Response generation with a multi-agent framework

Abstract

The present disclosure provides methods, apparatuses and non-transitory computer readable media for generating a response using a multi-agent framework. The multi-agent framework can include a planning agent, an evaluation agent and a discussion group. A topic can be obtained. In response to the topic, a plurality of rounds of discussions can be performed. In each round of discussion, the following operations can be performed: determining, using the planning agent, a discussion strategy for the round of discussion; determining, using the planning agent, a set of role-playing agents for forming the discussion group according to the discussion strategy; performing, using the discussion group, the round of discussion according to the discussion strategy; and generating, using the evaluation agent, an evaluation result for the round of discussion, the evaluation result being used to determine whether to perform a next round of discussion. A response to the topic can be generated according to outputs of the discussion group in the plurality of rounds of discussions.

Description

Background Technology

[0001] In the field of Artificial Intelligence (AI), agent technology has received widespread attention. An agent is an intelligent entity driven by an AI model, possessing capabilities such as perception, decision-making, and action, enabling it to autonomously complete given tasks without requiring human intervention at every step. Currently, agent technology is developing rapidly and showing broad application prospects in areas such as virtual assistants and robot control. Summary of the Invention

[0002] This invention is provided to introduce a set of concepts, which will be further described in the following detailed description. This invention is not intended to identify key or essential features of the protected subject matter, nor is it intended to limit the scope of the protected subject matter.

[0003] Embodiments of this disclosure provide a method, apparatus, and non-transitory computer-readable medium for generating responses using a multi-agent framework. The multi-agent framework may include a planning agent, an evaluation agent, and a discussion group. A topic can be obtained. In response to the topic, multiple rounds of discussion can be performed. In each round of discussion, the following operations can be performed: using the planning agent, determining a discussion strategy for the round; using the planning agent, determining a set of role-playing agents for forming the discussion group based on the discussion strategy; using the discussion group, performing the round of discussion according to the discussion strategy; and using the evaluation agent, generating an evaluation result for the round of discussion, the evaluation result being used to determine whether to proceed to the next round of discussion. Based on the output of the discussion group in the multiple rounds of discussion, a response to the topic can be generated.

[0004] It should be noted that one or more of the above aspects include the features specifically pointed out in the following detailed description and claims. Certain illustrative features of the one or more aspects are set forth in detail in the following specification and drawings. These features merely indicate various ways in which the principles of each aspect can be implemented, and this disclosure is intended to include all such aspects and their equivalents. Attached Figure Description

[0005] The following description will take into account several aspects disclosed, which are provided to illustrate rather than limit the aspects disclosed.

[0006] Figure 1 An exemplary process for generating responses using a multi-agent framework according to an embodiment is shown.

[0007] Figure 2 A schematic diagram of the structure of a multi-agent framework according to an embodiment is shown.

[0008] Figure 3 A schematic diagram of the structure of agents in a multi-agent framework according to an embodiment is shown.

[0009] Figure 4 A schematic diagram illustrating the training of a role-playing agent using multimodal / multi-source training knowledge data according to an embodiment is shown.

[0010] Figure 5 An exemplary process for performing the first round of discussion according to an embodiment is shown.

[0011] Figure 6 An exemplary process is shown in which subsequent discussions are performed iteratively after the first round of discussions, according to an embodiment.

[0012] Figure 7 An exemplary process for evaluating a discussion according to an embodiment is shown.

[0013] Figure 8 A schematic diagram of a discussion launch interface for an example application of the response generation mechanism utilizing a multi-agent framework for implementing the present disclosure, according to an embodiment, is shown.

[0014] Figure 9 A schematic diagram of a response output interface for an example application implementing the response generation mechanism of the present disclosure using a multi-agent framework, according to an embodiment, is shown.

[0015] Figure 10 A flowchart of an exemplary method for generating responses using a multi-agent framework, according to an embodiment, is shown.

[0016] Figure 11 An exemplary apparatus for generating responses using a multi-agent framework, according to an embodiment, is shown.

[0017] Figure 12 An exemplary apparatus for generating responses using a multi-agent framework, according to an embodiment, is shown. Detailed Implementation

[0018] This disclosure will now be discussed with reference to various exemplary embodiments. It should be understood that this discussion of embodiments is merely intended to enable those skilled in the art to better understand and thus implement the embodiments of this disclosure, and is not intended to teach any limitation on the scope of this disclosure.

[0019] Large Language Model (LLM)-based agent technology has been applied to general search, research, and other fields. In one example application scenario, a user can provide a query to the agent, which can automatically retrieve relevant knowledge from the knowledge base based on the query and generate a response (e.g., search results or research findings) to provide feedback to the user.

[0020] Existing agent technologies are facing significant challenges in real-world applications. On one hand, agents struggle to accurately handle queries involving novel questions. This is because agents typically cannot retrieve relevant existing knowledge from their knowledge base for such novel questions, thus failing to generate accurate responses. In this situation, agents are significantly affected by the AI ​​illusion effect. On the other hand, even for queries involving only obsolete questions, the responses generated by agents are often unsatisfactory. This is because users generally expect agents to generate responses that incorporate fresh knowledge for obsolete questions, rather than simply integrating existing knowledge like a search engine. These two examples illustrate that the agent's mechanism of simply integrating existing knowledge severely limits its response generation capabilities.

[0021] This disclosure proposes an improved response generation mechanism utilizing a multi-agent framework. The multi-agent framework of this disclosure is a structure formed by multiple agents, which may include a planning agent for planning multiple rounds of discussion, an evaluation agent for evaluating each round of discussion, and agents in a discussion group dynamically formed for each round of discussion. These agents can collaborate and interact with each other to automatically complete the objective task of generating responses to a topic.

[0022] This disclosed multi-agent framework can generate topic-specific responses by automatically organizing multiple rounds of discussion in response to a topic. In each round, a planning agent can determine the discussion strategy. In the first round, the planning agent can determine the discussion strategy based on the topic. In each subsequent round, the planning agent can determine an updated discussion strategy based on the evaluation results of the previous round. The evaluation results can be generated by an evaluation agent and used to assess the quality of each round. By evaluating each round and determining the updated discussion strategy based on the evaluation results of the previous round, the discussion strategy can be dynamically updated based on the situation of the previous round. Dynamically updating the discussion strategy can achieve the technical effect of gradually guiding the discussion in a direction conducive to the generation of new knowledge.

[0023] In each round of discussion, a discussion group can be formed to implement the discussion strategy. This discussion group can then execute the discussion in that round according to the strategy. The discussion group can include a set of role-playing agents. Each role-playing agent is trained using multimodal / multi-source training knowledge data associated with a specific "role," enabling it to generate an academic knowledge system and personalized output consistent with the corresponding "role" under the guidance of the discussion strategy in that round of discussion. The discussion group can be dynamically updated according to the determined discussion strategy. For example, for the current round of discussion, one or more role-playing agents that performed the previous round of discussion can be removed, and / or one or more other role-playing agents can be added, based on the discussion strategy. Dynamically updating the discussion group according to the determined discussion strategy helps to utilize role-playing agents more suitable for implementing the discussion strategy to execute the discussion, thereby achieving the technical effect of promoting the generation of new knowledge in multiple rounds of discussion.

[0024] In a single round of discussion, a group of role-playing agents can exchange their academic viewpoints, which helps generate new knowledge through critical thinking. Furthermore, each role-playing agent can absorb the academic viewpoints output by other agents to expand its own knowledge base. In other words, the role-playing agents in the discussion group can gradually evolve through multiple rounds of discussion. This helps improve the agents' capabilities, thereby further promoting the technical effect of generating new knowledge through multiple rounds of discussion.

[0025] Therefore, the response generation mechanism utilizing a multi-agent framework disclosed herein is no longer limited to simply integrating existing knowledge in a knowledge base. Instead, it promotes the generation of new knowledge during the discussion process by organizing multiple rounds of discussions, such as knowledge from fresh perspectives or knowledge with fresh viewpoints. By promoting the generation of new knowledge, the response generation mechanism utilizing a multi-agent framework disclosed herein can generate accurate responses to novel problems or responses containing new knowledge to old problems, thereby significantly improving the response generation capabilities of the agents.

[0026] The response generation mechanism utilizing a multi-agent framework of this disclosure will now be described exemplarily with reference to the accompanying drawings.

[0027] Figure 1 An exemplary process 100 for generating responses using a multi-agent framework according to an embodiment is shown.

[0028] exist Figure 1In the exemplary process 100 shown, user 110 can trigger the response generation mechanism of this disclosure utilizing a multi-agent framework by accessing application 120. In one example, application 120 can be implemented as a response generation application. In this example, user 110 can provide a topic 130 through the user interface of application 120. Application 120 can generate a response to topic 130 by implementing the response generation mechanism of this disclosure utilizing a multi-agent framework. In one example, the generated response can be fed back to user 110 through the user interface of application 120.

[0029] In response to topic 130, a multi-agent framework can be used to automatically organize multiple rounds of discussion 140 to generate a response to topic 130.

[0030] A multi-agent framework is a structure composed of multiple agents that interact and cooperate with each other. In one example, the agents in a multi-agent framework can be driven by an LLM (Language Model). An LLM is a large-scale language model with the ability to efficiently perform general language understanding and generation. This ability is acquired by learning a large number of parameters using a large amount of data during training. The LLM used to drive the agents in a multi-agent framework can employ any model architecture known in the art, such as the Transformer architecture.

