A method for generating content diversity by deep reasoning and agent synergistic enhancement model
By enhancing the model through deep reasoning and agent collaboration, complex problems are broken down into sub-problems and solved collaboratively by multiple intelligent agents. This solves the problem of single-model reasoning paths and achieves multi-dimensional understanding and high-quality answer generation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INSPUR SOFTWARE TECH CO LTD
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-29
AI Technical Summary
When dealing with complex problems, existing technologies rely on a single model with a limited reasoning path, resulting in incomplete content coverage and frequent information omissions. This makes it particularly difficult to comprehensively answer multi-faceted questions, especially in resource-constrained environments.
By enhancing the model through deep reasoning and agent collaboration, user questions are broken down into multiple sub-questions, which are then solved collaboratively by multiple intelligent agents to generate diverse answers. This includes a breakdown and expansion module, a problem-solving module, and an output module. Answer planning and identification are performed through moderator and expert models.
It enhances the model's ability to understand questions from multiple dimensions, improves the quality and diversity of responses in resource-constrained environments, and meets users' needs for high-quality output in intelligent question answering, content generation, and knowledge management.
Smart Images
Figure CN122114138A_ABST
Abstract
Description
Technical Field
[0001] This invention discloses a method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, which relates to the fields of artificial intelligence and natural language processing technology. Background Technology
[0002] In existing technologies, basic models often rely on a single inference path when dealing with complex problems, making it difficult to comprehensively cover the multiple dimensions of the problem. For example, while existing technologies such as Transformer-based models perform well in natural language processing tasks, their inference process is easily limited by task guidance in resource-constrained environments, resulting in insufficient breadth and depth of generated content. Specifically, Transformer models capture global dependencies in input sequences through self-attention mechanisms, achieving significant results in tasks such as translation, summarization, and question answering. However, when dealing with problems requiring multi-faceted analysis, single models are often limited by their fixed context windows and limited computing resources, unable to fully expand and deepen the reasoning. Especially in edge computing devices or low-power scenarios, the input limitations of the model make it exclusively focused on the problem result, thus affecting the completeness of the reasoning and the comprehensiveness of the answer. Furthermore, existing technologies often lack systematic decomposition and organization mechanisms when dealing with multifaceted problems, making it easy for models to miss key information or provide only superficial solutions when generating answers. For example, when faced with problems involving multiple domains or requiring multi-step reasoning, the model may only focus on the most direct answer path, ignoring other potentially important aspects. This limitation not only affects the model's practicality but also restricts its performance on complex tasks. Summary of the Invention
[0003] This invention provides a method to enhance the diversity of content generated by a model through deep reasoning and agent collaboration, which solves key problems in current artificial intelligence models when dealing with complex problems, such as incomplete content coverage, single reasoning path, and frequent information omission.
[0004] The specific solution proposed in this invention is as follows: This invention also provides a method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, comprising: Step 1 involves breaking down and expanding the user's question: Semantic analysis is performed on the input question, and related domains are expanded using a large language model. Multiple sub-questions are then generated and inserted into a question queue. Step 2: Create a moderator based on the large model. The moderator retrieves sub-problems from the problem queue, plans solutions, selects appropriate experts to solve the problems based on the solution plans, obtains answers, and determines whether the sub-problem answers are correct. If correct, they are used to generate the answer to the original problem; otherwise, the process is repeated. Step 3: Based on the correct answers to each sub-question, synthesize and provide a comprehensive natural language answer to the original question.
[0005] Furthermore, in step 1 of the method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, the generated sub-problems are sorted according to priority. The priority sorting includes sorting by dependency, sorting by execution efficiency and cost, and sorting by risk assessment and impact. Sorting by execution efficiency and cost includes prioritizing problems that can eliminate a large amount of uncertainty or prioritizing low-cost problems. Sorting by risk assessment and impact includes prioritizing problems with high feasibility.
[0006] Furthermore, in step 2 of the method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, creating a host based on a large model includes: The facilitator creates the planning and solution steps, selects experts, and evaluates the answers. The facilitator who plans the solution steps removes the sub-problems from the problem queue and plans solutions. The facilitator who selects experts chooses the appropriate experts according to the plan. The facilitator who identifies the answers conducts an initial assessment of the answers and provides feedback to the facilitator who plans the solution steps to determine whether the sub-problem answers are correct. If they are incorrect, the facilitator repeats the attempt.
