Background and significance of the project

In human society, intelligent individuals collaborate efficiently and safely through social cognitive abilities such as understanding of others' purposes, moral judgment, and emotional empathy. However, today's AI systems are still deficient in these dimensions of social intelligence, especially in multi-agent interaction scenarios. For example, it will be the norm for multiple AIs to plan, cooperate, and even compete at the same time in real-world applications, but a lack of inference about the purpose of others, moral assessment of the consequences of behavior, and understanding and response to emotions can lead to conflict, distrust, and even security risks. At present, some researchers have developed multi-agent simulation platforms such as "large-scale social simulators" to deduce the impact of group decision-making and assist social governance through value modeling and digital twin technology. This indicates that multi-agent social-level intelligence is becoming an important direction for the development of artificial intelligence. Therefore, endowing artificial intelligence with brain-like social cognitive capabilities not only has great academic value, but also an important way to ensure the safety and controllability of complex AI systems and benefit society.

Advances in brain science have provided inspiration for artificial intelligence: neuroscience research has shown that specific mechanisms in the human brain underpin higher-order social intelligence functions. For example, the ability of the Theory of Mind (ToM) enables humans to speculate about the mental state of others (beliefs, purposes), depending on the role of regions such as the prefrontal lobes of the brain; Empathy is related to physiological mechanisms such as the mirror neuron system, which allows us to perceive and share the emotions of others. More importantly, ToM and empathy together form the basis of human moral cognition—ToM provides cognitive understanding of others, empathy provides emotional drive, and the two synergistically shape moral judgments. These findings suggest that we can draw on the brain-like principle to construct an artificial social intelligence algorithm mechanism, so that AI can have human-like purpose recognition, ethical reasoning, and emotional understanding capabilities in a multi-agent environment.

From the perspective of application, AI with social intelligence will greatly expand the application depth and trust of multi-agent systems. In terms of human-computer interaction, social intelligence makes agents more empathetic and ethical, which can be used to accompany robots, intelligent customer service, etc., to improve user satisfaction and acceptance. In multi-robot collaboration and even unmanned system clusters, the addition of purpose inference and moral constraints can reduce internal friction conflicts and ensure the safety and reliability of collaborative tasks. In key fields such as medical care and education, the introduction of AI's empathy and moral judgment capabilities is expected to promote doctor-patient trust and teacher-student interaction, forming a new paradigm of "AI + human" collaboration and efficiency. Overall, based on the major needs of the country, this project aims to break through the bottleneck of artificial intelligence social cognition, build brain-like social intelligence algorithms and systems, and promote the safe, ethical and sustainable development of multi-agent technology.

Project objectives and key issues

The overall goal of this project is to build brain-inspired social reasoning capabilities through the DIKWP semantic system, and to develop brain-inspired social intelligence algorithms and prototype systems for multi-agent interaction. Specifically, the following three aspects of brain-like social intelligence capabilities are given to artificial intelligence, and on this basis, they are integrated and verified:

• Purpose Modeling and Understanding: Develop the machine's "Theory of Mind" function to enable agents to infer the implicit goal and belief states of others from interactions, and support prediction of others' behavior and appropriate responses.

• Moral reasoning and decision-making: The agent is given a mechanism for moral evaluation when making decisions, comprehensively considering the ethical norms and social consequences of behavior, and realizing the constraints on one's own behavior and the evaluation of others' behavior. So that AI decision-making is not only "effective" but also "ethical".

• Affective empathy and response: Enables agents to perceive and understand the emotional states of humans or other agents, appropriately produce empathetic responses, and incorporate this into interaction strategies to show human-like empathy and emotional care.

Key scientific questions that need to be addressed around these goals include:

(1) How to establish a unified knowledge representation and computational framework that integrates data, knowledge, and high-level semantics such as "Purpose" to support complex reasoning in the social context?

(2) How to simulate the reasoning mechanism in the brain that implies the purpose of others, and realize the purpose recognition and prediction of artificial intelligence?

(3) How to formalize human moral norms into computable rules or models, and combine them with machine learning to make agents embed ethical constraints in self-directed learning?

(4) How to simulate and measure the empathy ability of artificial intelligence so that it can respond appropriately to changes in the emotions of others and influence decision-making?

(5) How to ensure that the above process is explainable and trustworthy to avoid becoming a "black box"? The solution of these problems will provide a theoretical and technical basis for injecting brain-like social intelligence into multi-agent systems, and it is also the core of the innovation of this project.

Theoretical basis and innovation points

In order to realize brain-like social intelligence, this project will draw on the previous research results of the application team in artificial consciousness and semantic intelligence, and propose an innovative framework that integrates multi-level semantic representation, elastic semantic network and self-explanatory mechanism. Its theoretical basis mainly includes:

• DIKWP Semantic Model: Prof. Yucong Duan's team proposed the DIKWP model, which extends the classic information processing pyramid DIKW (Data-Information-Knowledge-Wisdom) and introduces a new element of "Purpose (Purpose)". DIKWP uses a hierarchical semantic structure to describe the cognitive process of agents: from raw data, through information extraction and knowledge condensation to wisdom decision-making, and the top-level Purpose drives and oversees the lower-level processes. This extension enables the model to embody a goal-oriented cognitive flow and to represent the internal motivation and purposiveness of the agent. The DIKWP model provides unified semantic coordinates for cognitive computing and realizes the semantic link from perceptual data to high-level decision-making, which is the fundamental framework for this project to build social reasoning ability.