[0031] A multi-agent framework can include a planning agent, an evaluation agent, and a discussion group dynamically formed for each round of discussion. The planning agent is an agent with task planning capabilities, used to plan, manage, and organize multiple rounds of discussion. The evaluation agent is an agent with evaluation capabilities, used to assess the quality of each round of discussion. The discussion group can include a set of role-playing agents, each generating output by playing a specific "role" in a corresponding round of discussion. In one example, the discussion group can also include a facilitator agent, used to guide the discussion of a set of role-playing agents in a corresponding round of discussion. In one example, the multi-agent framework can also include a summarizing agent. The summarizing agent is an agent with content summarizing capabilities, used to summarize the output of the discussion group in a corresponding round of discussion to generate a summary of that round of discussion. In one example, the summarizing agent can also be used to summarize the output of the discussion group in multiple rounds of discussion to generate responses to the topic. The following will combine... Figure 2 Further exemplary details related to the structure of the multi-agent framework are described below.

[0032] In one example, any agent in a multi-agent framework can be an agent system composed of multiple sub-agents. For instance, any agent in a multi-agent framework can include a master sub-agent, a memory organization sub-agent, and a tool organization sub-agent. The master sub-agent can have the ability to determine the agent's thought chain. The memory organization sub-agent can have the ability to organize long-term memory data and / or short-term memory data according to the thought chain. The tool organization sub-agent can have the ability to organize the invocation of external tools according to the thought chain. Each sub-agent can be specifically trained to possess the corresponding ability. Utilizing multiple sub-agents, each with corresponding abilities, to form any agent in a multi-agent framework can help improve the agent's ability to perform the corresponding task. The following will combine... Figure 3 Further exemplary details relating to the structure of agents in a multi-agent framework are described.

[0033] The aforementioned multi-agent framework can be used to conduct multi-round discussions.140 Figure 1 As shown, the operations performed in each round of discussion may include, for example, discussion strategy determination operation 141, role-playing agent determination operation 142, discussion operation 143, summarization operation 144, evaluation operation 145, etc.

[0034] A discussion strategy determination operation 141 can be performed to determine the discussion strategy for the current round of discussion. A discussion strategy is a plan for how to conduct the current round of discussion. The discussion strategy can broadly describe the topic-related content that should be discussed in the current round of discussion. The discussion strategy determination operation 141 can be performed using a planning agent within a multi-agent framework. If the current round of discussion is the first round in a multi-round discussion, the planning agent can determine the discussion strategy for the first round based on topic 130. If the current round of discussion is any round after the first round, the planning agent can determine an updated discussion strategy for the current round based on the evaluation results generated by performing the evaluation operation 145 in the previous round of discussion.

[0035] A role-playing agent determination operation 142 can be performed to determine a set of role-playing agents based on a discussion strategy. This set of role-playing agents is used to form a discussion group, which is used to perform the current round of discussion. A planning agent can be used to perform the role-playing agent determination operation 142. The planning agent can determine which role-playing agents are suitable for implementing the discussion strategy. For example, the planning agent can determine the role-playing agents associated with the content to be discussed indicated in the discussion strategy as a set of role-playing agents for performing the current round of discussion. In one example, the planning agent can select the aforementioned set of role-playing agents from a role-playing agent library 150. The role-playing agent library 150 can include multiple candidate role-playing agents. Each candidate role-playing agent can be trained using training knowledge data associated with a specific "role," such that once selected to perform a round of discussion, the candidate role-playing agent can produce output in that round of discussion by playing the specific "role." The following will combine... Figure 4 Further exemplary details related to training candidate role-playing agents are described.

[0036] In one example, although Figure 1 As not shown, the planning agent can also determine whether to include a moderator agent in the discussion group based on the discussion strategy. For example, if the planning agent determines, based on the discussion strategy, that adding a moderator agent to the discussion group is more conducive to the execution of the current round of discussion, then the planning agent can decide to add a moderator agent.

[0037] Discussion operation 143 can be executed. Discussion operation 143 refers to the operation used to execute the current round of discussion. In discussion operation 143, a discussion can be conducted according to a discussion strategy using a formed discussion group (which may include a group of role-playing agents and an optional facilitator agent).

[0038] A summarization operation 144 can be performed to generate a summary of the current round of discussion. Summarization operation 144 refers to the operation of extracting key and important information from the discussion group's output in the current round of discussion and forming a summary. Summarization operation 144 can be performed using a summarizing agent in a multi-agent framework. Summarization operation 144 is optional, and in one example, it may not be performed.

[0039] Evaluation operation 145 can be executed to generate an evaluation result for the current round of discussion. Evaluation operation 145 can be used to evaluate the quality of the current round of discussion. For example, the quality of the current round of discussion can be evaluated from dimensions such as completeness, hierarchy, freshness, and accuracy. Evaluation operation 144 can be executed using an evaluation agent in a multi-agent framework. The evaluation result can be fed back to the planning agent to determine whether to proceed to the next round of discussion. The planning agent can determine whether the quality of the current round of discussion meets the requirements based on the evaluation result according to predetermined rules. If the quality does not meet the requirements, the planning agent can determine that the next round of discussion needs to be continued. In the next round of discussion, operations 141 to 145 can be repeated. If the quality meets the requirements, the planning agent can determine that the next round of discussion does not need to be continued, and the multi-round discussion 140 can be stopped. The following will combine... Figures 5 to 7 Further exemplary details relating to the operations in multi-round discussion 140 are described.

[0040] The exemplary process 100 may also include a response generation operation 160. The response generation operation 160 is used to generate a topic-specific response based on the output of the discussion group during multiple rounds of discussion. The generated response can be used in various application scenarios.

[0041] In one example application scenario, the generated response can be output at point 171. In this scenario, the response can be output as knowledge data 172 generated for the topic. For example, the generated response can be returned to user 110 through the user interface of application 120. In one example, during a multi-round discussion, the output of the discussion group can be returned to user 110 in real time through the user interface of application 120. In this example, user 110 can pause the multi-round discussion and modify the output of the discussion group, thus achieving manual control over the multi-round discussion. For example, user 110 might find that the output of a role-playing agent in the discussion group is not highly relevant to the topic. In this case, user 110 can instruct the discussion to be paused and the output of that role-playing agent modified to include content closely related to the topic. Such manual control helps to avoid the multi-round discussion from deviating from the intended direction by leveraging human experience. The following section combines... Figures 8 to 9 Further exemplary details relating to this example are described below. In another example, a response can be generated after multiple rounds of discussion have concluded by summarizing the outputs of the discussion group during those rounds (e.g., using a summarizing agent in a multi-agent framework). For example, the response generated by the summarization described above could take the form of a research report, a paper, a mind map, etc.

[0042] In another example application scenario, the generated response can be stored at 181. For example, the response can be stored as a knowledge entry in knowledge base 182. This example application scenario can correspond to the process of preparing knowledge entries in knowledge base 182. In one example, this example application scenario can be used in conjunction with search engine technology. When a query is received in the search engine, knowledge entries matching the query can be retrieved in knowledge base 182. The matching knowledge entries can be returned as search results to the user who provided the query.

[0043] Figure 2 A schematic diagram of the structure of a multi-agent framework 200 according to an embodiment is shown. In one example, the multi-agent framework 200 may correspond to the combination described above. Figure 1 The described multi-agent framework is a multi-turn discussion framework for responding to topics. (e.g.) Figure 2 As shown, the multi-agent framework 200 may include a planning agent 210, a discussion group 220 dynamically formed for each round of discussion, a summarizing agent 230, and an evaluation agent 240. The functions of the corresponding agents in the multi-agent framework 200 and the interactions between these agents will be described in detail below.

[0044] Planning agent 210 is an agent with task planning capabilities, which can be used to plan, manage, and organize multi-round discussions in response to topics.

[0045] In one aspect, the planning agent 210 can be used to obtain topics from outside the multi-agent framework 200. In one example, the planning agent 210 can, for example, obtain topics from an application (e.g., in conjunction with the above) via an application programming interface (API). Figure 1 The discussion uses application 120) to obtain a topic. A topic can be open or closed. An open topic refers to a topic in which no clear conclusion has yet been reached within the relevant field. A closed topic refers to a topic in which a clear conclusion has been reached within the relevant field; this clear conclusion can be based on facts, common sense, axioms, etc. In this disclosure, for simplicity, the clear conclusion reached regarding a specific closed topic is referred to as the truth value of that specific closed topic.

[0046] On the other hand, the planning agent 210 can also be used to dynamically determine the discussion strategy for each round of discussion. The discussion strategy is a plan for how to execute the current round of discussion and may include a general description of the content to be discussed in the current round. In one example, the discussion strategy may include a description of one or more of the following aspects: the core concepts involved in the topic, possible connections between core concepts, viewpoints related to the corresponding core concepts (including, for example, supporting viewpoints, opposing viewpoints, neutral viewpoints, etc.), the application and development prospects of the corresponding core concepts, etc. If the current round of discussion is the first round, the planning agent 210 can determine the discussion strategy based on the obtained topic. If the current round of discussion is any round after the first round, the planning agent 210 can dynamically update the discussion strategy for the current round based on the situation of the previous round of discussion that occurred before the current round. In this case, the planning agent 210 can interact with the summarizing agent 230 to obtain a summary of the previous round of discussion. The planning agent 210 can also interact with the evaluation agent 240 to obtain an evaluation result of the previous round of discussion. The planning agent 210 can then determine an updated discussion strategy for the current round of discussion based on the summary and evaluation results of the previous round. Dynamically updating the discussion strategy for the current round of discussion based on the situation of the previous round helps to guide the discussion in the desired direction by adjusting the discussion strategy in a timely manner, thereby achieving the technical effect of promoting the generation of new knowledge during the discussion process.

