Multi-agent communication method and system based on communication intention-guided parallel diffusion

Abstract

The present invention belongs to the field of artificial intelligence and multi-agent collaboration technology, and provides a multi-agent communication method and system based on communication intention-guided parallel diffusion. Its technical solution is to receive the original sequence to be communicated and the communication intention sent by the source agent, and convert the communication intention into a structured conditional prompt; generate an initial noise communication sequence based on the original sequence to be communicated; input the initial noise communication sequence into a large-scale language diffusion model, and use the conditional prompt as conditional information to guide the large-scale language diffusion model to perform multi-step diffusion iterative denoising, ultimately generating a denoised communication sequence; format the denoised communication sequence according to the communication protocol between the source agent and the target agent, and send the formatted sequence to be communicated to the target agent. This method achieves rapid and high-quality generation and transmission of communication information between agents, thereby significantly improving the overall efficiency and performance of multi-agent collaboration.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and multi-agent collaboration technology, and in particular relates to a multi-agent communication method and system based on communication intention-guided parallel diffusion. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, breakthroughs have been made in artificial intelligence technology, particularly large language models (LLMs), such as the GPT series, which have demonstrated powerful capabilities in natural language processing and content generation. In this context, multi-agent systems (MAS), a distributed AI paradigm that mimics the collaborative behavior of multiple agents in human society, have been widely applied to solving complex problems, such as engineering development, web design, and complex scientific research. MAS can decompose complex tasks into multiple subtasks, each handled by a different agent, thus enabling seamless solutions to problems that are difficult to address with a single large model.

[0004] Current mainstream multi-agent systems still face significant challenges in inter-agent communication. Regardless of the collaborative approach, the work of downstream agents is entirely dependent on the processing results of upstream agents. Traditional communication mechanisms typically rely on large language models using autoregressive generation, generating communication content sequentially, word by word and token by token. While this generation method excels in ensuring language fluency and logical coherence, its serial nature requires that each token generated by the model depends on previously generated tokens. This prevents full parallelization of the generation process, resulting in linear growth in generation time and message length, making it difficult to achieve millisecond-level interactions. During the autoregressive generation process, if a token is generated incorrectly, subsequent generations may be affected, resulting in a decline in the quality of the entire communication content. Furthermore, as the scale of the multi-agent system expands and the number of agents increases, the communication link load increases dramatically, and bandwidth and computing costs rise, making it difficult to meet the requirements of large-scale multi-agent collaborative communication. Summary of the Invention

[0005] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a multi-agent communication method and system based on a diffusion model, which integrates communication intentions into the diffusion process and parallelly guides the large language diffusion model to realize the rapid and high-quality generation and transmission of communication information between agents, thereby significantly improving the overall efficiency and performance of multi-agent collaboration.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A first aspect of the present invention provides a multi-agent communication method based on communication intention-guided parallel diffusion, comprising the following steps:

[0008] Receive the original communication sequence and communication intention sent by the source agent, and convert the communication intention into a structured conditional prompt;

[0009] generating an initial noise communication sequence based on the original sequence to be communicated;

[0010] The initial noisy communication sequence is input into a large-scale language diffusion model. The conditional prompt is used as conditional information to guide the large-scale language diffusion model to perform multi-step diffusion iterative denoising. In each iteration, the context information of each word in the sequence is simultaneously considered based on a parallel structure. The context information and conditional prompt of the sequence after the noise removal in the previous iteration are used as input. The denoised word at each position in the sequence is predicted in parallel, and the sequence is corrected based on the denoised word at each position, ultimately generating a denoised communication sequence.

[0011] The denoised communication sequence is formatted according to the communication protocol between the source agent and the target agent, and the formatted sequence to be communicated is sent to the target agent.

[0012] Furthermore, structured conditional cues include the core semantics of the communication content as well as additional constraints, including the length, style, keywords, and target agent identity of the communication.