• Mesh Semantic Elastic Network: Different from the static hierarchical model, DIKWP will be implemented as a mesh semantic network in this project, and each hierarchical element forms an elastic connection through two-way interaction. Professor Yucong Duan proposed that in the mesh DIKWP model, there are 25 interactive modules between the five layers of data, information, knowledge, wisdom and purpose, so as to realize the two-way flow and feedback of semantics. This semantic elastic network allows conceptual boundaries to dynamically adjust to the context, and semantic associations are flexible and robust (i.e., semantics elastically change with context). This mechanism helps agents deal with semantic uncertainty in complex social environments, such as the change of meaning of the same behavior in different contexts, the elasticity of concept extension, etc. The introduction of the Semantic Elastic Network will ensure that our agents can flexibly update their understanding of other people's states and environmental contexts in multi-agent interactions, and are not bound by rigid rules.

• ACPU architecture (artificial consciousness processing unit): In order to support the efficient operation and interpretability of the above semantic computing model, we introduce the artificial consciousness computing architecture proposed by Yucong Duan's team。 The architecture includes a dedicated Artificial Awareness Processing Unit (ACPU) and a supporting operating environment and language to achieve integrated software and hardware design. The core idea is to integrate the support of state of consciousness and semantic concepts on the basis of classical computer architecture, and break through the limitations of traditional systems in cognitive computing. Specifically, ACPU supports the semantic flow of the DIKWP model at the hardware level (such as setting up processing units and control flows such as data, knowledge, and purpose in the chip), making the execution of artificial consciousness algorithms more efficient and controllable. This architecture is expected to improve the real-time inference capabilities of our social intelligence systems and lay the foundation for the development of specialized chips in the future. More importantly, the ACPU architecture aims to make AI systems "artificially aware" such as explainable, trustworthy, and consistent behavior. This explainable AI philosophy permeates the design of our systems, ensuring that the algorithmic decision-making process is transparent and traceable.

• Self-explanatory mechanism: In view of the high complexity and high risk of social intelligence decision-making, we have integrated a self-explainable mechanism into the architecture. Specifically, an interpretation module is set up inside the agent to perform metacognitive monitoring and causal chain extraction for key decisions, and generate human-understandable interpretation outputs. For example, when an agent makes a moral evaluation of another person's behavior or chooses a response, the system will simultaneously provide an explanation that clarifies the purpose assumptions, moral principle basis, and emotional considerations in its reasoning process. On the one hand, this self-explanatory ability improves the transparency and credibility of the system, and responds to the requirements of AI ethics for explainability. On the other hand, it also helps agents to engage in reflective learning and improve future performance by examining their own decision-making logic. This project will explore technologies such as generative language models to enrich the expression of explanations and make them easy for humans to understand, so as to build a truly "understandable" social agent.

In summary, the theoretical innovation of this project is embodied in the combination of the DIKWP artificial consciousness model and the multi-agent social intelligence needs, and the creative introduction of purpose-driven semantic representation, mesh elastic semantic network and interpretable artificial consciousness architecture to form a new paradigm of "comprehensive theory + technical framework". This paradigm is different from traditional pure data-driven or symbolic AI methods, emphasizing multi-level semantic fusion and cognitive mechanism biomimicry, which is expected to provide breakthroughs for AI to achieve advanced social behavior.

Brain-inspired social intelligence algorithm framework design

Based on the above theoretical basis, we design a brain-inspired social intelligence algorithm framework to endow agents with the ability to conduct social reasoning in human-computer/inter-machine interaction. The framework is shown in Figure 1 (omitted), using the DIKWP semantic system as the overall architecture: the bottom layer is the data layer (D) obtained by environmental awareness, which gradually rises to the information layer (I) of information processing and the knowledge layer (K) of knowledge abstraction, and forms the Wisdom layer (W) of comprehensive decision-making at the top level, and is composed of the top layer The Purpose layer (P) provides drive and regulation. In the framework, each layer is connected by a semantic network to support the two-way dynamic flow of cognitive processing: on the one hand, the high-level purpose and moral principles can constrain the bottom-level behavior choice from the top down; On the other hand, the underlying perception and interactive feedback update the cognitive state of the upper level from the bottom up. Focusing on the framework, we focus on designing the following core functional modules:

Purpose Modeling Mechanism (Brain-like "Theory of Mind")

The Purpose Modeling Module is designed to equip agents with the ability to understand and predict the behavior of other agents, which is equivalent to the theory of mind function of a machine. In the DIKWP framework, this module mainly involves the interaction between the knowledge layer and the purpose layer: the agent starts from the perceived behavior data of others, extracts key information, rises to the knowledge layer to form a representation of the state of others, and in the The Purpose layer generates assumptions about the goal/purpose of others. To achieve this, we will fuse a probabilistic graph model with deep learning techniques: construct a probability map of another person's mental state (including possible beliefs, wishes, purposes, etc.), and train a deep neural network to update the distribution of this graph from the observation sequence. Each agent maintains a mental model of the other body, which is constantly updated with interactions. It is worth noting that we introduce a hypothesis-testing reasoning process: the agent generates multiple hypotheses about the purpose of others based on current information, and tests the plausibility of these assumptions by predicting the next behavior of others, gradually converging to the best hypothesis. This process is similar to the repeated correction mechanism when humans speculate on the purpose of others, and it is also a manifestation of the brain-like characteristics.