[0007] Furthermore, in step 2 of the method for enhancing the diversity of content generated by the model through deep reasoning and agent collaboration, the selected experts solve the problem and obtain the answer. This includes: the moderator who selects the experts submits the problem to the experts, including expert models and corresponding services; the expert models fill in the terminology according to the sub-problem description and generate parameters; the parameters are passed to the corresponding tools; the results returned by the tools are returned to the expert models and corresponding services; the expert models and corresponding services professionally identify and filter the answers, filtering out sensitive or abnormal results; and then the relevant results are submitted to the moderator of the planning and solution steps. The moderator of the planning and solution steps solves the problem according to the problem-solving process until the problem queue is empty.
[0008] This invention also provides a device for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, including a splitting and expansion module, a problem-solving module, and an output module. The splitting and expansion module splits and expands user questions: it performs semantic analysis on the input question, expands it to related domains using a large language model, and then generates multiple sub-questions, which are then inserted into a question queue. The problem-solving module creates a moderator based on the overall model. The moderator then retrieves sub-problems from the problem queue, plans solutions, selects appropriate experts to solve the problems based on the solutions, obtains answers, and determines whether the sub-problem answers are correct. If correct, these answers are used to generate the answer to the original problem; otherwise, the process is repeated. The output module combines the correct answers to each sub-question to provide a comprehensive natural language answer to the original question.
[0009] Furthermore, in the device for enhancing the diversity of content generated by the model through deep reasoning and agent collaboration, the splitting and expansion modules sort the generated sub-problems according to priority. The priority sorting includes sorting according to dependency, sorting according to execution efficiency and cost, and sorting according to risk assessment and impact. Sorting according to execution efficiency and cost includes prioritizing problems that can eliminate a large amount of uncertainty or prioritizing low-cost problems. Sorting according to risk assessment and impact includes prioritizing problems with high feasibility.
[0010] Furthermore, the problem-solving module in the device for enhancing the diversity of content generated by the model through deep reasoning and agent collaboration creates a host based on the large model, including: The facilitator creates the planning and solution steps, selects experts, and evaluates the answers. The facilitator who plans the solution steps removes the sub-problems from the problem queue and plans solutions. The facilitator who selects experts chooses the appropriate experts according to the plan. The facilitator who identifies the answers conducts an initial assessment of the answers and provides feedback to the facilitator who plans the solution steps to determine whether the sub-problem answers are correct. If they are incorrect, the facilitator repeats the attempt.
[0011] Furthermore, in the device for enhancing the diversity of content generated by the model through deep reasoning and agent collaboration, the problem-solving module solves problems and obtains answers by selecting experts. This includes: a moderator who selects experts submits the problem to the experts, including expert models and corresponding services; the expert models fill in the terminology based on the sub-problem description and generate parameters; the parameters are passed to the corresponding tools; the results returned by the tools are then returned to the expert models and corresponding services; the expert models and corresponding services professionally identify and filter the answers, filtering out sensitive or abnormal results; and then the relevant results are submitted to the moderator who plans the solution steps. The moderator then solves the problem according to the problem-solving process until the problem queue is empty.
[0012] The advantages of this invention are: This invention introduces a three-stage deep reasoning mechanism: problem decomposition and expansion, organization and confirmation of the aspects involved in the problem, and solution planning for each step. Through a structured, hierarchical, and collaborative processing flow, the model can gradually deepen its understanding of the complexity of the problem and systematically organize the reasoning path. Under this mechanism, the model first decomposes the input problem into several related sub-problems, thus breaking through the limitations of a single input. Finally, with the collaborative action of multiple agents, reasoning and solving are performed on different sub-problems respectively, and the outputs of each agent are integrated to form a final answer that is clearly structured, rich in content, and diverse in perspective. This method not only enhances the model's ability to understand problems from multiple dimensions but also significantly improves the quality and diversity of its answers in resource-constrained environments, thereby better meeting users' needs for high-quality and comprehensive output in applications such as intelligent question answering, content generation, and knowledge management. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0014] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example
[0015] This invention also provides a method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, comprising: Step 1 involves breaking down and expanding the user's problem: Semantic analysis is performed on the input problem, and related domains are expanded using a large language model to generate multiple sub-problems, which are then inserted into a problem queue. These sub-problems are then prioritized, including by dependency, execution efficiency and cost, and risk assessment and impact. Prioritization by execution efficiency and cost prioritizes problems that minimize uncertainty or have low cost. Prioritization by risk assessment and impact prioritizes problems with high feasibility.