[0013] Furthermore, the formula for generating the initial noise communication sequence based on the original sequence to be communicated is:

[0014] ,

[0015] in, is the initial noise communication sequence, is a masked sequence, is the original sequence to be communicated, is the mask symbol, L is the number of word tokens in the information sequence to be communicated, is the Dirac function, indicating that the 𝑖th position of the sequence is not masked and the original content is maintained. Indicates that the 𝑖th position of the sequence is masked and replaced by the mask symbol; t is the mask intensity ratio parameter.

[0016] Furthermore, when considering the contextual information of each word in the sequence based on the parallel structure, a bidirectional Transformer structure is adopted, and the cross entropy loss of its training objective is:

[0017] ,

[0018] in, is the cross entropy loss, , , To represent the parameters 𝑡, the original sequence And the corresponding mask sequence Find the mathematical expectation, where the masking ratio 𝑡 obeys the uniform distribution on the interval [0,1], and the original sequence From real communication Sampling from the data distribution, Predict the probability of the true content of the first 𝑖 position for the model.

[0019] Furthermore, the contextual information and conditional prompts of the sequence after iterative denoising in the previous step are used as input. When predicting the denoised word at each position in the sequence in parallel, the word probability distribution of the predicted denoised word is estimated. Combined with the word probability distribution estimation results, the final denoised word at each position is screened out.

[0020] Furthermore, when the final denoised word-unit at each position is screened and combined with the word-unit probability distribution estimation results, the estimation of the word-unit probability distribution is selected through sampling or Argmax operation.

[0021] Furthermore, when formatting the denoised communication sequence according to the communication protocol between the source agent and the target agent, the denoised communication sequence includes adding a message header, a sender / receiver ID, a timestamp, and message type metadata to the denoised communication sequence.

[0022] Furthermore, the method also includes sharing the state information generated by the agent during the task execution with related agents through the LLaDa model, including:

[0023] Filter out key status information important to the collaborative task through preset rules to form an initial status information set;

[0024] The initial state information set is input into the LLaDa diffusion model, and the state information text is generated with the guidance of task-related condition prompts;

[0025] The agent broadcasts or transmits the generated status information text to other related agents through a preset communication channel;

[0026] After receiving the above status information, the relevant agents parse the text content and extract valid status information, and update the collaborative knowledge base or local status view maintained by themselves to achieve synchronization of the internal status of the multi-agent system.

[0027] A second aspect of the present invention provides a multi-agent communication system based on a language diffusion model, comprising:

[0028] The communication task receiving module is used to receive the original communication sequence and communication intention sent by the source agent, and convert the communication intention into a structured conditional prompt;

[0029] The communication content generation module is used to generate an initial noisy communication sequence based on the original sequence to be communicated. The initial noisy communication sequence is input into a large language diffusion model, and conditional prompts are used as conditional information to guide the large language diffusion model to perform multi-step diffusion iterative denoising. In each iteration, the context information of each word in the sequence is simultaneously considered based on a parallel structure. The context information and conditional prompts of the sequence after the noise is removed in the previous iteration are used as input. The denoised word at each position in the sequence is predicted in parallel, and the sequence is corrected according to the denoised word at each position, ultimately generating a denoised communication sequence.

[0030] The communication content transmission module is used to format the denoised communication sequence according to the communication protocol between the source agent and the target agent, and send the formatted sequence to be communicated to the target agent.

[0031] A third aspect of the present invention provides a program product.

[0032] A program product, which is a computer program product, includes a computer program. When the computer program is executed by a processor, it implements the steps of the multi-agent communication method based on communication intention-guided parallel diffusion as described in the first aspect.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention maps the communication tasks between multiple agents into the parallel denoising process of the Large Language Diffusion Model (LLaDa), introduces communication intention as a conditional prompt, and dynamically injects context at each diffusion stage to achieve precise control of communication content. It solves the real-time, high-throughput, and robust communication problems of large-scale access systems, and improves the efficient and reliable collaborative communication between multiple agents.