In terms of algorithm implementation, the method based on rules and Bayesian inference can be used to design algorithm judgment rules for typical scenarios (for example, a predefined purpose library can be matched by other people's action sequences). With the accumulation of data, reinforcement learning or meta-learning mechanisms are introduced, so that agents can learn more complex purpose inference strategies autonomously through trial-and-error interaction. We draw on Stanford's "Hypothetical Minds" framework, in which a dedicated ToM module generates hypotheses about other agents' strategies and goals, improving adaptability in multi-agent collaboration and adversarial scenarios. Experiments show that the performance of this kind of agent based on mental modeling is significantly better than that of traditional deep reinforcement learning agents in competition and cooperation environment. Inspired by this, the Purpose modeling module of this project will make full use of large models and knowledge graph technology to achieve highly generalized Purpose recognition and behavior prediction of others in a complex environment with multiple agents.

Mechanisms of Moral Reasoning (Embedded Ethical Decision-making)

The moral reasoning module gives agents the ability to make ethical judgments when making decisions, ensuring that their actions are in line with the moral norms and values of human society. In the DIKWP architecture, moral reasoning mainly acts on the Wisdom layer (W): as an evaluation constraint mechanism for final decision-making, it evaluates and scores the moral value of alternative behavior schemes, and influences the goal setting of the Purpose layer (e.g., eliminating immoral goals). To construct this module, we will integrate symbolic logical reasoning and data-driven learning methods to form a blended ethical decision-making system.

First, a formal moral knowledge base is established at the symbolic level. Refer to ethical theories (such as responsibility ethics, consequentialism, virtue ethics, etc.) to extract the principles and rules applicable to AI behavior. For example, classical ethical norms (the principle of non-harm, the principle of fairness, etc.) can be encoded as logical rules, or a system of rules for tolerance of non-compliance (defeasible rules) can be used to deal with conflict of principles in complex situations. Secondly, a value-aligned learning method is introduced at the learning level, so that agents can learn moral preferences from human feedback. Specifically, reverse reinforcement learning (IRL) can be used to summarize the implicit ethical norms from human behavior, or reinforcement learning training with human preference feedback can be used to make agents consider the reward and punishment of the moral dimension when optimizing multiple objectives. Previous reviews have shown that the current popular practice is often limited to pure rules or pure learning, and lacks an intermediate solution that is both adaptable and controllable. Therefore, the innovation of this project is to design a hybrid architecture at the top level: the agent not only refers to the built-in ethical rules when making decisions, but also dynamically adjusts the weight of moral preferences through learning, so as to achieve robust and explainable moral reasoning.

In specific applications, the Moral Reasoning module will be used to assess the social impact of multi-agent action scenarios. For example, in the group decision-making of unmanned vehicles, the introduction of moral evaluation can avoid undesirable group behaviors such as vehicle rushing. In medical AI, embedding ethical judgments prevents decisions that violate medical ethics (e.g., excessive diagnosis and treatment or privacy breaches). We plan to design a series of moral dilemma scenarios (such as the famous tram problem) to test the agent's ability to make ethical decisions to ensure that it meets human moral expectations when weighing its own interests against the interests of others. At the same time, with the help of the method of moral judgment agency, an independent adjudication sub-module is used to monitor the agent's behavior and give ethical judgment from the perspective of a third party. The result of this ruling can be used to correct the decision-making of the main agent and enhance the security and credibility of the system architecture.

Affective Empathy Mechanism (Artificial Empathy Computation)

The Affective Empathy module enables agents to perceive the emotional states of others and generate corresponding emotional responses, simulating human empathy. In the DIKWP architecture, the empathy process involves the extraction of emotional cues from others at the information layer (I) and the representation of emotional semantics by the knowledge layer (K), and finally feedback to influence the decision-making of the Wisdom layer (e.g., adjusting the original plan due to empathy). Our implementation will draw on affective computing and cognitive psychology models to give agents the ability to infer emotions from multimodal information and regulate their own behavior to respond to the emotions of others.

First of all, in terms of emotion perception, we will use multimodal emotion recognition technology: combined with text semantic analysis, voice intonation features, facial expression recognition (during human-computer interaction) or other proxy behavior signals, to judge the category and intensity of others' emotions at the information layer. For purely virtual agent interaction situations, emotions can be inferred by deviations from their behavioral strategies (e.g., sudden conservatism may indicate fear). Secondly, in terms of emotional response, we design an empathy strategy for the agent: when the agent detects that the other person is in a certain emotional state, the agent simulates the corresponding emotional resonance through internal state regulation (for example, when the other person is depressed, he or she feels sympathy and comforts), and considers this emotion in decision-making (such as slowing down the pace of competition, providing help, etc.). This mechanism is similar to the phenomenon of members of a human team influencing each other's emotions to coordinate and cooperate.