[0047] On the other hand, the planning agent 210 can also be used to form a discussion group 220 for carrying out the discussion in each round of discussion, based on the determined discussion strategy.

[0048] In one example, planning agent 210 can determine which role-playing agents to use to form discussion group 220 based on discussion strategies. In another example, planning agent 210 can select a set of role-playing agents from a role-playing agent library containing multiple trained candidate role-playing agents, as described above. Figure 1The discussion is ongoing. In one example, candidate role-playing agents in the role-playing agent library can be divided into multiple sets according to knowledge domains. These knowledge domains may include, for example, philosophy, biology, history, computer science, etc. The planning agent 210 can first determine the knowledge domains involved in the current round of discussion based on the discussion strategy. Then, the planning agent 210 selects role-playing agents from the set of candidate role-playing agents in the role-playing agent library corresponding to the determined knowledge domains to form the discussion group. For example, the planning agent 210 can determine, based on the discussion strategy, that the current round of discussion involves the knowledge domains of biology and computer science, and select a set of role-playing agents from the sets of candidate role-playing agents corresponding to the knowledge domains of biology and computer science, respectively. Figure 2 The discussion group 220 formed for a round of discussion can include a set of role-playing agents 221-1 to 221-M, where M is an integer greater than 1.

[0049] In one example, the planning agent 210 can also determine whether a moderator agent should be included in the discussion group based on the discussion strategy. If the planning agent 210 determines, based on the discussion strategy, that adding a moderator agent to the discussion group is more conducive to the execution of the current round of discussion, then the planning agent 210 can determine to add a moderator agent to the discussion group. Figure 2 The discussion group 220 formed for a round of discussion may include a moderator agent 222. The moderator agent 222 and a group of role-playing agents 221-1 to 221-M together form the discussion group for a round of discussion.

[0050] A role-playing agent is an intelligent agent that produces output by playing a specific "role." This "role" can correspond to a real person or a virtual character. In one example, the "role" could correspond to a user (e.g., in conjunction with the above). Figure 1 (Described user 110). In other words, a role-playing agent can participate in multiple rounds of discussion as a representative of the user. Playing a specific "role" means that the output generated by the role-playing agent conforms to the academic knowledge system and personality of the corresponding "role". The role-playing agent can acquire the ability to play a specific "role" through a training process. During the training process, data associated with a specific "role" is used to train the role-playing agent corresponding to that "role".

[0051] Any one of a group of role-playing agents 221-1 to 221-M can interact with a host agent 222 to generate an academic knowledge system and personalized output consistent with its corresponding "role" under the guidance of the host agent 222. In one example, the group of role-playing agents 221-1 to 221-M can also interact with each other. In this example, a specific role-playing agent (e.g., role-playing agent 221-1) can receive academic viewpoints output by other role-playing agents (e.g., role-playing agents 221-2 to 221-M) in a round of discussion. This specific role-playing agent can memorize the academic viewpoints of other role-playing agents to expand its own academic knowledge system. This specific role-playing agent can generate output based on the expanded academic knowledge system in subsequent discussions. In other words, in this example, a specific role-playing agent can achieve self-evolution during the discussion process by absorbing the output of other role-playing agents.

[0052] The facilitator agent 222 is an agent with discussion guidance capabilities, which can be used to guide a group of role-playing agents 221-1 to 221-M in conducting a discussion according to a discussion strategy. The facilitator agent 222 can interact with the planning agent 210 to obtain the discussion strategy for the current round of discussion from the planning agent 210. The facilitator agent 222 can determine the flow and / or form of the current round of discussion based on the discussion strategy. In one example, the facilitator agent 222 can determine the flow of the current round of discussion based on the discussion strategy, which may involve stages such as self-statement, rebuttal of opposing viewpoints, and self-summary. In another example, the facilitator agent 222 can determine the form of the current round of discussion based on the discussion strategy, such as one-on-one debate, free discussion, or speaking in a designated order. The facilitator agent 222 can interact with each of the role-playing agents 221-1 to 221-M to guide them in executing the discussion according to the determined flow and / or form.

[0053] The summarizing agent 230 is an agent with content summarizing capabilities, which can be used to summarize the output of the discussion group 220 in a single round of discussion and / or summarize the output of the corresponding discussion group in multiple rounds of discussion.

[0054] In one aspect, the summarizing agent 230 can be used to summarize the outputs of agents in discussion group 220 during the current round of discussion to generate a summary of the current round of discussion. The summarizing agent 230 can interact with a group of role-playing agents 221-1 to 221-M and the facilitating agent 222 in discussion group 220 respectively to obtain the corresponding outputs of each agent during the current round of discussion. The summarizing agent 230 can then extract key, important information from the outputs of discussion group 220 during the current round of discussion and generate a summary of the current round of discussion based on the extracted information. The summarizing agent 230 can also interact with the planning agent 210 and the evaluation agent 240. The summarizing agent 230 can provide the generated summary of the current round of discussion to the planning agent 210 to determine an updated discussion strategy for the next round of discussion. The summarizing agent 230 can also provide the generated summary of the current round of discussion to the evaluation agent 240 to generate an evaluation result for the current round of discussion. By summarizing the outputs of agents in discussion group 220 in the current round of discussion using summarizing agent 230, key and important information can be provided to planning agent 210 and evaluation agent 240. This helps planning agent 210 to efficiently and accurately determine updated response strategies and helps evaluation agent 240 to efficiently and accurately generate evaluation results.

[0055] On the other hand, the summarizing agent 230 can also be used to summarize the output of the corresponding discussion group in multiple rounds of discussion to generate a response to the topic. The summarizing agent 230 can extract key and important information from the output of the corresponding discussion group in multiple rounds of discussion and generate a response based on the extracted information. The summarizing agent 230 can generate a response with a specific format, which may include, for example, research reports, papers, mind maps, etc. By using the summarizing agent 230 to summarize the output of the corresponding discussion group in multiple rounds of discussion, the generated response can include filtered key and important information. In one example, the summarizing agent 230 can also be used to output the generated response outside the multi-agent framework.

[0056] The evaluation agent 240 is an agent capable of evaluating discussions, which can be used to generate evaluation results for the current round of discussion. In one example, the evaluation agent 240 can interact with the summarizing agent 230 to obtain a summary of the current round of discussion from the summarizing agent 230. The evaluation agent 240 can then evaluate the quality of the current round of discussion based on the summary.

[0057] In one aspect, the evaluation agent 240 can assess the discussion process in the current round of discussion. The evaluation agent 240 can evaluate the discussion process from dimensions such as completeness, hierarchy, novelty, and accuracy. Completeness measures whether the discussion process is comprehensive and sufficient. In one example, the evaluation agent 240 can determine which aspects the discussion process should cover for a given topic. These aspects may include, for example, discussions of concepts, current applications, and development trends—aspects that could potentially foster new knowledge. Regarding completeness, the evaluation agent 240 can compare the outputs of the current round of discussion with these aspects to determine the coverage of the discussion process. Hierarchy measures whether the discussion process progresses logically and progressively. In one example, regarding hierarchy, the evaluation agent 240 can assess whether the discussion process progresses from basic concepts to cutting-edge concepts, from current applications to future applications, and from applications in the original domain to applications in extended domains. Novelty measures whether fresh perspectives / ideas emerge during the discussion process. In one example, the evaluation agent 240 can identify existing viewpoints / perspectives for a topic, which are viewpoints / perspectives already existing in the evaluation agent 240's own knowledge base or external data it can access. Regarding freshness, the evaluation agent 240 can compare viewpoints / perspectives emerging during the discussion with the aforementioned existing viewpoints / perspectives to determine similarity, and can identify viewpoints / perspectives emerging during the discussion with similarity below a predetermined threshold as fresh viewpoints / perspectives. Accuracy is used to measure the accuracy of the outputs generated during the discussion. In one example, regarding accuracy, the evaluation agent 240 can determine, based on its own knowledge base or external data it can access, whether the outputs generated during the discussion are clear, accurate, and free from "illusions."

[0058] For each of the dimensions mentioned above, evaluation results can be represented in various ways. In one example, quantitative methods can be used to represent evaluation results; for example, the evaluation results can be represented as grade levels or ratings such as "low," "medium," or "high." In another example, qualitative methods can be used to represent evaluation results; for example, the evaluation results can be represented as descriptive comments. In yet another example, the qualitative and quantitative methods mentioned above can also be used in combination; for example, the evaluation results can include both grades and descriptive comments.

[0059] On the other hand, the evaluation agent 240 can also evaluate the discussion results of the current round of discussion. Discussion results, also known as discussion conclusions, refer to the explicit conclusions reached by the discussion group through the execution of the discussion. Evaluating the discussion results can be applied to situations where the topic is not open-ended. In this case, the evaluation agent 240 can determine whether the current round of discussion produced a discussion result, and whether the produced discussion result conforms to the truth value of the non-open-ended topic.