[0035] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1This is a flow chart of a multi-agent communication method based on communication intention-guided parallel diffusion provided by an embodiment of the present invention;

[0038] Figure 2 This is a block diagram of a multi-agent communication system based on communication intention-guided parallel diffusion provided by an embodiment of the present invention;

[0039] Figure 3 It is a block diagram of an electronic device according to multiple embodiments provided by the embodiments of the present invention. DETAILED DESCRIPTION

[0040] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0041] In the description of the embodiments of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to." The term "based on" should be understood as "based at least in part on." The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment." The terms "first," "second," etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0042] Current mainstream multi-agent systems still face significant challenges in inter-agent communication. Regardless of the collaborative approach, the work of downstream agents is completely dependent on the processing results of upstream agents. Traditional communication mechanisms typically rely on large language models using autoregressive generation, and communication content is generated word by word and token by token. For example, when agent A needs to send a message to agent B, the LLM will start with the first word of the message and predict and generate each subsequent word in sequence until the message ends. Although this generation method performs well in ensuring language fluency and logical coherence, its inherent serial nature leads to the following core problems:

[0043] 1. Inefficient Generation: Autoregressive generation requires the model to rely on previously generated tokens for each token generated, making the generation process inefficiently parallelizable. Generation time increases linearly with the length of the communication content, severely impacting the real-time nature of communication, especially in scenarios requiring rapid response.

[0044] 2. High communication latency: Due to generation efficiency limitations, there is a significant delay between agents sending and receiving messages. In dynamic, real-time multi-agent collaborative tasks, such as autonomous vehicle fleet coordination or robot formation control, high latency can lead to outdated information, poor decision-making, and even system crashes.

[0045] 3. Limited scalability: As the scale of multi-agent systems increases, the number of agents increases, and the frequency and volume of communication increase dramatically, the efficiency bottlenecks of traditional autoregressive communication become more pronounced. The overall communication bandwidth and processing power of the system are unable to meet the needs of large-scale collaboration, limiting the application of MAS in complex, large-scale systems.

[0046] 4. Insufficient robustness and fault tolerance: During the autoregressive generation process, if a word is generated incorrectly, subsequent generation may be affected, resulting in a drop in the quality of the entire communication content. This can reduce the reliability of communication in noisy environments or adversarial scenarios.

[0047] In addition, the inventors found that text is a discrete sequence of symbols, which is different from continuous image pixels. Directly applying the denoising mechanism of the continuous diffusion model becomes complicated. At the same time, how to effectively integrate the communication intention (Prompt) into the diffusion process and guide the model to generate communication content that meets specific semantics and task requirements is also a key issue.

[0048] To address at least one of the aforementioned issues, an improved method for multi-agent communication is proposed according to an embodiment of the present disclosure. This method innovatively combines the parallel generation mechanism of a diffusion model with the semantic understanding and generation capabilities of a large language model. This method leverages the ability of the large language diffusion model (LLaDa) to process all tokens in parallel during the denoising process, breaking the serial generation limitations of traditional autoregressive models. By using communication task requirements as conditional prompts to guide the diffusion process, the resulting communication content efficiently and accurately meets the collaborative needs of the agents, enabling rapid, high-quality generation and transmission of communication information between agents, thereby significantly improving the overall efficiency and performance of multi-agent collaboration.

[0049] An architecture in which various embodiments of the present disclosure can be implemented includes intelligent agent nodes, a communication network, and a diffusion model training and management platform;

[0050] Agent Node: Each agent node is an independent entity with the capabilities of perception, decision-making, action, and communication. For example, in autonomous driving scenarios, each autonomous vehicle can be an agent node; in smart manufacturing, each robot or production unit can be an agent node.

[0051] Communication Network: Connects all agent nodes and transmits information between agents after it has been processed by a large-scale language diffusion model. This network can be a local area network, wide area network, cellular network (such as 5G / 6G), satellite network, or a combination thereof. The communication network should have low latency and high bandwidth to support efficient collaboration among large-scale agents.

[0052] Diffusion Model Training and Management Platform: An independent, centralized or distributed platform that can generate and update diffusion models based on different application scenarios and data modalities.

[0053] Please refer to the figure, Figure 1 A multi-agent communication method based on a large-scale language diffusion model is shown, comprising the following steps:

[0054] S101, receiving the original communication sequence and communication intention sent by the source agent, and converting the communication intention into a structured conditional prompt;

[0055] In this embodiment, the original communication initiated by the source agent can be multimodal information. For example, text, images, audio, video, sensor data, or a combination thereof. In this embodiment, the original communication sequence is text data. Word embedding can be generated through the Transformer model. After preprocessing, all information is converted into a unified, fixed-dimensional communication sequence. .