Algorithmically, we will construct an artificial empathy model that integrates cognitive empathy (understanding the reasons for others' emotions) and emotional empathy (empathizing with others' emotions). One possible solution is to use a fuzzy logic system to deal with the uncertainty of emotional information: for example, to take other people's emotions and situations as fuzzy inputs and deduce the appropriate level of response. Previous studies have suggested that integrating empathy into multi-agent decision-making can significantly improve the efficiency of team collaboration. Our system will verify this in a multi-agent collaboration task: the group of agents expected to have empathy mechanisms outperform cold rational agents in terms of communication costs, conflict resolution, and synergistic benefits. A study pointed out that the introduction of an empathy layer for autonomists to evaluate and respond to the state of others can provide a dynamic way of information sharing, which is especially beneficial for large heterogeneous agent teams to complete tasks efficiently. Based on this, an empathy algorithm is designed to enable agents to optimize the group decision-making process through the "emotion perception-feedback" loop.

Semantic fusion vs. self-explanatory integration

The above three mechanisms, Purpose, Morality, and Empathy, do not exist in isolation, but need to be organically integrated and interacted within the semantic framework of DIKWP. We will build a semantic blackboard or shared knowledge graph as the center, and the agent's various cognitions of the environment and others (including others' purpose assumptions, moral evaluations, and emotional states) will be uniformly represented in the knowledge layer, which will be comprehensively considered by the wisdom layer when making decisions. For example, when the Purpose Reasoning module judges that a partner's purpose is "asking for help" and the empathy module detects that the partner is depressed, the moral reasoning module will assign a higher ethical value to "help the partner", and finally the Wisdom layer chooses to assist in action. This decision-making method of multi-dimensional factors is in line with the human social behavior model, and also gives full play to the advantages of the context association of the DIKWP model.

At the same time, all key inference processes record their semantic inference links through the self-explanatory module. As soon as the agent makes an important decision or behaves abnormally, the system generates an explanatory report that states: "Because the agent predicts that the other person's purpose is X, and according to the moral principle Y, and perceives its emotion Z, the agent chooses action Q." “。 This interpretation includes both purpose, morality, and emotion, as well as their specific impact on decision-making, enabling human supervisors to understand the logic behind AI behavior. The interpretation module will be reviewed by experts during the testing phase of the system to iteratively improve its clarity and accuracy, ensuring that it ultimately reaches a level that can be used for safety oversight and user communication.

Through the above framework design, we will implement a brain-like agent system with deep semantic understanding (modeling others and situations through DIKWP), value orientation (embedding ethical norms), and emotional intelligence (empathy-driven coordination). It is capable of social reasoning and decision-making close to the human level in complex interactions of multiple agents, and interprets its own reasoning process in a way that is understandable to humans. This lays the architectural foundation for subsequent system implementation and application.

System platform construction and typical application scenarios

In order to verify and apply the above algorithm framework, this project will build a brain-like social intelligence system platform, including a simulation environment, algorithm software and demonstration application prototypes. The content of platform construction is divided into the following aspects:

• Multi-agent semantic interaction simulation platform: Develop a visual simulation platform that supports real-time interaction of multiple agents. The platform will be presented in the form of a virtual environment, which can simulate different scenarios such as social, medical, and education, and allow the embedding of the agent algorithm developed in this project. We will integrate a physics simulation engine and interactive interfaces to enable virtual agents to perceive changes in the environment, communicate with humans or other agents (verbal conversations or non-verbal signals), perform action tasks, and more. The platform will also have built-in semantic monitoring and visualization tools to display the changes in the internal state of each agent's DIKWP model in real time, such as the purpose reasoning process, moral evaluation values, emotional parameters, etc. This helps developers debug algorithms and visualizes the "ideas" and "emotions" of AI to domain experts. The simulation platform will serve as a testbed for the integration and verification of various technologies of the project, and will be scalable to continuously add new scenarios and new agent models.

• DIKWP multi-agent semantic interaction system: Develop system software to support the operation of DIKWP model, including core components such as semantic network management, central information blackboard, and interpreter. On the one hand, the system provides an internal cognitive computing framework for each agent (implementing the processing units and interaction logic of each layer of DIKWP), and on the other hand, manages the communication protocol and shared environment interaction between multiple agents. We will define a semantic communication language that allows agents to pass high-level semantic information between them (e.g., sharing part of knowledge, requesting assistance, etc.), beyond the limitations of traditional multi-agents exchanging simple signals. This semantic interaction mechanism will significantly improve the efficiency of group collaboration and the ability to solve complex tasks. In addition, the system also includes a safety sandbox module to restrain agent behavior in the simulation to avoid loss of control, and a logging module to completely record the decision trajectory and interpretation of the agent in each experiment, which is convenient for offline analysis and continuous improvement.

On the above platform, we will select three typical application scenarios for technical verification, each of which corresponds to major practical needs and can fully reflect the value of brain-like social intelligence. The specific scenarios and demonstration schemes are as follows:

Scenario 1: Social service robots

Scenario description: Several intelligent service robots interact with human users in a semi-structured environment, for example, in the nursing home care scenario, the robots undertake tasks such as companionship and communication, reminding to take medication, and emergency calls, and multiple robots also need to work together to complete tasks such as handling. The human-machine ratio may be many-to-many, and the environment is unpredictable.

