and rendering layer employs an encrypted computing protocol when executing multi-agent collaborative logic. The virtual agent's profile data and the user's logical reasoning chain are protected during transmission using a 128-bit symmetric encryption algorithm. The system allocates an isolated region in memory specifically for handling logical semantic analysis tasks, preventing data leakage caused by memory overflow attacks.

[0060] The strategy management layer utilizes a decentralized storage solution when storing user evaluation data. Each user's critical thinking ability development trajectory is encrypted and stored in a distributed database, with each user possessing a unique private key for access control. When generating evaluation reports, the system only uses anonymized statistical indicators, ensuring that educators cannot deduce users' raw physiological privacy data while obtaining teaching feedback. Furthermore, the system has a built-in abnormal behavior detection program. If unauthorized sensor access commands or abnormal data flow are detected, the system will cut off the power supply to the hardware interface layer within 5 milliseconds and issue an alert to the strategy management layer. Example 7

[0061] Building upon Example 1, this example describes the application of a multi-user collaboration module in a large-scale debate scenario. Please refer to the appendix. Figure 1 The system supports up to 50 users online simultaneously through a message bus based on a publish-subscribe model.

[0062] In this scenario, the system constructs a hierarchical virtual interactive space. The central area serves as the core debate zone, displaying the core logic graph of the current debate topic. The surrounding areas are for information retrieval and observation. The system assigns different permissions to users with different roles. Debaters have the authority to modify logic nodes. Observers can like or mark logic nodes; this feedback is overlaid on the logic graph in the form of a heatmap, providing debaters with real-time audience feedback pressure.

[0063] In multi-player mode, the virtual agents perform more complex functions. One agent labeled "judge" uses a logical semantic analysis module to perform real-time logical compliance checks on the statements of each participant. Whenever a participant commits a logical fallacy, the judge agent immediately generates a prominent red warning mark in the 3D logical structure mapping module. Another agent labeled "logic disruptor" is responsible for posing provocative questions when the debate reaches a stalemate, sparking new ideas.

[0064] The system ensures real-time collaboration through a dynamic bandwidth allocation algorithm. As user density increases, the system automatically reduces the rendering complexity of the background virtual scene, prioritizing bandwidth requirements for logical node positioning and user gesture synchronization. The system also supports multi-user collaborative logic decomposition tasks. A group of users can work together to decompose a complex philosophical proposition, with each person responsible for a branch. The system automatically merges the decomposition results from all parties, uses a logical consistency check algorithm to identify conflicting viewpoints between different users, and highlights these conflicts as key points for further discussion.

[0065] Through the detailed descriptions of the above embodiments, this invention demonstrates a critical thinking training system that deeply integrates mixed reality technology, multi-agent collaboration, and multimodal perception. This system not only solves the learner's dependence on algorithms but also achieves personalized and efficient cognitive training through logical visualization and adaptive intervention. The modules work closely together to form a complete closed loop from data perception, logical analysis, interactive display to quantitative evaluation. This system has broad application prospects in fields such as education and training, decision support, and psychological counseling.

[0066] During system operation, every parameter setting underwent rigorous engineering verification. For example, the accuracy of eye tracking must be maintained within 0.5 degrees to ensure that the false trigger rate when the user selects a logical node is less than 1%. The latency of speech recognition must be controlled within 500 milliseconds to ensure that the virtual agent's rebuttals have a conversational feel. The refresh rate of the 3D logic graph is set to 90 Hz to eliminate the motion lag felt by the user when grasping nodes. These technical details collectively ensure the system's stability in practical applications and a deeply immersive user experience.

[0067] 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 it. 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. 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. Those skilled in the art can flexibly combine the technical features of the different embodiments described above according to actual needs to construct a critical thinking training environment adapted to different application scenarios.