[0056] It should be noted that the communication intention sent by the source agent is an unstructured "information request" or "task goal"; this information needs to be abstracted into a structured "information request" or "task goal" and then sent to one or more target agents;

[0057] For example, in a task decomposition scenario, the information requirement might be "decompose task X into subtasks Y and Z, and assign them to agents B and C." In a decision negotiation scenario, the information requirement might be "request agents B and C to evaluate decision plan A and provide their opinions." These explicit information requirements will be converted into conditional prompts in natural language.

[0058] The prompt not only contains the core semantics of the communication content, but can also include additional constraints such as the length, style, keywords, target agent identity, etc. The quality of the prompt directly affects the accuracy and relevance of the communication content generated by the LLaDa model.

[0059] S102, generating an initial noise communication sequence based on the original sequence to be communicated;

[0060] In this example, when generating the initial noise communication sequence based on the original sequence to be communicated, the sequence can be composed of random tokens or special placeholders representing "unknown" or "noise". The length of this original sequence to be communicated matches the maximum length of the communication content to be generated;

[0061] The initial noise communication sequence is the starting point of the diffusion process. The LLaDa model will gradually transform this meaningless noise sequence into meaningful communication information through iterative denoising.

[0062] Specifically, the generation formula of the initial noise communication sequence is:

[0063] ,

[0064] in, is the initial noise communication sequence, is a masked sequence, is the original sequence to be communicated, is the mask symbol, L is the number of word tokens in the information sequence to be communicated, is the Dirac function, indicating that the 𝑖th position of the sequence is not masked and the original content is maintained. Indicates that the 𝑖th position of the sequence is masked and replaced by the mask symbol; t is the mask (or noise) intensity ratio parameter; the range is [0,1];

[0065] Through this initialization method, the original sequence to be communicated is completely fuzzified.

[0066] S103: Input the initial noisy communication sequence into the large language diffusion model LLaDa, use the conditional prompt as conditional information, and guide the large language diffusion model to perform multi-step diffusion iterative denoising. In each iteration, based on a parallel structure, the context information of each word in the sequence is simultaneously considered. The context information and conditional prompt of the sequence after the noise removal in the previous iteration are used as input. The denoised word at each position in the sequence is predicted in parallel, and the sequence is corrected according to the denoised word at each position, ultimately generating a denoised communication sequence.

[0067] In this invention, the initial noisy communication sequence is converted into understandable, high-quality communication information in parallel through the iterative denoising process of the LLaDa model. This process takes advantage of the parallelism of the diffusion model and the semantic understanding ability of the LLaDa model.

[0068] The specific steps include:

[0069] S301, setting the number of iteration steps;

[0070] Communication content generation is a multi-step iterative process with a finite number of steps (e.g., 5 to 20 steps, based on experience). In each iteration, the LLaDa model receives the current state of the communication sequence (initially a noisy sequence, followed by a progressively denoised sequence) and a Conditional Prompt as input. The model's goal is to predict and remove the "noise" in the current sequence, bringing it closer to the final, clear communication content.

[0071] By setting a limited number of iterations, rapid convergence of communication information is achieved. Compared to image diffusion models, which require hundreds or even thousands of steps, text diffusion models, due to their discrete nature, can typically achieve high-quality generation within a relatively small number of iterations. For example, 5-20 iterations are sufficient to significantly reduce communication generation time while maintaining generation quality. This rapid convergence is key to achieving efficient intelligent agent communication.

[0072] S302, predicting the context information of each word in the sequence in parallel based on the bidirectional Transformer structure;

[0073] In this embodiment, the LLaDa model uses a bidirectional Transformer structure. , the training target is cross entropy loss:

[0074] ,

[0075] in, is the cross entropy loss, , , To represent the parameters 𝑡, the original sequence And the corresponding mask sequence Find the mathematical expectation, where the masking ratio 𝑡 obeys the uniform distribution on the interval [0,1], the original sequence From real communication Sampling from the data distribution, masking the sequence According to the aforementioned forward diffusion process, from the given and The mask ratio 𝑡 is generated, Predict the probability of the true content at position 𝑖 for the model (Transformer). The model attempts to predict and restore the original sequence based on the masked sequence.