[0060] The evaluation agent 240 can also interact with the planning agent 210. The evaluation agent 240 can provide the generated evaluation results to the planning agent 210 to determine whether to proceed to the next round of discussion. The planning agent 210 can determine whether to proceed to the next round of discussion based on predetermined rules, which may be set for one or more of the multiple dimensions (e.g., completeness, hierarchy, freshness, accuracy) used to evaluate the discussion process. In one example, the predetermined rule may stipulate that if the evaluation result of the current round of discussion for any dimension does not reach a predetermined threshold, then proceed to the next round of discussion. In another example, the predetermined rule may stipulate that if the final evaluation result of the current round of discussion for multiple dimensions does not reach a predetermined threshold, then proceed to the next round of discussion. The final evaluation result for multiple dimensions can be calculated by combining the evaluation results for each dimension. For example, a weighted approach can be used to calculate the final evaluation result of the current round of discussion for multiple dimensions by assigning appropriate weights to each dimension. In other examples, other rules can be flexibly set in addition to the example rules described above. If it is determined that the next round of discussion will be conducted, the generated evaluation results can be further used by the planning agent 210 to determine an updated discussion strategy for the next round of discussion. In one example, the planning agent 210 can determine an updated discussion strategy based on one or more dimensions that are not in compliance as indicated in the evaluation results.

[0061] Figure 2 The structure of the multi-agent framework 200 shown is only an example. In other examples, the multi-agent framework may be omitted. Figure 2 One or more intelligent agents are shown, or include, in addition to Figure 2Other agents besides the one shown. In one example, the multi-agent framework may not include a summarizing agent. In this example, the evaluating agent may not generate an evaluation result based on the summary of the current round of discussion, but rather directly based on the output of the agents in the discussion group during the current round of discussion. Alternatively, in this example, other agents or content summarizing tools outside the multi-agent framework with content summarizing capabilities can be used to summarize the output of the discussion group in multiple rounds of discussion to generate a response. In another example, the multi-agent framework may not include a facilitating agent. In this example, the discussion guidance function of the facilitating agent described above can be implemented by a planning agent. For example, the planning agent can directly interact with a group of role-playing agents to guide the group of role-playing agents in discussing according to a determined discussion strategy.

[0062] Figure 3 A schematic diagram 300 illustrates the structure of an agent 310 in a multi-agent framework according to an embodiment. In one example, agent 310 may correspond to the combination described above. Figure 1 and Figure 2 The discussed multi-agent framework can include any agent, such as a planning agent, a role-playing agent, a facilitating agent, a generalizing agent, or an evaluation agent. Agent 310 may include a controlling sub-agent 311, a memory-organizing sub-agent 312, and a tool-organizing sub-agent 313. In other words, agent 310 can be an agent system composed of three sub-agents.

[0063] The master sub-agent 311 can serve as the decision-making center for agent 310. The master sub-agent 311 is used to determine the chain of thought (COT) of agent 310 for the target task. The target task refers to the operation that agent 310 needs to autonomously perform, for example, in conjunction with the above... Figure 1 and Figure 2 This describes the operations performed by the respective agents in a multi-agent framework. COT (Cognitive Order) refers to the reasoning process determined by agent 310 for a target task, involving a series of logical steps. Determining the COT enables agent 310 to perform the target task in a manner that simulates human problem-solving, for example, not jumping directly from input to output, but through a series of intermediate reasoning steps. These intermediate reasoning steps collectively constitute agent 310's "thought chain."

[0064] The memory organization sub-agent 312 is used to organize memory data 320 according to a thought chain. Memory data 320 may include long-term memory data 321 and short-term memory data 322. Long-term memory data 321 refers to data that can be retained for a long time without changing or being lost over time. In one example, long-term memory data 321 may be stored in the form of a knowledge base, where a knowledge base is a structured data collection. Short-term memory data 322 refers to data that is cleared after a short period of retention, such as data used during the execution of a specific objective task and cleared after the objective task is completed. In one example, short-term memory data 322 may be stored in the form of temporary variables, cached data, etc. The memory organization sub-agent 312 can determine which long-term memory data 321 and / or short-term memory data 322 the master sub-agent 311 needs according to the thought chain and provide the required data to the master sub-agent 311. The memory organization sub-agent 312 can also determine which data to store as long-term memory data 321 or short-term memory data 322 according to the thought chain.

[0065] The memory organization sub-agent 312 can be used in conjunction with various memory technologies to achieve long-term or short-term memory of data. These memory technologies may include those for implementing local or remote storage (e.g., cloud storage). In one example, if the memory organization sub-agent 312 determines, based on a thought chain, that the master sub-agent 311 needs certain long-term memory data 321, the memory organization sub-agent 312 can read the long-term memory data 321 from a corresponding memory used to implement long-term memory (e.g., various types of non-volatile memory). The memory organization sub-agent 312 can then provide the read long-term memory data 321 to the master sub-agent 311. Similarly, if the memory organization sub-agent 312 determines, based on a thought chain, that the master sub-agent 311 needs certain short-term memory data 322, the memory organization sub-agent 312 can read the short-term memory data 322 from a corresponding memory used to implement short-term memory (e.g., various types of volatile memory). Then, the memory organization sub-agent 312 can provide the read short-term memory data 322 to the master control sub-agent 311. In another example, if the memory organization sub-agent 312 determines, based on the thought chain, to save certain data as long-term memory data 321, the memory organization sub-agent 312 can write this data into the corresponding memory used to implement long-term memory. Similarly, if the memory organization sub-agent 312 determines, based on the thought chain, to save certain data as short-term memory data 322, the memory organization sub-agent 312 can write this data into the corresponding memory used to implement short-term memory.

[0066] In one example, agent 310 may be a planning agent in a multi-agent framework. In this example, the long-term memory data 321 may include, for example, knowledge data from various domains, data describing the academic viewpoints of candidate role-playing agents in a role-playing agent library, etc. The short-term memory data 322 may include, for example, data used by the planning agent during multi-round discussions, such as topics and discussion strategies, summaries, and evaluation results in each round. The exemplary long-term memory data 321 and short-term memory data 322 help the planning agent perform operations such as the discussion strategy determination and role-playing agent selection described above.

[0067] In another example, agent 310 can be a role-playing agent in a multi-agent framework.

[0068] In this example, the aforementioned long-term memory data 321 may include at least one of, for example, identity data, academic viewpoint data, and cultural background data. Identity data refers to data describing the identity of the "role" played by the role-playing agent, and may include data describing social identities (e.g., occupation, membership, etc.), personal identities (e.g., personal habits, personality, interests, etc.), and other identities. Academic viewpoint data refers to data describing the academic viewpoints of the "role" played by the role-playing agent, and may include data describing knowledge domains, educational experiences, academic achievements, academic social relationships, etc. Cultural background data may include data describing the cultural background of the "role" played by the role-playing agent, and may include data describing language abilities, upbringing, literary works, etc. In one example, the aforementioned identity data, academic viewpoint data, and cultural background data for a specific "role" may be obtained based on the "role's" self-description. Self-description refers to data generated by the "role" itself describing the characteristics of the "role," which may include descriptive data provided in the form of, for example, the "role's" autobiography, self-introduction, personal academic achievements, etc. In another example, the aforementioned identity data, academic viewpoint data, and cultural background data for a specific "role" can be obtained based on external descriptions of that "role." External descriptions refer to data generated from the outside world that describes the characteristics of the "role," and may include descriptive data provided in the form of, for example, external comments or introductions of the role. Organizing the above exemplary data into long-term memory data 321 for a role-playing agent helps the agent generate an academic knowledge system and personalized output that conforms to the corresponding "role" during discussions.

[0069] In this example, the aforementioned short-term memory data 322 may include at least one of the following: role-playing interaction data, modifications to the output of the role-playing agent, evaluations of the output of the role-playing agent, and historical self-reflection data. Role-playing interaction data refers to the output generated by other role-playing agents during a discussion, such as academic viewpoints contained in the output of other role-playing agents. Modifications to the output of the role-playing agent refer to, for example, modifications made by a user to the output of the role-playing agent (e.g., in conjunction with the above). Figure 1 (Example: Users can pause the discussion and modify the output of the role-playing agent). Evaluation of the role-playing agent's output refers to the assessment of that output by other agents (e.g., other role-playing agents) in a multi-agent framework, or by users. Evaluations can indicate the degree to which other agents or users approve of the output. Historical self-reflection data refers to data that the role-playing agent acquires from past discussions to improve its performance in subsequent discussions. Role-playing agents can generate historical self-reflection data by analyzing the outputs of other role-playing agents, user modifications, and evaluations from other agents or users obtained during past discussions. In one example, the role-playing agent can form historical self-reflection data based on academic viewpoints output by other role-playing agents. In this example, the role-playing agent can use the academic viewpoints of other role-playing agents to expand its own academic knowledge. In another example, the role-playing agent can indicate in its historical self-reflection data, based on user modifications, that the output generated in subsequent discussions does not conflict with those modifications. In another example, a role-playing agent can use historical self-reflection data, based on evaluations from other agents or users, to instruct itself to avoid producing outputs similar to those that received unfavorable evaluations in subsequent discussions. The role-playing agent can thus evolve itself over multiple rounds of discussion using this historical self-reflection data.

[0070] Whether the exemplary data described above is used as long-term memory data 321 or short-term memory data 322 can be flexibly set as needed. For example, in one example, the aforementioned historical self-reflection data can be used as long-term memory data, thereby further enhancing the self-evolution capability of the role-playing agent. The exemplary long-term memory data 321 and short-term memory data 322 described above help the role-playing agent perform operations such as conducting discussions according to the discussion strategy described above.