Demands and challenges: Service robots need to deeply understand the needs and purposes of elderly users (e.g., whether the user's fall is an accident or attracting attention), abide by medical and nursing ethics (respect the user's privacy and dignity, do not read the user's personal belongings at will, etc.), and make appropriate comfort to the user's emotional changes (such as soft comfort when the user is frustrated). In the event of an emergency (someone falls), which robots are responsible for comforting and which ones are going to pick up the medical kit need to have moral priorities (saving people over daily chores) and purpose coordination. Traditional robots lack these social intelligences, making it difficult for them to cope with this complex scenario.

This project is to embed DIKWP social agents in each robot. When interacting with the elderly, the robot obtains user behavior and expression data through voice and visual perception, identifies the user's emotions (peace of mind/anxiety) at the internal information layer, and speculates on its current purpose (whether it needs help) at the knowledge layer. The Purpose layer, combined with the built-in care purpose drive (the robot has a care purpose), determines the next course of action. If a conflict situation is detected (e.g., two elderly people asking for help at the same time), the ethics module makes decisions based on the principles of risk and fairness: priority is given to those who are at high risk in an emergency, but also shows empathy by verbally reassuring the other person and asking them to wait. Multiple bots negotiate a division of labor through the platform's semantic communication: one bot sends a message "I'm in charge of patient A, you're in charge of patient B" with a moral assessment of both, ensuring that all requests for help are covered and resources are allocated fairly. During the whole process, the self-explanatory module records the decision-making basis of each robot, such as "Because it is detected that elderly person A may have a fracture (Purpose = medical treatment is needed), follow the principle of emergency priority, I will go immediately; At the same time, inform the companion to take care of the elderly person B. This explanation can later be used by caregivers to review whether the robot's response is appropriate.

Argument: Through simulation experiments and actual scenario tests, we expect that compared with traditional rule-based robots, robots that introduce social intelligence will significantly improve user satisfaction, emergency response success rate and other indicators, and there will be no ethical violations. User surveys are expected to show that older adults have a higher sense of trust in robots with emotional empathy and ethical principles, and are willing to accept their care. The scenario verification will prove that the technology of this project has practical value in real human-computer interaction.

Scenario 2: Intelligent medical collaboration

Scenario description: In the hospital diagnosis and treatment scenario, AI diagnostic assistants and nursing robots are introduced to form a medical team with doctors and nurses. Multiple AI assistants need to work together to handle several patients, participate in decision-making throughout the entire process, from consultation and diagnosis to medication, while working closely with human care.

Needs and challenges: Medical AI must strictly abide by medical ethics (such as informed consent, privacy protection, and avoiding decision-making bias), understand the purpose of doctors' diagnosis and treatment, and provide valuable references, rather than mechanically giving conclusions that conflict with people. At the same time, it is necessary to be sensitive to the patient's emotions: in the face of anxious patients, the AI should have soothing wording in answering questions; When the patient is in pain, the nursing robot should adjust the intensity of the movement and tone to express care. Multiple AI assistants also need to collaborate on resource scheduling and diagnosis and treatment plans to avoid giving conflicting recommendations and confusing doctors.

This project solution: Deploy the Purpose recognition module in the AI diagnostic assistant, which can infer the doctor's concerns from his questions and reactions, so as to provide auxiliary decision-making instead of dominating the situation. For example, when a doctor is doing ward rounds, the assistant presumes the doctor's initial diagnosis intention based on the doctor's examination sequence and expression, and searches the relevant cases and precautions in the knowledge base for the doctor's reference. The assistant analyzes the patient's language and facial emotions (empathy module) at the same time, and if the patient is nervous because of the worry, it advises the doctor to explain or suspend certain stimulating topics in a more understandable way. This reflects the positive moderating effect of AI on the doctor-patient relationship. Nursing robots, on the other hand, use built-in ethics (such as the "Do No Harm Principle") to ensure the safety and comfort of patients during handling or treatment, such as detecting painful expressions and immediately slowing down and soothing them. Multiple AIs share patient status semantics and resource allocation purposes through the platform , for example, two diagnostic AIs can inform each other of different assumptions about the same patient and agree on them, or automatically negotiate which one will prioritize new emergency patients (following the ethical guidelines of critical priority).

Argument: We will test the solution in a simulated hospital environment and compare it with real doctor feedback. Expected results: Medical AI with the blessing of social intelligence can more accurately grasp the purpose of doctors, and the adoption rate of diagnostic suggestions will be increased; Emotional response to patients leads to improved patient satisfaction and compliance, and reduces medical disputes. Under the multi-AI collaboration, the efficiency and accuracy of diagnosis and treatment will be better than that of a single machine or no collaboration, and there will be no violation of medical ethics. This validation will demonstrate the potential of our technology in a serious industry (healthcare) and can be used as a reference paradigm for future medical AI system development.

Scenario 3: Intelligent education interactive system

Scenario description: An intelligent education platform that includes AI teacher assistants and several virtual student agents, as well as real student participation. AI teachers are responsible for explaining and tutoring, and student agents simulate different personalities and learning statuses to interact with real students. The whole system is like a hybrid virtual classroom.