[0076] Unlike traditional autoregressive Transformers (such as GPT), which can only process sequences in one direction (from left to right), bidirectional Transformers can simultaneously consider the contextual information of each word in the sequence, regardless of its position in the sequence;

[0077] Therefore, at each denoising step, the model can predict and correct tokens at all positions in the sequence in parallel, enabling the parallel generation of communication information. This parallelism greatly improves generation efficiency and avoids the serial bottleneck of autoregressive models.

[0078] S303, predicting the denoised word at each position in the sequence based on the context information and conditional prompts of each word in the sequence;

[0079] In each iteration, the LLaDa model predicts the denoised word at each position in the sequence based on the contextual information and conditional prompts of the current communication sequence. The denoised word is the word that the model believes is closest to the actual communication content at that position.

[0080] In this embodiment, when denoising a word at each position in the sequence, the probability distribution of the predicted denoised word is estimated, and the denoised word is selected based on the estimation result;

[0081] It should be noted that the estimation of word unit probability distribution of the predicted denoised word units depends on the specific data format, and those skilled in the art can select the corresponding word unit probability distribution according to the specific data;

[0082] In this embodiment, the denoised word-unit at each position is selected based on an estimation of the word-unit probability distribution through sampling or Argmax operation.

[0083] For example, if the Argmax operation is selected, when selecting based on the result of the estimation of the word unit probability distribution by the Argmax operation, the final denoised word unit at each position is selected based on the result corresponding to 1 after the Argmax operation is amplified.

[0084] S304, outputting the denoised sequence to be communicated;

[0085] After the LLaDa model completes the pre-set iterative denoising process, the resulting communication sequence is the understandable and executable communication content between agents, which is efficiently delivered to the target agent.

[0086] This invention effectively integrates communication intention (Prompt) into the diffusion process, guiding the model to generate communication content that meets specific semantics and task requirements. It overcomes the efficiency and real-time bottlenecks of traditional autoregressive communication and fully utilizes the advantages of parallel generation and high-quality output of the diffusion model to meet the development needs of future large-scale, high-real-time multi-agent systems.

[0087] S104, formatting the output denoised sequence to be communicated according to the communication protocol between the multiple agents to obtain a denoised sequence to be communicated after formatting, and sending the formatted sequence to be communicated to the target agent;

[0088] In this embodiment, in order to ensure that the target agent can correctly parse and process the sequence to be communicated, the output denoised sequence to be communicated is formatted according to the communication protocol between multiple agents;

[0089] Specifically, when the denoised communication sequence is formatted according to the communication protocol between the source agent and the target agent, metadata such as the message header, sender / receiver ID, timestamp, and message type are added to the denoised communication sequence to ensure that the message can be correctly parsed and processed by the target agent.

[0090] The formatted communication content is sent directly to the intended target agent. Due to the efficiency of the generation process, the time delay from communication intent input to final content output is extremely low, enabling rapid interaction and information exchange between agents. This fast interaction capability is crucial for multi-agent applications that require high-frequency communication and real-time decision-making.

[0091] S105. When new tasks appear or the environment changes, the large language diffusion model LLaDa is used to generate and broadcast task decomposition and allocation instructions, and adjustments are made according to the process of S102-S104 to achieve task allocation and dynamic adjustment;

[0092] Efficient communication enables more refined and dynamic task allocation. When new tasks emerge or the environment changes, the task coordination agent can quickly generate and broadcast task decomposition and allocation instructions through LLaDa, which other agents can quickly receive and respond to. Furthermore, problems encountered during task execution or status updates can be promptly reported, facilitating dynamic adjustments to task allocation.

[0093] Furthermore, the state information generated by the agent during task execution (such as its own progress, environmental perception, resource occupancy, etc.) can be efficiently generated and shared with related agents through the LLaDa model.