[0071] In another example, agent 310 may be a facilitator agent in a multi-agent framework. In this example, the long-term memory data 321 may include, for example, knowledge data from various domains. The short-term memory data 322 may include, for example, data used by the facilitator agent during multi-round discussions, such as discussion strategies in each round, and the corresponding outputs of a group of role-playing agents. The exemplary long-term memory data 321 and short-term memory data 322 help the facilitator agent perform operations such as guiding a group of role-playing agents in a discussion based on the discussion strategies described above.

[0072] In another example, agent 310 may be a generalizing agent within a multi-agent framework. In this example, the aforementioned long-term memory data 321 may include, for example, knowledge data from various domains. The aforementioned short-term memory data 322 may include, for example, data used by the generalizing agent during multi-round discussions, such as the corresponding outputs of agents in the discussion group. The aforementioned exemplary long-term memory data 321 and short-term memory data 322 help the generalizing agent perform operations such as generalizing for a single round of discussion and / or generalizing for multiple rounds of discussion, as described above.

[0073] In another example, agent 310 may be an evaluation agent in a multi-agent framework. In this example, the long-term memory data 321 may include, for example, knowledge data from various domains. The short-term memory data 322 may include, for example, data used by the evaluation agent during multi-round discussions, such as a summary of a round of discussion, the corresponding outputs of a group of role-playing agents, etc. The exemplary long-term memory data 321 and short-term memory data 322 described above help the evaluation agent perform operations such as generating an evaluation result for a round of discussion as described above.

[0074] The tool organization sub-agent 313 is used to organize the invocation of external tools according to the thought chain. The tool organization sub-agent 313 can invoke tools in the tool library 330. The tool library 330 can be a collection of external tools that agent 310 can use. The tool library 330 can include any type of tool that can be invoked by the tool organization sub-agent 313, such as a retrieval-augmented generation (RAG) tool 331, an optical character recognition (OCR) tool 332, a speech conversion tool 333, etc. In one example, the tool organization sub-agent 313 can access agent 310's knowledge base or external data by invoking tools in the tool library 330.

[0075] Figure 4A schematic diagram 400 is shown illustrating the training of a role-playing agent using multimodal / multi-source training knowledge data according to an embodiment.

[0076] Figure 4 A role-playing intelligent agent library 410 is shown. In one example, the role-playing intelligent agent library 410 can correspond to the combination described above. Figure 1 The described role-playing agent library 150. The role-playing agent library 410 may include multiple (e.g., N, which are positive integers greater than M) candidate role-playing agents 411-1 to 411-N. Each candidate role-playing agent 411-1 to 411-N, once selected to perform a round of discussion, can perform the discussion in a manner that plays a specific "role." For example, the candidate role-playing agent can generate an academic knowledge system and personalized output consistent with that "role" during the discussion. Each candidate role-playing agent can be trained to acquire the ability to play a specific "role." Figure 4 The diagram illustrates how candidate role-playing agents 411-2 are trained. It should be understood that other candidate role-playing agents in the role-playing agent library 410 can be trained in a similar manner to candidate role-playing agent 411-2.

[0077] like Figure 4 As shown, candidate role-playing agents 411-2 can be trained by executing role-playing agent training operations 420 using training knowledge data 430.

[0078] The role-playing agent training operation 420 can be performed using any technique in the art for training agents. For example, reinforcement learning techniques can be used to perform the role-playing agent training operation 420.

[0079] Training knowledge data 430 is data associated with the "role" to be played by the candidate role-playing agent 411-2, and may include one or more of the following: academic data 431, personalized data 432, etc. Academic data 411 refers to data associated with the "role's" academic knowledge, and may include, for example, academic achievement data (e.g., papers, technical articles, etc.), academic influence data (e.g., paper impact factors, industry evaluations, etc.), and academic research data (e.g., research fields, methodologies, core viewpoints, etc.). Personalized data 412 refers to data associated with the "role's" personalities, and may include, for example, personal background (e.g., educational background, upbringing, professional experience, cultural background, etc.) and personal style (e.g., personality, language habits, behavioral habits, etc.). Training knowledge data 430 may also include any other data associated with the "role" besides the data described above. In one example, training knowledge data 430 may be obtained based on the "role's" self-description. Self-description refers to data generated by the "role" itself that describes the characteristics of the "role," which may include descriptive data provided in the form of, for example, the "role's" autobiography, self-introduction, personal academic achievements, etc. In another example, training knowledge data 430 may be obtained based on external descriptions of the "role." External descriptions refer to data generated from the outside world that describes the characteristics of the "role," which may include descriptive data provided in the form of, for example, external comments or introductions of the role.

[0080] In one example, the training knowledge data 430 can be multimodal. Modality, also known as format or style, specifies how data is organized and stored. For example, multimodal training knowledge data could include text training knowledge data, video training knowledge data, audio training knowledge data, etc.

[0081] In one example, the training knowledge data 430 can be from multiple sources. Multi-source training knowledge data can include public data, semi-public data, and private data. Public data refers to data that can be publicly obtained, such as internet data. Semi-public data refers to data that can be publicly obtained under certain restrictions, such as copyrighted data. Private data refers to data that cannot be publicly obtained, such as private knowledge data of a specific user or expert knowledge data.

[0082] Using multimodal / multi-source training knowledge data to train role-playing agents can help improve the agents' ability to accurately play the corresponding "roles".

[0083] In one example, the candidate role-playing agents in the role-playing agent library 410 may include public role-playing agents and private role-playing agents. Public role-playing agents may be those not limited to use by a specific user. In one example, a public role-playing agent may be trained using publicly available training knowledge data 430. Private role-playing agents may be those limited to use only by a specific user. In one example, a private role-playing agent may be trained using semi-public and / or private training knowledge data 430. For example, a private role-playing agent limited to use only by a specific user can be trained using semi-public and / or private training knowledge data 430 provided by that specific user. In one example, a private role-playing agent limited to use by a specific user can be converted into a public role-playing agent through user authorization. In this example, a specific user can grant authorization to, for example, another user, to allow that user to use their private role-playing agent. Conversely, if a specific user does not grant authorization, their private role-playing agent remains limited to use only by that specific user.

[0084] Figure 5 An exemplary process 500 is shown, which utilizes a multi-agent framework to perform a first round of discussion according to an embodiment.

[0085] At point 510, the topic can be obtained. This topic corresponds to the combination mentioned above. Figure 1 Topic 130 is described. For ease of understanding, the following description will use an exemplary open topic—"What is the relationship between memory in LLM and DNA in biology"—to illustrate the operation of exemplary process 500.

[0086] At point 520, the discussion strategy for the first round can be determined based on the topic. This operation corresponds to the combination of the above. Figure 1 The discussion strategy described in the first round of discussion determines operation 141, and can be determined by a planning agent in a multi-agent framework (e.g., combined with the above). Figure 2 The discussion is carried out by the planning agent 210. For example, for the example topic above, the discussion strategy can be summarized as follows: In the first round of discussion, the concepts of memory in LLM, DNA in biology, the possible connections between these concepts, the current applications and development prospects of memory technology in LLM, and the current applications and development prospects of DNA technology should be discussed.

[0087] At point 530, the agents in the discussion group can be determined based on the discussion strategy defined at point 530. This operation can be performed by the planning agent described above. In one example, at point 530, the discussion strategy can be used to determine which role-playing agents to form the discussion group. In this example, the operation at point 530 can correspond to the combination of the above... Figure 1 The described role-playing agent determines operation 142. For example, for the example above, the planning agent can determine to form a discussion group using the following four role-playing agents that are conducive to realizing the above discussion strategy: a first role-playing agent whose knowledge system involves research on deep learning technology; a second role-playing agent whose knowledge system involves research on the future development direction of neural networks; a third role-playing agent whose knowledge system involves research on the double helix structure of DNA; and a fourth role-playing agent whose knowledge system involves research on the principles of storage and replication of genetic information. In one example, at 530, it can also be determined whether to add a chairing agent to the discussion group based on the discussion strategy.

[0088] At point 540, the first round of discussion can be performed. This operation corresponds to the combination mentioned above. Figure 1 The discussion operation 143 in the first round of discussion is described, and can be formed by a discussion group in a multi-agent framework for the first round of discussion (e.g., combined with the above). Figure 2 The discussion group 220 described above is executed. In one example, a group of role-playing agents in the discussion group can discuss the content that needs to be discussed as indicated in the discussion strategy above, under the guidance of a host agent (if present).

[0089] At position 550, a summary of the first round of discussion can be generated. This operation corresponds to the combination mentioned above. Figure 1 The generalization operation 144 described in the first round of discussion can be performed by a generalizing agent in a multi-agent framework (e.g., combined with the above). Figure 2 The summarizing agent (230) performs the task. The summarizing agent can extract key and important information from the output of the discussion group in the first round of discussion to generate a summary. For example, the summary may include concepts involved in the first round of discussion, the corresponding viewpoints output by the role-playing agents, etc.

[0090] At point 560, the evaluation results of the first round of discussions can be generated based on the summary. This operation corresponds to the combination mentioned above. Figure 1 The evaluation operation 145 described in the first round of discussion can be performed by an evaluation agent in a multi-agent framework (e.g., combined with the above). Figure 2 The evaluation agent 240 performs the discussion. For example, for the example above, an exemplary evaluation result of the first round of discussion could include:

[0091] Completeness: "High" - The first round of discussions covered most aspects, including discussions on the concept of memory in LLM and the concept of DNA in biology, discussions on the current applications of memory technology and DNA technology, and discussions on the development trends of memory technology and DNA technology.

[0092] Level: "Low" - The first round of discussions initially explored the potential connection between memory technology and DNA technology, but did not delve into the development prospects that this connection might bring.