Needs and challenges: In the field of education, which emphasizes individualized teaching and personality development, AI teachers need to understand students' learning purposes (e.g., asking questions for further study or seeking help due to confusion), integrate moral education (e.g., honesty and trustworthiness, cooperative spirit) into teaching strategies, and respond to students' emotional changes (frustration or boredom) in a timely manner and adjust the teaching pace. When multiple students (real + virtual) exist at the same time, AI teachers should pay fair attention to everyone, be impartial, and create a good class atmosphere. This places high demands on the social intelligence of AI.

This project plan: build a virtual AI teacher image and run the DIKWP social intelligence model internally. In the teaching process, the teacher's AI infers the status of each student in real time through the students' facial expressions and answering questions: Purpose, identify which students are eager to challenge more difficult and which need basic explanation; In terms of ethics, when students are found to be cheating or conflicting, AI teachers implement pre-set ethical intervention strategies (such as emphasizing the importance of integrity or advocating the spirit of mutual assistance); In terms of emotion, the AI teacher will slow down the tone, give encouragement, and may adjust the teaching plan to reduce the difficulty when detecting that individual students show frustrated expressions due to repeated mistakes, so as to prevent them from losing confidence. Virtual student agents have different parameters (diligent, naughty, etc.) to test the adaptability of AI teachers. AI teachers need to coordinate the interaction between real and virtual students, such as when a naughty virtual student interjects and causes trouble, the AI teacher politely but firmly stops them, and takes the opportunity to guide the class to discuss the topic of discipline or respect for others, so as to achieve a subtle and silent moral education.

Argument: By comparing the experimental class (with AI teachers) and the ordinary class, we will evaluate the learning effect and the change in students' moral character. It is expected that AI teachers will be able to significantly improve the degree of personalization of teaching, improve learning performance, and contribute to the aspect of invisible moral education (e.g., students are more willing to cooperate, and the cheating rate is reduced). In particular, we pay attention to whether AI teachers respond appropriately to unexpected situations (disputes, mistakes) in the classroom, and ensure that AI's moral judgments are in line with human teacher norms. If successful, the scenario will demonstrate the feasibility of brain-inspired social intelligence in the field of education, showing how AI teaching assistants can both teach and educate.

The above three scenarios cover the representative fields of human-machine and machine-machine interaction, and verify the role of this project technology in improving the social adaptability and credibility of AI. These application demonstrations also lay the foundation for subsequent industrialization.

Three-year research plan and phases

The implementation period of the project is planned to be three years (2025.7-2028.6), and it will be carried out in three stages: basic theoretical breakthrough, key technology research, and system integration demonstration, and each stage will focus on tasks and milestone goals as follows

Phase 1 (Year 1, 2025.7–2026.6): Theoretical framework and prototype building

• Task 1: Improve the design of the DIKWP semantic model. Combined with the needs of social intelligence, the formal definitions and interfaces of semantic elements at each layer of DIKWP are clarified, including the representation method of the Purpose layer and the inference framework of the Wisdom layer integrating moral constraints. Output: A description of the overall technical scheme and architecture design of the project.

• Task 2: Purpose, Ethics, and Empathy Algorithm Prototype Development. Preliminary algorithm models were developed respectively: the other people's purpose reasoning prototype based on rules + Bayes, the small moral rule set was constructed and the role of verification in simple decision-making problems was constructed, and the basic emotion recognition and simple empathy response modules were developed. The functional correctness of each module is verified separately in a closed test environment. Output: Prototype code of algorithm modules, respective performance reports.

• Task 3: Integrate the simulation platform with a single agent. Develop a basic framework for a multi-agent visualization simulation platform to support environment construction and simple interaction. A scenario (e.g., a social robot) is selected for single-agent integration testing: the three modules of task 2 are integrated into a DIKWP agent, and a human agent is interacted in a simulated environment to verify that the framework is basically feasible. Output: Simulation platform V1.0 version, integrated test report.

• Milestone 1 (June 2026): Completion of the preliminary architecture design of DIKWP+ social intelligence; The accuracy rate of Purpose recognition reaches ≥60% (in simple scenarios), the correct application rate of moral rule reasoning ≥ 70%, and the accuracy rate of emotion recognition ≥70%. The simulation platform can run demonstration scenarios where a single agent interacts with a human. Passed the mid-term review of experts in the first year.

Phase 2 (Year 2, 2026.7–2027.6): Key Algorithm Optimization and Multi-agent Collaboration

• Task 4: Algorithm Deepening and Learning Enhancement. In view of the shortcomings of the prototype in the first year, a data-driven method was introduced to optimize the performance of each module: adversarial training and inverse reinforcement learning were used to improve the generalization of purpose reasoning, and human feedback reinforcement learning (RLHF) was introduced to optimize moral decision-making, so that the empathy response parameters were automatically adjusted by simulating user feedback. Expand the training data (including simulated interaction logs and public datasets) to improve the robustness of each algorithm. Output: Upgraded algorithm module, performance improvement report.

• Task 5: Development of multi-agent collaboration mechanism. The extended simulation platform supports multi-agent parallel interaction and communication. Implement inter-agent semantic communication protocols and collaboration strategies, including purpose negotiation, task assignment, and conflict resolution. Ensure that the agent has some adversarial capability in a non-cooperative environment (e.g., in the event that an individual agent does not comply with the protocol). Output: multi-agent interaction module code, collaborative mechanism test report.