[0094] Specifically, the state information sharing process includes the following steps:

[0095] Acquiring and organizing state information: The agent continuously monitors internal state parameters such as task progress, resource usage, and environmental status, triggering shared events periodically or when significant state changes occur. The agent first organizes its own state data and, using pre-set rules, selects key state information important to the collaborative task, forming an initial state information set.

[0096] If instructions are required, the preset rules can be set according to the actual situation.

[0097] Diffusion model parallel information generation: The agent inputs the above initial state information set into the LLaDa diffusion model, using task-related conditional prompts to guide the generation model. LLaDa completes the text output of the overall state description in a short time through a parallel non-autoregressive generation mechanism. During the iteration process, it automatically corrects inconsistencies or ambiguities in the state description to ensure the consistency and clarity of the generated information.

[0098] Status information sharing and dissemination: The agent broadcasts or transmits the generated status information text to other relevant agents through a preset communication channel. The specific transmission path can be determined based on the topological structure or task relationship between the agents to ensure that the relevant agents can quickly receive the latest status messages.

[0099] Receiving and updating state information: After receiving the state information, the relevant agents first parse the text content and extract valid state information. They then update their collaborative knowledge base or local state view to synchronize the internal state of the multi-agent system. If the parsing process discovers inconsistencies or conflicts in the information from different agents, the agents can further trigger a conflict identification mechanism.

[0100] Furthermore, in scenarios requiring group decision-making, multiple agents can generate their own opinions or evaluation results in parallel and quickly share them through the LLaDa mechanism. The specific implementation process of this parallel communication and decision-making mechanism includes:

[0101] Parallel opinion generation: Each agent independently generates its own opinion text in parallel using the LLaDa model based on shared environment and task information. This process leverages the non-autoregressive generation mechanism of the diffusion model to ensure that each agent efficiently generates opinions simultaneously, avoiding the delays associated with traditional one-by-one communication.

[0102] Rapid Opinion Sharing: After all agents generate their opinions almost simultaneously, they quickly broadcast or send their textual opinions to all decision-related agents, enabling rapid and parallel exchange of opinions. After receiving the opinions of other agents, each agent immediately aggregates all of them.

[0103] Conflict identification and analysis: The aggregated opinions are analyzed by the agent's built-in or external semantic analysis module. The specific analysis includes:

[0104] Semantic contradiction detection: By semantically matching the opinion text, we can identify direct conflicts or inconsistencies between opinions, such as diametrically opposed action recommendations.

[0105] Confidence inconsistency judgment: Extract the intrinsic confidence generated by the model or the explicit confidence assigned by the agent in the opinion text, judge the situation where the confidence deviates significantly, and mark it as a potential conflict.

[0106] Rapid conflict resolution and consensus formation: Once a conflict is discovered, the agent will execute a rapid negotiation and resolution process based on the LLaDa model. Specific methods include:

[0107] Iterative regeneration negotiation: Conflicting opinions are re-input into the LLaDa model, and new opinions that eliminate the conflict are regenerated through additional conditional prompts (such as "comprehensive solution" and "compromise suggestion");

[0108] Confidence voting mechanism: When conflicts are difficult to resolve completely, agents can make quick voting decisions based on the confidence of their opinions;

[0109] Priority decision-making: Based on the predefined agent priority, the opinions of high-priority agents are directly adopted;

[0110] Conversational negotiation: Leveraging the LLaDa model’s multi-round parallel generation capabilities, intelligent agents conduct fast and efficient multi-round negotiation dialogues to gradually reach consensus.

[0111] Ultimately, through this conflict identification and resolution mechanism, multiple agents can quickly reach a collective decision-making consensus and jointly adopt a consistent action strategy. The aforementioned technical details demonstrate the technical implementation of agent communication and collaborative decision-making based on the LLaDa diffusion model, ensuring the rapid and efficient operation of multi-agent systems.

[0112] Through this mechanism, the present invention significantly improves the overall task execution efficiency of multi-agent systems. Fast, high-quality communication reduces information transmission latency and increases agent response speed, enabling the entire system to adapt more quickly to environmental changes and complete complex tasks more efficiently.