Claims

1. A multi-agent critical thinking training system based on mixed reality, characterized in that, include: The spatial environment perception and virtual-real mapping module is used to perform real-time spatial scanning through the depth sensor integrated into the mixed reality terminal to generate point cloud data of the physical environment. Simultaneous localization and mapping algorithms are used to determine the 6-DOF pose of the mixed reality terminal in space; virtual debate scenes, logic nodes and virtual intelligent agents are fixed at specific positions on the physical surface through spatial anchoring technology; and infrared sensors and visible light cameras are used to capture the user's gaze focus, gestures and facial expressions to convert multimodal behavioral data into interactive commands that the system can recognize. A multi-agent logical dialectic engine is used to construct multiple virtual agents with different logical attributes and knowledge backgrounds, and generate conflicting logical assertions and debate strategies based on the current training topic. The logical semantic analysis and reasoning chain extraction module is used to perform semantic parsing on the text or speech input by the user, extract the core assumptions, reasoning paths and conclusions in the user's argumentation process, and use a logical fallacy classifier based on a converter architecture to match the user's argument with 12 preset logical fallacy templates to identify the logical fallacy types. The 3D logical structure mapping module is used to transform the logical elements extracted by the logical semantic analysis and reasoning chain extraction module into a visualized spatial topology structure, and to display it in a spatially anchored manner in the virtual interactive space; the cognitive load monitoring and adaptive scaffolding module is used to assess the cognitive load level of users in real time based on their physiological feedback and interactive performance during the training process, and to dynamically adjust the challenge intensity and the level of detail of logical prompts of the virtual intelligent agent accordingly. The Critical Thinking Evolution Evaluation Module is used to quantitatively evaluate a user's logical rigor, evidentiary validity, preference recognition ability, and metacognitive monitoring ability throughout the training cycle. The system operates within a layered, collaborative computing architecture; The layered collaborative computing architecture includes a bottom hardware interface layer, a middle logic processing and rendering layer, and a top policy management layer; the bottom hardware interface layer is responsible for driving the sensor array of the mixed reality terminal. The intermediate logic processing and rendering layer is responsible for executing multi-agent collaboration logic, semantic parsing, and 3D graphics rendering; the top-level policy management layer is responsible for controlling the training process and evaluating the algorithm's operation.

2. The multi-agent critical thinking training system based on mixed reality according to claim 1, characterized in that, When the spatial environment perception and virtual-real mapping module collects multimodal behavioral data: gaze tracking data includes pupil position and gaze duration; gesture data includes the spatial coordinates of 21 key skeletal points; facial expressions are parsed into discrete emotional features based on action units; the multimodal behavioral data, after normalization processing, serves as the raw input for logical reasoning and cognitive assessment.

3. The multi-agent critical thinking training system based on mixed reality according to claim 1, characterized in that, The multi-agent logical dialectical engine includes a role profile generation unit, a conflict logic modeling unit, and a dialogue strategy scheduling unit. The role profile generation unit assigns a unique cognitive label to each virtual agent. The cognitive label includes logical tendencies, professional domain background, and debating style. The logical tendencies include confirmation bias, skepticism, deductive reasoning, and empiricism. The conflict logic modeling unit retrieves mutually exclusive arguments from the knowledge base for a preset training topic and uses a generative language model to construct a structured argument that includes preconditions, logical operators, and conclusions. The dialogue strategy scheduling unit schedules virtual agents with different cognitive labels to participate in the debate in real time, based on the user's argumentation logic.

4. The multi-agent critical thinking training system based on mixed reality according to claim 3, characterized in that, The working process of the conflict logic modeling unit is as follows: It acquires a set of core concepts for the training topic and expands the relationships between concepts using knowledge graph technology; it constructs positive argumentation paths and negative challenge paths for the training topic based on predicate logic; according to a preset cognitive challenge level, it implants specific types of logical weaknesses into the argumentation paths, and these logical weaknesses are encoded as special nodes in the logic graph that have identification features and rebuttal paths; it converts the logic paths into natural language expressions and distributes them to virtual agents with corresponding attributes for expression.