• Task 6: System integration and verification of medium-complexity scenarios. The optimized Purpose, Morality, and Empathy modules are integrated into each agent to realize the social intelligence system of N agents. Two scenarios (e.g., social robotics and education) were selected for mesoscale verification: for example, 3 robots interacted with 3 simulated elderly people to complete the task; Or an AI teacher interacting with several students. Evaluate the stability and performance of the system in a multiplayer environment. Output: System integration version V2.0, scenario test data and analysis report.

• Task 7: Self-explanatory and safety mechanism improvement. Develop decision-making process records and natural language interpretation generation components, collect interpretation cases in scenario validation, ask domain experts to evaluate the accuracy and comprehensibility of explanations, and iteratively improve the algorithm. At the same time, the security sandbox is improved to enforce the correction strategy for detected unethical or abnormal behaviors. Output: Interpret module code, interpret quality assessment report, security mechanism test report.

• Milestone 2 (June 2027): The multi-agent DIKWP system is fully functional and exhibits significant social intelligence characteristics in typical scenarios: 80% of the Purpose recognition ≥accuracy, 75% of the ethical decision-making agreement rate with human experts≥≥ and 85% of the empathy response rate (according to the user questionnaire score). The success rate of multi-agent cooperative tasks is ≥30% higher than that of the baseline without social intelligence. Explanation systems are able to generate textual explanations of key decisions, which correctly reflect the internal logic ≥ 80% of the time. Passed the second annual milestone review to lay the foundation for the final demonstration.

Phase 3 (Year 3, 2027.7–2028.6): Comprehensive Demonstration and Test Evaluation

• Task 8: Demonstration of full-scale application scenarios. Deploy our system for comprehensive demonstration testing in all three selected scenarios. Including: multi-person and multi-machine collaborative experiments in the real environment of nursing homes (test beds can be built with cooperative units); Continuous scenario testing that simulates hospital processes; Teaching experiments that combine virtual classrooms with real students. Collect quantitative metrics (task completion rate, response delay, user satisfaction, etc.) and qualitative feedback to comprehensively evaluate system performance.

• Task 9: Performance Optimization and Engineering Improvement. In response to the problems found in the demonstration, the system is optimized from an engineering perspective: such as improving the efficiency of the algorithm to meet real-time performance (consider lightweight model deployment or using ACPU hardware to accelerate prototype testing), improve the human-machine interface friendliness, and enhance robustness to avoid errors under extreme conditions. Output: Optimized final system V3.0, performance optimization report.

• Task 10: Verification and acceptance preparation of assessment indicators. According to the requirements of the guidelines, test and verify each item against the assessment indicators and form a report. Organize project technical documents and manuals, and write white papers for industrial applications. Preparation of project acceptance presentations, including on-site presentations and video materials. Outputs: Project summary report, technical white paper, acceptance presentation materials.

• Milestone 3 (June 2028): All assessment indicators of the project meet or exceed the requirements: for example, in the real-world scenario test, the system decision-making accuracy rate ≥ 90%, the multi-agent collaboration efficiency is improved by more than 50%, and there are no serious ethical violations; The average score of the user satisfaction survey ≥ 85; The accuracy rate of interpretation was ≥ 90% by third-party evaluation; The algorithm runs to meet the quasi-real-time requirements (decision delay < 0.5 seconds). Form a mature technology package that can be used for the next step of industrialization. Pass the final acceptance of the project.

Key technical nodes and assessment indicators

In order to ensure that the project is carried out according to the phased goals, the project has set up a number of key technical nodes and corresponding quantitative assessment indicators:

• Node 1: DIKWP Semantic Framework Design Completed (Month 6) – Indicators: Submission of framework design documents and prototype systems; The completeness of the expert argumentation framework has reached the recognized level.

• Node 2: The performance of the prototypes of the three algorithm modules meets the standard (Month 12) – Indicators: PurposeThe inference accuracy rate ≥ 60%, the correct application rate of moral rules ≥70%, and the accuracy rate of emotion recognition ≥70%; Each module functions properly in the single-agent test.

• Node 3: Verification of multi-agent collaboration mechanism (Month 18) – Indicators: The success rate of 3-5 agent collaboration tasks in the simulation scenario increased by ≥30% (compared with no synergy); The resolution rate of conflicts ≥ 80 per cent.

• Node 4: Mid-term system integration performance (Month 24) – Indicators: In the comprehensive scenario (e.g., Scenario 1), the agent exhibits expected social intelligence: 80% of the purpose recognition ≥, 75% of the moral judgment is consistent with humans, and the appropriateness of empathy response ≥≥ 85%; The system runs stably for 1 hour without crashing.

• Node 5: Interpretation and Safety Function Effectiveness (Month 24) – Indicators: 10 system decisions were randomly selected, and 8 were correctly explained by the explanation module≥ Bad behavior induction is introduced, and the successful interception rate of the security mechanism is 100%.

• Node 6: Final Demonstration Indicator (Month 33) – Indicators: The comprehensive test of different scenarios meets the requirements of the guideline: including a task success rate of ≥ 90%, a user satisfaction rate of ≥ of 85%, a multi-agent efficiency improvement of ≥50%, and no ethical violations; The system decision-making delay < 500ms to meet the requirements of real-time applications.