[0016] Step 2: Create a moderator based on the large model. The moderator retrieves sub-problems from the problem queue, plans solutions, and selects appropriate experts to solve the problems and obtain answers. This process involves the moderator submitting the problem to the experts (including expert models and their corresponding services). The expert models then fill in the sub-problem descriptions with terminology and generate parameters. These parameters are passed to the appropriate tools, and the results are returned to the expert models and their corresponding services. The expert models and their services professionally evaluate and filter the answers, removing sensitive or abnormal results. These results are then submitted back to the moderator planning the solution steps. The moderator then solves the problems according to the problem-solving process until the problem queue is empty.
[0017] Determine if the answer to the subproblem is correct. If it is, use it to generate the answer to the original problem; otherwise, repeat the process.
[0018] The creation of hosts based on the large model includes: The facilitator creates the planning and solution steps, selects experts, and evaluates the answers. The facilitator who plans the solution steps removes the sub-problems from the problem queue and plans solutions. The facilitator who selects experts chooses the appropriate experts according to the plan. The facilitator who identifies the answers conducts an initial assessment of the answers and provides feedback to the facilitator who plans the solution steps to determine whether the sub-problem answers are correct. If they are incorrect, the facilitator repeats the attempt.
[0019] Step 3: Based on the correct answers to each sub-question, synthesize and provide a comprehensive natural language answer to the original question.
[0020] This invention also provides a device for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, including a splitting and expansion module, a problem-solving module, and an output module. The splitting and expansion module splits and expands user questions: it performs semantic analysis on the input question, expands it to related domains using a large language model, and then generates multiple sub-questions, which are then inserted into a question queue. The problem-solving module creates a moderator based on the overall model. The moderator then retrieves sub-problems from the problem queue, plans solutions, selects appropriate experts to solve the problems based on the solutions, obtains answers, and determines whether the sub-problem answers are correct. If correct, these answers are used to generate the answer to the original problem; otherwise, the process is repeated. The output module combines the correct answers to each sub-question to provide a comprehensive natural language answer to the original question.
[0021] The information interaction and execution process between the modules in the above-mentioned device are based on the same concept as the method embodiment of the present invention, and the specific details can be found in the description in the method embodiment of the present invention, and will not be repeated here.
[0022] Similarly, the device of this invention introduces a three-stage deep reasoning mechanism: problem decomposition and expansion, organization and confirmation of the aspects involved in the problem, and solution planning for each step. Through a structured, hierarchical, and collaborative processing flow, the model can gradually deepen its understanding of the complexity of the problem and systematically organize the reasoning path. Under this mechanism, the model first decomposes the input problem into several related sub-problems, thereby breaking through the limitation of a single input. Finally, with the collaborative action of multiple agents, reasoning and answering different sub-problems are performed separately, and the outputs of each agent are integrated to form a final answer that is clearly structured, rich in content, and diverse in perspective. This method not only enhances the model's ability to understand problems from multiple dimensions but also significantly improves the quality and diversity of its answers in resource-constrained environments, thereby better meeting users' needs for high-quality and comprehensive output in applications such as intelligent question answering, content generation, and knowledge management.
[0023] It should be noted that not all steps and modules in the above processes and device structures are mandatory; some steps or modules may be omitted as needed. The execution order of the steps is not fixed and can be adjusted as required. The device structures described in the above embodiments can be physical or logical structures. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.
[0024] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.
Claims
1. A method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, characterized by: include: Step 1 involves breaking down and expanding the user's question: Semantic analysis is performed on the input question, and related domains are expanded using a large language model. Multiple sub-questions are then generated and inserted into a question queue. Step 2: Create a moderator based on the large model. The moderator retrieves sub-problems from the problem queue, plans solutions, selects appropriate experts to solve the problems based on the solution plans, obtains answers, and determines whether the sub-problem answers are correct. If correct, they are used to generate the answer to the original problem; otherwise, the process is repeated. Step 3: Based on the correct answers to each sub-question, synthesize and provide a comprehensive natural language answer to the original question.