[0093] Novelty: "High" - The first round of discussions brought forth the novel perspective of "interdisciplinary integration of memory in LLM with DNA in biology".

[0094] Accuracy: "High" - There are no obvious errors in the output of each role-playing agent.

[0095] At point 570, it can be determined whether to proceed with the next round of discussion. This operation can be performed by the planning agent described above. (As mentioned above...) Figure 2 The planning agent can determine whether to proceed with the next round of discussion based on predefined rules, considering one or more dimensions such as completeness, hierarchy, freshness, and accuracy. For example, in the above scenario, the planning agent can determine, based on predefined rules, that the first round of discussion does not meet the requirements in the hierarchy dimension, and thus determine that the next round of discussion needs to be executed.

[0096] In one example, in response to determining to proceed with the next round of discussion at 570 ( Figure 5 If the branch "is" is selected, process 500 can proceed to 580. At 580, the next round of discussion can be performed.

[0097] In one example, in response to determining at 570 not to perform the next round of discussion ( Figure 5 If the branch "No" is selected, process 500 can proceed to 590. At 590, a response can be generated. This operation corresponds to the combination above. Figure 1 The described response generation operation 160, and in one example, can utilize a generalized agent in a multi-agent framework (e.g., Figure 2 The generalizing agent (240) performs this action. The generalizing agent can generate a response by summarizing the output of the discussion group. For example, the response generated through this generalization can take the form of a research report, a paper, or a mind map.

[0098] Figure 6 An exemplary process 600 is shown in which subsequent discussions are iteratively executed after a first round of discussion, according to an embodiment. For clarity, process 600 is described with reference to the current round of discussion. This current round of discussion can be any round of discussion following the first round.

[0099] At point 610, an updated discussion strategy for the current round of discussion can be determined, at least based on the evaluation results of the previous round. In one example, the planning agent can determine the updated discussion strategy based on one or more dimensions that are deemed unacceptable in the evaluation results. For instance, for the example above, the updated strategy determined in the second round of discussion could broadly describe the possible directions of connection between memory technology and DNA technology that should be further discussed in the second round, and the potential development prospects brought about by this connection. In another example, the planning agent can determine the updated discussion strategy based on both the evaluation results and the summary of the previous round of discussion. For example, the planning agent can identify content that still needs to be discussed from the summary and combine this content with the content that needs to be discussed based on the evaluation results to generate an updated discussion strategy.

[0100] At point 620, agents in the discussion group can be determined based on the updated discussion strategy. For example, in the second round of discussion, the planning agent can determine to add a fifth role-playing agent to the discussion group, in addition to the four role agents mentioned above, whose knowledge base involves research on interdisciplinary applications of deep learning technology. In one example, at point 620, it can also be determined whether the discussion group includes a facilitator agent based on the updated discussion strategy.

[0101] At position 630, the current round of discussion can be executed. In one example, a group of role-playing agents in the discussion group can discuss the content indicated in the updated discussion policy above, guided by a moderator agent (if present).

[0102] At point 640, a summary of the current round of discussion can be generated. This summary can be generated using a summarizing agent that extracts key and important information from the discussion group's output in the current round.

[0103] At point 650, an evaluation result for the current round of discussion can be generated based on the summary. For example, for the example above, an exemplary evaluation result for the second round of discussion could include:

[0104] Completeness: "High" - The second round of discussions covered most aspects, including the concept of the connection between memory technology and DNA technology, the current application of the connection between memory technology and DNA technology, and the development trend of the connection between memory technology and DNA technology.

[0105] Level: "High" - The second round of discussions delved into the potential development prospects arising from the connection between memory technology and DNA technology, proposing that the replication and repair mechanisms of DNA could be used to enhance the stability of long-term memory.

[0106] Novelty: "High" - The first round of discussions introduced a novel perspective: "using DNA replication and repair mechanisms to enhance the stability of long-term memory."

[0107] Accuracy: "High" - There are no obvious errors in the output of each role-playing agent.

[0108] At point 660, it can be determined whether to proceed with the next round of discussion. A planning agent can be used to determine this based on predefined rules for one or more dimensions such as completeness, hierarchy, freshness, and accuracy. For example, in the above scenario, the planning agent can determine, based on predefined rules, that the second round of discussion meets the requirements in all dimensions, and therefore determine that no further discussion is necessary.

[0109] In one example, in response to determining to proceed with the next round of discussion at 660 ( Figure 6 If the branch "is" is selected, process 600 can proceed to the next round of discussion. In the next round of discussion, operations 610 to 660 are executed iteratively.

[0110] In another example, in response to determining at 660 not to perform the next round of discussion ( Figure 6 If the branch "No" is selected, process 600 can proceed to 670. At 670, the multi-round discussion process can be terminated. In one example, after determining to terminate the multi-round discussion process, a response to the topic can be generated based on the discussion group's output during the multi-round discussion.

[0111] Figure 7 An exemplary process 700 for evaluating a discussion according to an embodiment is shown.

[0112] In one instance, the operation shown in process 700 can be performed by an evaluation agent in a multi-agent framework (e.g., combined with the above). Figure 2 The evaluation agent 240 discussed was executed.

[0113] The exemplary process 700 may include obtaining topic 710. Topic 710 may correspond to the combination above. Figure 1 The topic of discussion is 130. In one example, the evaluation agent can obtain topic 710 from the planning agent.

[0114] At point 720, we can determine whether topic 710 is open. (This is in conjunction with the above.) Figure 2 For discussion, whether topic 710 is an open topic can be determined based on whether a clear conclusion has been reached in the field involved.

[0115] In one example, in response to determining at 720 that topic 710 is not open, that is, topic 710 is a non-open topic, process 700 can proceed to 730.

[0116] At point 730, an evaluation can be performed on the results of the current round of discussion. The discussion results are the explicit conclusions reached by the discussion group through the current round of discussion. In one example, evaluating the discussion results could include comparing the results with the truth values ​​of non-open topics to determine the similarity between the two.

[0117] At point 740, it can be determined whether the discussion result matches the true value. In one example, if the similarity between the discussion result and the true value is higher than a predetermined threshold, the discussion result can be determined to match the true value. Otherwise, if the similarity between the discussion result and the true value is not higher than the predetermined threshold, the discussion result can be determined to not match the true value.

[0118] In one example, in response to determining that the discussion result is true at 740, process 700 can proceed to 750.

[0119] At point 750, the discussion process of the current round can be evaluated. In one example, the evaluation can be performed on dimensions such as completeness, hierarchy, freshness, and accuracy, as described above. Figure 2 Discussed.

[0120] In another example, in response to determining at 740 that the discussion result does not meet the truth value, process 700 can proceed to 760. In this example, the evaluation of the discussion process can be skipped to quickly move on to the next round of discussion.

[0121] In another example, in response to determining that the topic is open at 720, process 700 can proceed to 750 to evaluate the discussion process.

[0122] At point 760, an evaluation result can be generated. The evaluation result can be generated based on an evaluation of the discussion process and / or an evaluation of the discussion results. In one example, if an evaluation of the discussion process is performed but not of the discussion results, the generated evaluation result may include evaluation results for example dimensions such as completeness, hierarchy, freshness, and accuracy. In another example, if an evaluation of the discussion results is performed but not of the discussion process, the generated evaluation result may include an evaluation result indicating whether the discussion results conform to truth. In yet another example, if both an evaluation of the discussion process and an evaluation of the discussion results are performed, the generated evaluation result may include evaluation results for example dimensions such as completeness, hierarchy, freshness, and accuracy, as well as an evaluation result indicating whether the discussion results conform to truth.

[0123] The evaluation results generated by process 700 can be fed back to the planning agent by the evaluation agent to determine whether to continue the next round of discussion. If the planning agent determines to continue the next round of discussion, it can further determine an updated discussion strategy for the next round based on the evaluation results.

[0124] The following will combine Figures 8 to 9 This section describes an example application for implementing the response generation mechanism utilizing a multi-agent framework as described above. This example application can correspond to the combination described above. Figure 1 Application 120 is discussed.

[0125] Figure 8 A schematic diagram of a discussion launch interface 800 for implementing the response generation mechanism of the multi-agent framework of this disclosure, according to an embodiment, is shown. This example application can correspond to the above-described combination of... Figure 1 Application 120 is discussed.

[0126] The discussion launch interface 800 may include a topic input box 810. The user accessing the sample application (e.g., the user can correspond to the above combination) Figure 1 User 110 can enter a topic in the topic input box 810. The discussion startup interface 800 may include a public role-playing agent selection area 820. The public role-playing agent selection area 820 can display selectable public role-playing agents. The selectable public role-playing agents can correspond to the above-mentioned combination. Figure 4 The role-playing agents discussed in the library are not limited to those used by a specific user. The public role-playing agent selection area 820 may include a domain selection box 830 and an agent selection area 840. The domain selection box 830 can be used to select a knowledge domain. For example, Figure 8 The example illustrates the selection of a computer domain. Different domains can be switched by interacting with the domain selection box. The agent selection area 840 can be used to present a collection of candidate role-playing agents from a role-playing agent library corresponding to the knowledge domain selected in the domain selection box 830. Taking role-playing agent 841 as an example, it can present the name 851 of the "role" played by the agent, a brief description 852 of the "role," and a selector 853 for selecting the role-playing agent. The discussion startup interface 800 may also include a private role-playing agent selection area 860. The private role-playing agent selection area 860 can present private role-playing agents for selection. The private role-playing agents for selection can correspond to the combinations described above. Figure 4The role-playing agents discussed in the library are limited to those used by the user in question. Private role-playing agents can be presented in a customized manner for the user accessing the sample application. In another example, if another user accesses the sample application, other private role-playing agents for that other user can be presented in the private role-playing agent selection area 860. For each private role-playing agent, the name of the "role" played by that role-playing agent, a brief description of the "role," and a selector for selecting the role-playing agent can be presented in a manner similar to that of role-playing agent 850.