5. The multi-agent critical thinking training system based on mixed reality according to claim 1, characterized in that, When identifying logical fallacies, the logical semantic analysis and reasoning chain extraction module adopts a multi-feature fusion judgment method: the system extracts semantic features at the text level, and combines them with the acoustic features of the user's intonation fluctuations and stress distribution when speaking, as well as the logical node position features pointed to by gestures during the argumentation process, to construct a multi-dimensional feature vector. The multidimensional feature vectors are input into a pre-trained deep belief network. Based on the conditional probability distribution of given text features, acoustic features, and gesture features, the probability that a user's argument belongs to a specific type of logical fallacy is calculated to identify the type of logical fallacy. The extracted reasoning chain is stored in a structured form as a directed graph, where nodes represent statements, edges represent logical deduction relationships, and each edge is associated with a weight value representing the strength of the deduction.

6. The multi-agent critical thinking training system based on mixed reality according to claim 1, characterized in that, The process of mapping the reasoning chain by the 3D logical structure mapping module is as follows: Proposition nodes are assigned different geometric shapes according to their importance in the argument, with axiomatic nodes being spherical, hypothetical nodes being semi-transparent cubes, and conclusion nodes being pyramidal; the luminous intensity of the proposition node is positively correlated with the evidence support of the proposition; logical deduction relationships are connected by light strips with a sense of flow, the thickness of the light strips representing the logical support, and the movement speed of the light flow representing the deduction frequency of the reasoning chain.

7. A multi-agent critical thinking training system based on mixed reality according to claim 6, characterized in that, The 3D logical structure mapping module is also used to: respond to user gestures to grasp, scale, or reorganize logical nodes in space; when a user clicks on a specific proposition node, pop up the underlying assumptions on which the proposition depends and the corresponding evidence tracing interface; execute dynamic time-series backtracking function, automatically record the logical evolution trajectory during the training process, and respond to the user's sliding operation on the virtual timeline, and use linear interpolation algorithm to reproduce the debate process at a specific decision point.

8. The multi-agent critical thinking training system based on mixed reality according to claim 1, characterized in that, The cognitive load monitoring and adaptive support module assesses cognitive load levels through the following process: collecting user skin conductance response and heart rate variability data through bioelectric sensors embedded in the mixed reality terminal; and extracting feature statistics of physiological signals using a sliding window algorithm, including mean, variance, and power spectral density. The difference between the real-time measured pupil diameter and the mean pupil diameter under the baseline state is obtained and standardized, and used as the first feature component; the reciprocal of the heart rate variability rate is obtained as the second feature component; and the slope of the change in skin conductance response is obtained as the third feature component. The cognitive load index is obtained by weighted summation of the first, second, and third characteristic components.

9. A multi-agent critical thinking training system based on mixed reality according to claim 8, characterized in that, The cognitive load monitoring and adaptive scaffolding module is also used to: classify the cognitive load index into three levels: low load, suitable load, and overload; when the user is in a low load state, instruct the multi-agent logic dialectic engine to increase the depth of rebuttal; when the user is in an overload state, trigger the adaptive scaffolding mechanism, highlight the current logical bottleneck point in the 3D logic structure mapping module, and provide guiding questions by a mentor-type agent; the adaptive scaffolding mechanism includes three modes: heuristic inquiry, logical path simplification, and contradiction point focusing. The system uses a reinforcement learning algorithm to assign weights to the three modes based on the user's historical performance.

10. A multi-agent critical thinking training system based on mixed reality according to claim 1, characterized in that, The critical thinking evolution evaluation module constructs an evaluation model with four levels: Level 1 is logical completeness, used to calculate the proportion of valid reasoning paths in the total paths in a user's argument; Level 2 is evidence strength, used to assess the correlation strength and reliability level between the factual evidence cited by the user and the conclusion; Level 3 is perspective switching rate, used to count the frequency with which users correct their own positions or absorb dissenting opinions when facing virtual intelligent agents with different logical attributes; Level 4 is metacognitive awareness, used to record the time efficiency of users in self-monitoring and repairing their own logical flaws; The system also includes a multi-user collaboration module, used to support multiple users entering the same virtual interactive space, and uses a distributed synchronization protocol to ensure that the virtual logical structure seen by all users remains consistent in spatial location and state.

Images