2. The method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration as described in claim 1, characterized in that in step 1, the generated multiple sub-problems are sorted according to priority, wherein the priority sorting includes sorting according to dependency, sorting according to execution efficiency and cost, and sorting according to risk assessment and impact; sorting according to execution efficiency and cost includes prioritizing problems that can eliminate a large amount of uncertainty, or prioritizing low-cost problems; sorting according to risk assessment and impact includes prioritizing problems with high feasibility.
3. The method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration as described in claim 1, characterized in that: Step 2 involves creating a presenter based on the large model, including: The facilitator creates the planning and solution steps, selects experts, and evaluates the answers. The facilitator who plans the solution steps removes the sub-problems from the problem queue and plans solutions. The facilitator who selects experts chooses the appropriate experts according to the plan. The facilitator who evaluates the answers conducts an initial evaluation of the answers and provides feedback to the facilitator who plans the solution steps to determine whether the sub-problem answers are correct. If they are incorrect, the facilitator repeats the attempt.
4. The method for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration as described in claim 1, characterized in that: Step 2 involves obtaining answers by selecting experts. This includes: the moderator who selected the experts submits the problem to the experts, who include expert models and their corresponding services. The expert models fill in the terminology based on the sub-problem description and generate parameters. These parameters are then passed to the appropriate tools. The results returned by the tools are then sent back to the expert models and their corresponding services. The expert models and their corresponding services professionally identify and filter the answers, removing sensitive or abnormal results. The relevant results are then submitted to the moderator who plans the solution steps. The moderator then solves the problem according to the problem-solving process until the problem queue is empty.
5. A device for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration, characterized in that: It includes a splitting and expansion module, a problem-solving module, and an output module. The splitting and expansion module splits and expands user questions: it performs semantic analysis on the input question, expands it to related domains using a large language model, and then generates multiple sub-questions, which are then inserted into a question queue. The problem-solving module creates a moderator based on the overall model. The moderator then retrieves sub-problems from the problem queue, plans solutions, selects appropriate experts to solve the problems based on the solutions, obtains answers, and determines whether the sub-problem answers are correct. If correct, these answers are used to generate the answer to the original problem; otherwise, the process is repeated. The output module combines the correct answers to each sub-question to provide a comprehensive natural language answer to the original question.
6. The apparatus for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration as described in claim 5, characterized in that: The splitting and expansion module sorts the generated sub-problems according to priority. Priority sorting includes sorting by dependency, by execution efficiency and cost, and by risk assessment and impact. Sorting by execution efficiency and cost includes prioritizing problems that can eliminate a large amount of uncertainty or prioritizing low-cost problems. Sorting by risk assessment and impact includes prioritizing problems with high feasibility.
7. The apparatus for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration as described in claim 5, characterized in that the problem... The solution module creates a host based on the large model, including: The facilitator creates the planning and solution steps, selects experts, and evaluates the answers. The facilitator who plans the solution steps removes the sub-problems from the problem queue and plans solutions. The facilitator who selects experts chooses the appropriate experts according to the plan. The facilitator who evaluates the answers conducts an initial evaluation of the answers and provides feedback to the facilitator who plans the solution steps to determine whether the sub-problem answers are correct. If they are incorrect, the facilitator repeats the attempt.
8. The apparatus for enhancing the diversity of content generated by a model through deep reasoning and agent collaboration as described in claim 5, characterized in that the problem... The problem-solving module obtains answers by selecting experts. This process includes: the moderator who selected the experts submits the problem to the experts, who include expert models and their corresponding services; the expert models fill in the terminology based on the sub-problem description and generate parameters; the parameters are passed to the appropriate tools; the results returned by the tools are then sent back to the expert models and their corresponding services; the expert models and their corresponding services professionally identify and filter the answers, removing sensitive or abnormal results; and the relevant results are then submitted to the moderator who plans the solution steps. The moderator then solves the problem according to the problem-solving process until the problem queue is empty.