[0127] In one example, in response to a user entering a topic in the topic input box 810, the response generation mechanism of this disclosure utilizing a multi-agent framework can automatically select a role-playing agent by performing the role-playing agent determination operation described above. Figure 8 Three role-playing agents, 841, 842, and 843, are shown as an example of being automatically selected. In one example, the user can modify the automatically selected role-playing agents. For example, Figure 8 An example is shown where a user can additionally select a role-playing agent 861.

[0128] Users can then begin the discussion by interacting with the start button 870. In one example... Figure 8 The selected role-playing agents 841, 842, 843, and 861 shown are used to perform the first round of discussion in a multi-round discussion. In subsequent discussions, the response generation mechanism of this disclosure utilizing a multi-agent framework can determine the role-playing agents to perform the discussion based on an updated discussion strategy, as described above.

[0129] Figure 9 A schematic diagram of a response output interface 900 for an example application implementing the response generation mechanism of the present disclosure using a multi-agent framework, according to an embodiment, is shown.

[0130] like Figure 9 As exemplarily shown, the output of the discussion group can be returned to the user in real time via the response output interface 900 during a multi-round discussion. In one example, the user can manually control the discussion process. For example, the user may find that the output 910 of the role-playing agent playing "Role 1" is not very relevant to the topic. In this case, the user can pause the multi-round discussion process by interacting with the pause button 920 and modify the output 910 to include content closely related to the topic. In one example, this modification can be stored as short-term memory data of the role-playing agent playing "Role 1," as described above. Figure 3 The discussion continues. Users can then interact with the continue button 930 to continue the discussion in multiple rounds.

[0131] In one example, the user can also evaluate the output of the role-playing agent. For instance, the user can interact with indicator 940 to indicate approval of output 910, or with indicator 950 to indicate disapproval of output 910. In one example, this evaluation can also be stored as short-term memory data of the role-playing agent acting as "Role 1," as described above. Figure 3 Discussed.

[0132] Figure 10 A flowchart is shown of an exemplary method 1000 for generating a response using a multi-agent framework according to an embodiment. The multi-agent framework for implementing method 1000 may include a planning agent, an evaluation agent, and a discussion group.

[0133] At point 1010, you can obtain the topic.

[0134] At point 1020, multiple rounds of discussion can be initiated in response to a topic. Within each round of discussion, actions 1021 through 1024 can be performed.

[0135] At point 1021, the planning agent can be used to determine the discussion strategy for the first round of discussion.

[0136] At 1022, the planning agent can be used to determine a set of role-playing agents for forming the discussion group, based on the discussion strategy.

[0137] At point 1023, the discussion group can be used to perform the round of discussion according to the discussion strategy.

[0138] At position 1024, the evaluation agent can be used to generate an evaluation result for the first round of discussion, which is used to determine whether to proceed with the next round of discussion.

[0139] At point 1030, a response to the topic can be generated based on the output of the discussion group in the multiple rounds of discussion.

[0140] In one implementation, method 1000 may further include: determining whether to execute the next round of discussion based on an evaluation result; and in response to determining to execute the next round of discussion based on the evaluation result, wherein determining the discussion strategy for the previous round of discussion includes: determining an updated discussion strategy based at least on the evaluation result; or in response to determining not to execute the next round of discussion based on the evaluation result, ending the multi-round discussion.

[0141] In one implementation, the multi-agent framework may further include a generalizing agent. Method 1000 may further include: using the generalizing agent to generate a summary of the first round of discussion by summarizing the output of the discussion group in the first round of discussion; and generating an evaluation result for the first round of discussion includes: generating the evaluation result based on the summary.

[0142] In one implementation, determining the updated discussion strategy based at least on the evaluation results may include: determining the updated discussion strategy based on the evaluation results and the summary.

[0143] In one implementation, determining a set of role-playing agents for forming the discussion group according to the discussion strategy may include: selecting the set of role-playing agents from a library of role-playing agents that includes multiple candidate role-playing agents, according to the discussion strategy.

[0144] In one implementation, method 1000 may further include: determining whether the discussion group includes a moderator agent based on the discussion strategy, wherein the moderator agent is used to guide the group of role-playing agents to perform the first round of discussion based on the discussion strategy.

[0145] In one implementation, generating the evaluation result of the first round of discussion may include: determining whether the topic is an open topic or a closed topic; and in response to determining that the topic is an open topic, generating the evaluation result by evaluating the discussion process of the first round of discussion; or in response to determining that the topic is a closed topic, generating the evaluation result by evaluating the discussion process of the first round of discussion and / or evaluating the discussion result of the first round of discussion.

[0146] In one implementation, evaluating the discussion process of the first round of discussion may include evaluating the discussion process for at least one of completeness, hierarchy, freshness, and accuracy.

[0147] In one implementation, determining the discussion strategy for the first round of discussion may include: if the first round of discussion is the first round of discussion, determining the discussion strategy based on the topic.

[0148] In one implementation, each agent in the multi-agent framework may include: a master sub-agent for determining the agent's thought chain; a memory organization sub-agent for organizing long-term memory data and / or short-term memory data according to the thought chain; and a tool organization sub-agent for organizing the invocation of external tools according to the thought chain.

[0149] In one implementation, for one of the set of role-playing agents: the long-term memory data may include at least one of identity data, academic viewpoint data, and cultural background data, and the short-term memory data may include at least one of role interaction data, modifications to the agent's output, evaluations of the output, and historical self-reflection data.

[0150] In one implementation, each role-playing agent may be trained using training knowledge data, which may include at least one of academic data and personalized data, and may include multimodal data and / or multi-source data.

[0151] In one implementation, generating a response to the topic based on the output of the discussion group in the multiple rounds of discussion may include: using a generalizing agent in the multi-agent framework to generate the response by generalizing the output.

[0152] In one implementation, the response can be stored as a knowledge entry in a knowledge base; or the response can be output as knowledge data.

[0153] In one implementation, method 1000 may further include: obtaining a modification to the output of the role-playing agent in the discussion group; and storing the modification as short-term memory data of the role-playing agent.

[0154] It should be understood that method 1000 may also include any steps / processes for generating a response using a multi-agent framework according to the embodiments of this disclosure described above.

[0155] Figure 11 An exemplary apparatus 1100 for generating responses using a multi-agent framework, according to an embodiment, is shown.

[0156] The apparatus 1100 may include: a topic acquisition module 1110 for acquiring a topic; a discussion execution module 1120 for executing multiple rounds of discussion in response to the topic, wherein in each round of discussion: using the planning agent, a discussion strategy for the round of discussion is determined; using the planning agent, a set of role-playing agents for forming the discussion group is determined according to the discussion strategy; using the discussion group, the round of discussion is executed according to the discussion strategy; and using the evaluation agent, an evaluation result for the round of discussion is generated, the evaluation result being used to determine whether to execute the next round of discussion; and a response generation module 1130 for generating a response to the topic based on the output of the discussion group in the multiple rounds of discussion.

[0157] In addition, the apparatus 1100 may also include any other module configured to perform any operation of the method for generating a response using a multi-agent framework according to the above embodiments of the present disclosure.

[0158] Figure 12 An exemplary apparatus 1200 for generating responses using a multi-agent framework, according to an embodiment, is shown.

[0159] The apparatus 1200 may include at least one processor 1210. The apparatus 1200 may also include a memory 1220 connected to the at least one processor 1210. The memory 1220 may store computer-executable instructions that, when executed, cause the at least one processor 1210 to: obtain a topic; and, in response to the topic, perform multiple rounds of discussion, wherein, in each round of discussion: using the planning agent, determine a discussion strategy for the round of discussion; using the planning agent, determine a set of role-playing agents for forming the discussion group based on the discussion strategy; using the discussion group, perform the round of discussion according to the discussion strategy; and using the evaluation agent, generate an evaluation result for the round of discussion, the evaluation result being used to determine whether to proceed to the next round of discussion; and generate a response to the topic based on the output of the discussion group in the multiple rounds of discussion. Furthermore, the at least one processor 1210 may also be configured to perform any other operations of the method for generating responses using a multi-agent framework according to the above embodiments of this disclosure.

[0160] Embodiments of this disclosure can be implemented in a non-transitory computer-readable medium. This non-transitory computer-readable medium may include instructions that, when executed, cause at least one processor to: obtain a topic; and, in response to the topic, perform multiple rounds of discussion, wherein, in each round of discussion: using the planning agent, determine a discussion strategy for the round of discussion; using the planning agent, determine a set of role-playing agents for forming the discussion group based on the discussion strategy; using the discussion group, perform the round of discussion according to the discussion strategy; and using the evaluation agent, generate an evaluation result for the round of discussion, the evaluation result being used to determine whether to proceed to the next round of discussion; and generate a response to the topic based on the output of the discussion group in the multiple rounds of discussion. Furthermore, when the instructions are executed, at least one processor may also be caused to perform any step / process of the method for generating a response using a multi-agent framework according to the embodiments of this disclosure described above.