• Node 7: Outcome Output Indicators (Month 36) – Indicators: ≥5 high-level papers published (at least 1 CCF-A journal/conference); applied for 3 invention patents ≥; Train X number of doctoral/master's students≥; Build demonstration systems and datasets that are open and shared with scientific research and industry.

The design of the above indicators fully considers the technical achievement and application effect, including not only algorithm performance indicators (accuracy, improvement rate, etc.), but also user experience and security compliance indicators to ensure that the scientific value and practical benefits of the project results are taken into account.

Expected results of the project and the path of industrialization

Expected Results: This project will achieve innovative results with international leading level in both theory and application. In terms of theory, the brain-inspired social intelligence computing framework is proposed, which enriches the research field of artificial intelligence cognitive computing and multi-agent. It is expected to produce a series of academic papers and monograph chapters to enhance China's discourse power in the direction of artificial consciousness and social AI. In terms of technology, a prototype system platform and algorithm library have been developed, including modules such as Purpose Recognition, Moral Decision-making, and Empathy Computing, which can be widely reused in different AI systems. In terms of application, the verification and demonstration of three typical scenarios were completed to form a generalizable solution. The project will also produce datasets and benchmarks (e.g., a multi-agent social reasoning test set) that will contribute to the establishment of industry standards. In terms of intellectual property rights, it is expected to apply for a number of patents (algorithms, devices, etc.) and software copyrights to lay the foundation for subsequent industrial transformation.

Industrialization path: The project results have a clear industrial application prospect, and we will promote its transformation and implementation through the following ways:

• Establish an open platform and ecosystem: At the end of the project, the "Brain-like Social Intelligence Development Framework" was released, including some open source code, interface specifications, and demonstration cases, to lower the threshold for industry adoption. Relying on the DIKWP laboratory and international standards work led by Professor Yucong Duan, we will promote the inclusion of project results in AI evaluation standards and tool libraries. Hold seminars and training courses to attract developers and enterprises to participate, and gradually form an ecosystem.

• Demonstration application incubation: Cooperate with enterprises/institutions with a foundation for cooperation to carry out pilot applications in the fields of elderly care robots, medical AI auxiliary diagnosis, and intelligent teaching. Through small-scale deployment, the system is verified in the real environment and user feedback is collected to continuously polish the product. Success stories will serve as a benchmark to promote market awareness and confidence. For example, cooperate with elderly care equipment companies to create a prototype of an emotional escort robot; Cooperate with hospital information vendors to integrate AI ethical decision-making modules; Jointly launched a trial version of the intelligent teaching assistant system with an educational technology company.

• Technical cooperation and licensing: For the core algorithm modules of the project (such as the Purpose inference engine, moral evaluation module, etc.), we seek to cooperate with large AI companies and embed them into existing product systems to improve performance. For example, our Purpose recognition technology is licensed to driverless or intelligent customer service vendors to improve decision-making collaboration in multi-agent scenarios. Provide the empathy interaction module to the social robot company to strengthen the human-machine affinity of the product. Gain industry benefits and expand your influence through technology licensing and services.

• Entrepreneurship and industrial base: Depending on the maturity of the project results and market opportunities, the team does not rule out the establishment of a start-up company or the establishment of a brain-like social intelligence industrial base with industry alliances. Make full use of the Hainan Free Trade Port policy and the special support to transform scientific research results on the spot and form a new growth point for artificial intelligence. Focus on promoting the industrialization of products in the field of people's livelihood such as medical care and elderly care, and strive to give birth to a number of innovative products with demonstration effect within 3-5 years to serve the national strategic needs such as Healthy China and Wisdom Education.

• Continuous R&D and standardization: After the project is completed, we will continue to track the cutting-edge technology and continuously optimize the algorithm model. For example, the introduction of newer large models to enhance language empathy capabilities, or the experimentation of brain-like chip implementations to improve energy efficiency. At the same time, it actively participates in the formulation of AI ethics and multi-agent system standards at home and abroad, elevates project experience to norms, and enhances China's dominant position in relevant standards. It is expected that the results of the project will have a far-reaching impact on AI security, swarm intelligence and other fields, and have sustainable vitality.

In summary, this project will produce a set of available, credible and scalable brain-like social intelligence technologies and systems, which will provide key support for the development of multi-agent artificial intelligence. In scientific research, it fills the gap in the social cognition of AI, and promotes the better integration of AI into human society in its application. With solid theoretical innovation and clear application direction, we are confident that we will complete the task on time, achieve the expected indicators, and push the results to practical application, creating significant social and economic benefits.

References: The key references in this Recommendation are well documented, and some of the literature and sources are as follows (excerpts):

• Yucong Duan et al., "DIKWP Artificial Consciousness Model Research Report", 2023

• K. Wu, Y. Duan et al., "Computer Architecture and Chip Design for DIKWP Artificial Consciousness", 2023

• Liang He, Science and Technology Daily, "China's First General Artificial Intelligence Large-scale Social Simulator Released", 2025

• Elizaveta Tennant et al., "Hybrid Approaches for Moral Value Alignment in AI Agents", 2024

• Synced, “Stanford’s Hypothetical Minds: Multi-Agent AI with Theory of Mind”, 2024.

• J. Siwek et al., "Artificial Empathy in Multi-Agent Systems", 2023

• Du Jun et al., "A Study on the Relationship between Theory of Mind and Moral Sensitivity", BMC Psychology, 2024