[0113] Figure 2 FIG. 2 shows a schematic block diagram of a multi-agent communication system 200 based on a language diffusion model according to an embodiment of the present disclosure. Figure 2 As shown, the communication system 200 includes a communication task receiving module 201, a communication content generating module 202 and a communication content transmitting module 203;

[0114] The communication task receiving module 201 is used to receive the original communication sequence and communication intention sent by the source agent, and convert the communication intention into a structured conditional prompt;

[0115] The communication content generation module 202 is configured to generate an initial noisy communication sequence based on the original sequence to be communicated; the initial noisy communication sequence is input into a large language diffusion model, and conditional prompts are used as conditional information to guide the large language diffusion model to perform multi-step diffusion iterative denoising. In each iteration, the context information of each word in the sequence is simultaneously considered based on a parallel structure, and the context information and conditional prompts of the sequence after the noise is removed in the previous iteration are used as input. The denoised word at each position in the sequence is predicted in parallel, and the sequence is corrected based on the denoised word at each position, ultimately generating a denoised communication sequence;

[0116] The communication content transmission module 203 is used to format the denoised communication sequence according to the communication protocol between the source agent and the target agent, and send the formatted sequence to be communicated to the target agent.

[0117] This paper maps the communication tasks between multiple agents into the parallel denoising process of the Large Language Diffusion Model (LLaDa), introduces communication intention as a conditional prompt, and dynamically injects context into each diffusion stage to achieve precise control of communication content. It solves the real-time, high-throughput, and robust communication problems of large-scale access systems and improves the efficient and reliable collaborative communication between multiple agents.

[0118] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0119] Figure 3 A schematic block diagram of an example electronic device 300 that can be used to implement an embodiment of the present disclosure is shown. Figure 1 A multi-agent communication method based on communication intention-guided parallel diffusion in

[15] . Electronic devices are intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0120] like Figure 3As shown, the device 300 includes a computing unit 301, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. Various programs and data required for the operation of the device 300 can also be stored in the RAM 303. The computing unit 301, the ROM 302, and the RAM 303 are connected to each other via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0121] Various components in device 300 are connected to I / O interface 305, including: an input unit 306, such as a keyboard, mouse, etc.; an output unit 307, such as various types of displays, speakers, etc.; a storage unit 308, such as a magnetic disk, optical disk, etc.; and a communication unit 309, such as a network card, modem, wireless communication transceiver, etc. The communication unit 309 allows device 300 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0122] The computing unit 301 can be various general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 301 performs the various methods and processes described above. For example, in some embodiments, a multi-agent communication method based on communication intention-guided parallel diffusion can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 300 via the ROM 302 and / or the communication unit 309. When the computer program is loaded into the RAM 303 and executed by the computing unit 301, one or more steps of the multi-agent communication method based on communication intention-guided parallel diffusion described above can be performed. Alternatively, in other embodiments, the computing unit 301 can be implemented in any other appropriate manner.

[0123] It is configured (for example, by means of firmware) to execute a multi-agent communication method based on communication intention-guided parallel diffusion.

[0124] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0125] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0126] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0127] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0128] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0129] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0130] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not a limitation herein.

[0131] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A multi-agent communication method based on communication intention-guided parallel diffusion, characterized by: include: Receive the original communication sequence and communication intention sent by the source agent, and convert the communication intention into a structured conditional prompt; generating an initial noise communication sequence based on the original sequence to be communicated; The initial noisy communication sequence is input into a large-scale language diffusion model. The conditional prompt is used as conditional information to guide the large-scale language diffusion model to perform multi-step diffusion iterative denoising. In each iteration, the context information of each word in the sequence is simultaneously considered based on a parallel structure. The context information and conditional prompt of the sequence after the noise removal in the previous iteration are used as input. The denoised word at each position in the sequence is predicted in parallel, and the sequence is corrected based on the denoised word at each position, ultimately generating a denoised communication sequence. The denoised communication sequence is formatted according to the communication protocol between the source agent and the target agent, and the formatted sequence to be communicated is sent to the target agent.