[0161] Embodiments of this disclosure propose a computer program product comprising a computer program executable by a processor to: obtain a topic; and, in response to the topic, perform multiple rounds of discussion, wherein, in each round of discussion: using a planning agent, determine a discussion strategy for the round of discussion; using the planning agent, determine a set of role-playing agents for forming the discussion group based on the discussion strategy; using the discussion group, perform the round of discussion according to the discussion strategy; and using an evaluation agent, generate an evaluation result for the round of discussion, the evaluation result being used to determine whether to proceed to the next round of discussion; and generate a response to the topic based on the output of the discussion group in the multiple rounds of discussion. Furthermore, the computer program may also be executable by the processor to perform any steps / processes of the method for generating responses using a multi-agent framework according to the embodiments of this disclosure described above.

[0162] It should be understood that all operations in the methods described above are merely exemplary, and this disclosure is not limited to any operation in the methods or the order of such operations, but should cover all other equivalent transformations under the same or similar conception.

[0163] Furthermore, unless otherwise specified or clearly indicated from the context that the application is to the singular form, the articles “a” and “an” as used in this specification and the appended claims should generally be interpreted as meaning “one” or “one or more”.

[0164] It should also be understood that all modules in the apparatus described above can be implemented in various ways. These modules can be implemented as hardware, software, or a combination thereof. Furthermore, any of these modules can be further functionally divided into sub-modules or combined together.

[0165] Processors have been described in conjunction with various devices and methods. These processors can be implemented using electronic hardware, computer software, or any combination thereof. Whether these processors are implemented as hardware or software will depend on the specific application and the overall design constraints imposed on the system. As an example, the processors, any portions of processors, or any combinations of processors given in this disclosure can be implemented as microprocessors, microcontrollers, digital signal processors (DSPs), field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), state machines, gate logic, discrete hardware circuits, and other suitable processing units configured to perform the various functions described in this disclosure. The functionality of the processors, any portions of processors, or any combinations of processors given in this disclosure can be implemented as software executed by a microprocessor, microcontroller, DSP, or other suitable platform.

[0166] Software should be broadly considered as representing instructions, instruction sets, code, code segments, program code, programs, subroutines, software modules, applications, software applications, software packages, routines, subroutines, objects, running threads, procedures, functions, etc. Software may reside on a computer-readable medium. Computer-readable media may include, for example, memory, which may be, for example, magnetic storage devices (e.g., hard disks, floppy disks, magnetic stripes), optical disks, smart cards, flash memory devices, random access memory (RAM), read-only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), registers, or removable disks. Although memory is shown as separate from the processor in several aspects set forth in this disclosure, memory may also reside within the processor (e.g., in caches or registers).

[0167] The above description is provided to enable any person skilled in the art to implement the various aspects described herein. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects. Therefore, the claims are not intended to be limited to the aspects shown herein. All structural and functional equivalents of the elements of the various aspects described in this disclosure that are known or about to be known to those skilled in the art shall be covered by the claims.

Claims

1. A method for generating responses using a multi-agent framework, wherein, The multi-agent framework includes a planning agent, an evaluation agent, and a discussion group, and the method includes: Gain a topic; In response to the topic, multiple rounds of discussion are conducted, in which: Using the planning agent, determine the discussion strategy for the first round of discussion; Using the planning agent, a set of role-playing agents is determined according to the discussion strategy to form the discussion group; Using the discussion group, the first round of discussion is conducted according to the discussion strategy; and Using the aforementioned evaluation agent, an evaluation result for the first round of discussion is generated, and this evaluation result is used to determine whether to proceed to the next round of discussion; and Based on the output of the discussion group in the multiple rounds of discussion, a response is generated for the topic.

2. The method according to claim 1, further comprising: Determine whether to proceed with the next round of discussions based on the evaluation results; as well as In response to determining to proceed with the next round of discussion based on the evaluation results, the discussion strategy for determining the first round of discussion includes: The updated discussion strategy should be determined based at least on the evaluation results; or In response to the determination based on the evaluation results that the next round of discussion will not be carried out, the multi-round discussion ends.

3. The method according to claim 2, wherein, The multi-agent framework also includes a generalized agent, and wherein, The method further includes: Using the summarizing agent, a summary of the first round of discussion is generated by summarizing the output of the discussion group in that round; and The evaluation results generated in the first round of discussion include: The evaluation results are generated based on the summary.

4. The method according to claim 3, wherein, The determination of the updated discussion strategy based at least on the evaluation results includes: The updated discussion strategy will be determined based on the evaluation results and the summary.

5. The method according to claim 1, wherein, The step of determining a set of role-playing agents for forming the discussion group based on the discussion strategy includes: According to the discussion strategy, the set of role-playing agents is selected from a library of role-playing agents that includes multiple candidate role-playing agents.

6. The method according to claim 1, further comprising: Based on the discussion strategy, it is determined whether the discussion group includes a moderator agent, wherein the moderator agent is used to guide the group of role-playing agents to perform the first round of discussion according to the discussion strategy.

7. The method according to claim 1, wherein, The evaluation results generated in the first round of discussion include: Determine whether the topic is open or closed; and In response to determining that the topic is an open topic, the evaluation result is generated by assessing the discussion process of the first round of discussion; or In response to determining that the topic is a non-open topic, the evaluation result is generated by evaluating the discussion process of the first round of discussion and / or evaluating the discussion results of the first round of discussion.

8. The method according to claim 7, wherein, The evaluation of the discussion process in the aforementioned round of discussions includes: The discussion process shall be evaluated for at least one of the following: completeness, hierarchy, freshness, and accuracy.

9. The method according to claim 1, wherein, The discussion strategy for determining the first round of discussion includes: If the aforementioned round of discussion is the first round of discussion, the discussion strategy shall be determined based on the aforementioned topic.

10. The method according to claim 1, wherein, Each agent in the multi-agent framework includes: A master-control sub-agent, which is used to determine the thought chain of the agent; A memory organization sub-agent, configured to organize long-term memory data and / or short-term memory data according to the thought chain; and The tool organization sub-agent is used to organize the invocation of external tools according to the thought chain.

11. The method according to claim 10, wherein, For one of the aforementioned set of role-playing agents: The long-term memory data includes at least one of identity data, academic opinion data, and cultural background data, and The short-term memory data includes at least one of the following: role interaction data, modifications to the agent's output, evaluations of the output, and historical self-reflection data.

12. The method according to claim 1, wherein, Each role-playing agent is trained using training knowledge data. The training knowledge data includes at least one of academic data and personalized data, and The training knowledge data includes multimodal data and / or data from multiple sources.

13. The method according to claim 1, wherein, The step of generating a response to the topic based on the output of the discussion group in the multiple rounds of discussion includes: The response is generated by summarizing the output using the generalizing agent in the multi-agent framework.

14. The method according to claim 1, wherein, The response is stored as a knowledge entry in the knowledge base; or The response is output as knowledge data.

15. The method according to claim 1, further comprising: Obtain modifications to the output of the role-playing agents in the aforementioned discussion group; as well as The modifications are stored as short-term memory data of the role-playing agent.

16. An apparatus for generating responses using a multi-agent framework, wherein, The multi-agent framework includes a planning agent, an evaluation agent, and a discussion group, and the device includes: At least one processor; and A memory storing computer-executable instructions that, when executed, cause the at least one processor to: Gain a topic; In response to the topic, multiple rounds of discussion are conducted, in which: Using the planning agent, determine the discussion strategy for the first round of discussion; Using the planning agent, a set of role-playing agents is determined according to the discussion strategy to form the discussion group; Using the discussion group, the first round of discussion is conducted according to the discussion strategy; and Using the aforementioned evaluation agent, an evaluation result for the first round of discussion is generated, and this evaluation result is used to determine whether to proceed to the next round of discussion; and Based on the output of the discussion group in the multiple rounds of discussion, a response is generated for the topic.

17. The apparatus according to claim 16, wherein, When the computer-executable instructions are executed, the at least one processor is also caused to: Determine whether to proceed with the next round of discussions based on the evaluation results; as well as In response to determining to proceed with the next round of discussion based on the evaluation results, the discussion strategy for determining the first round of discussion includes: The updated discussion strategy should be determined based at least on the evaluation results; or In response to the determination based on the evaluation results that the next round of discussion will not be carried out, the multi-round discussion ends.

18. The apparatus according to claim 17, wherein, The multi-agent framework also includes a generalized agent, and wherein, When the computer-executable instructions are executed, the at least one processor is also caused to: Using the summarizing agent, a summary of the first round of discussion is generated by summarizing the output of the discussion group in that round; and The evaluation results generated in the first round of discussion include: The evaluation results are generated based on the summary.

19. The apparatus according to claim 18, wherein, The determination of the updated discussion strategy based at least on the evaluation results includes: The updated discussion strategy will be determined based on the evaluation results and the summary.

20. A non-transitory computer-readable medium comprising instructions that, when executed, cause at least one processor to: Gain a topic; In response to the topic, multiple rounds of discussion were conducted, in which, In each round of discussion: The planning agent in the multi-agent framework is used to determine the discussion strategy for the first round of discussion; Using the planning agent, a set of role-playing agents are determined according to the discussion strategy to form a discussion group in the multi-agent framework; The discussion group is used to conduct the first round of discussion according to the discussion strategy. as well as Using the evaluation agent in the multi-agent framework, an evaluation result for the first round of discussion is generated, and the evaluation result is used to determine whether to proceed to the next round of discussion; as well as Based on the output of the discussion group in the multiple rounds of discussion, a response is generated for the topic.

Images