2. The multi-agent communication method based on communication intention-guided parallel diffusion as claimed in claim 1, characterized in that: Structured conditional cues include the core semantics of the communication content as well as additional constraints, including the length, style, keywords, and identity of the target agent.

3. The multi-agent communication method based on communication intention guided parallel diffusion as claimed in claim 1, characterized in that: The formula for generating the initial noise communication sequence based on the original sequence to be communicated is: ， in, is the initial noise communication sequence, is a masked sequence, is the original sequence to be communicated, is the mask symbol, L is the number of word tokens in the information sequence to be communicated, is the Dirac function, indicating that the 𝑖th position of the sequence is not masked and the original content is maintained. Indicates that the 𝑖th position of the sequence is masked and replaced by the mask symbol; t is the mask intensity ratio parameter.

4. The multi-agent communication method based on communication intention guided parallel diffusion as claimed in claim 1, characterized in that: When considering the contextual information of each word in the sequence based on a parallel structure, a bidirectional Transformer structure is adopted, and the cross entropy loss of its training objective is: ， in, is the cross entropy loss, , , To represent the parameters 𝑡, the original sequence And the corresponding mask sequence Find the mathematical expectation, where the masking ratio 𝑡 obeys the uniform distribution on the interval [0,1], and the original sequence From real communication Sampling from the data distribution, Predict the probability of the true content at the 𝑖th position for the model.

5. The multi-agent communication method based on communication intention-guided parallel diffusion as claimed in claim 1, characterized in that: The contextual information and conditional prompts of the sequence after iterative denoising in the previous step are used as input. When predicting the denoised word at each position in the sequence in parallel, the word probability distribution of the predicted denoised word is estimated. Combined with the word probability distribution estimation results, the final denoised word at each position is screened.

6. The multi-agent communication method based on communication intention-guided parallel diffusion as claimed in claim 5, characterized in that: Combined with the word unit probability distribution estimation results, when screening the final denoised word unit at each position, the estimation of the word unit probability distribution is selected through sampling or Argmax operation.

7. The multi-agent communication method based on communication intention guided parallel diffusion as claimed in claim 1, characterized in that: When formatting the denoised communication sequence according to the communication protocol between the source agent and the target agent, the method includes adding a message header, a sender / receiver ID, a timestamp, and a message type metadata to the denoised communication sequence.

8. The multi-agent communication method based on communication intention guided parallel diffusion as claimed in claim 1, characterized in that: The method further includes sharing state information generated by the agent during task execution with related agents through the LLaDa model, including: Filter out key status information important to the collaborative task through preset rules to form an initial status information set; The initial state information set is input into the LLaDa diffusion model, and the state information text is generated with the guidance of task-related condition prompts; The agent broadcasts or transmits the generated status information text to other related agents through a preset communication channel; After receiving the above status information, the relevant agents parse the text content and extract valid status information, and update the collaborative knowledge base or local status view maintained by themselves to achieve synchronization of the internal status of the multi-agent system.

9. A multi-agent communication system based on a language diffusion model, characterized in that: include: The communication task receiving module is used to receive the original communication sequence and communication intention sent by the source agent, and convert the communication intention into a structured conditional prompt; A communication content generation module, which is used to generate an initial noise communication sequence based on an original sequence to be communicated; The initial noisy communication sequence is input into a large-scale language diffusion model. The conditional prompt is used as conditional information to guide the large-scale language diffusion model to perform multi-step diffusion iterative denoising. In each iteration, the context information of each word in the sequence is simultaneously considered based on a parallel structure. The context information and conditional prompt of the sequence after the noise removal in the previous iteration are used as input. The denoised word at each position in the sequence is predicted in parallel, and the sequence is corrected based on the denoised word at each position, ultimately generating a denoised communication sequence. The communication content transmission module is used to format the denoised communication sequence according to the communication protocol between the source agent and the target agent, and send the formatted sequence to be communicated to the target agent.

10. A program product, wherein the program product is a computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the multi-agent communication method based on communication intention-guided parallel diffusion as described in any one of claims 1 to 8 are implemented.

Images