A multi-agent large language model reasoning method and system based on high-order cooperation

Abstract

The application discloses a kind of based on high-order cooperation's multi-agent large language model inference method and system, comprising: obtaining the input of to-be-processed task, and the mapping relationship between the task organization hypergraph, hyperedge dependency graph, hyperedge and task unit corresponding to to-be-processed task input;Extract task signal and determine the cooperation mode corresponding to each hyperedge;Distribute collaboration role and generate corresponding collaboration agreement;Control the plurality of agents in each hyperedge around corresponding task unit to perform collaborative reasoning, obtain corresponding intermediate reasoning result;Standardize and encapsulate each intermediate reasoning result, generate structured result record;The structured result record of current hyperedge and the structured result record of its predecessor hyperedge are aggregated consistently in dependence, obtain the aggregated result record corresponding to each hyperedge;Generate the final reasoning result corresponding to the input of to-be-processed task.The application can improve the matching degree between cooperation mode and task unit features, improve the organization of hyperedge internal cooperation process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a reasoning method and system based on a high-order collaborative multi-agent large language model. Background Technology

[0002] In multi-agent large language model systems, multiple agents typically need to perform collaborative processing around different task units, and further integrate the results after completing local processing to form a holistic output. In existing technologies, collaboration among multiple agents usually employs a unified protocol, such as fixed-round discussions, voting, sequential calls, or free dialogue. However, different task units differ significantly in terms of constraint coupling, exploration needs, stage refinement needs, and result review needs. If all task units adopt the same collaboration method, it becomes difficult to simultaneously address the processing requirements of different task scenarios.

[0003] Specifically, for task units with strong constraint coupling, multiple agents often need to coordinate constraints and reach a consensus; for task units with clear stage divisions, it is more suitable to process them step by step in sequence; for task units that need to generate multiple candidate solutions, it is more suitable to explore in parallel and then integrate them. If there is a lack of a collaboration mode selection mechanism and role allocation mechanism based on the characteristics of the task unit, problems such as low collaboration efficiency, unstable quality of candidate results, or mismatch between local processing methods and task requirements can easily occur.

[0004] On the other hand, in existing technologies, the intermediate results generated by each local collaborative unit are usually output in a non-uniform format, and there are differences between different results in terms of field structure, expression granularity, and reusability. Due to the lack of a unified encapsulation method, subsequent modules often need to perform additional parsing or conversion when calling the predecessor results, which can easily lead to problems such as difficulty in result inheritance, difficulty in identifying conflicts, and unclear reasoning chains.

[0005] Furthermore, in complex tasks involving multiple task units, there are often predecessor dependencies between the local collaborative units. Existing technologies typically integrate intermediate results using methods such as direct concatenation, unified summarization, or free referencing, lacking a constraint-based aggregation mechanism based on dependencies. This can easily lead to subsequent processing units prematurely introducing information they shouldn't access, or ignoring predecessor constraints during result integration, thus affecting the consistency and traceability of the final output.

[0006] Therefore, existing technologies require a new multi-agent collaborative reasoning scheme to achieve adaptive collaborative mode selection and role allocation for different task units based on the existing collaborative organizational structure, and to uniformly encapsulate and consistently aggregate the intermediate results of each local collaborative unit, thereby improving the task adaptability, result reusability and global consistency in the multi-agent collaborative reasoning process. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-agent large language model reasoning method and system based on high-order collaboration, in order to solve the problems of mismatch between collaboration methods and task requirements, inconsistent intermediate result formats, lack of dependency constraints in cross-task unit aggregation, and insufficient consistency of final results in existing multi-agent systems under complex task scenarios, thereby improving task adaptability, result reusability, and global consistency in the multi-agent collaborative reasoning process.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A reasoning method for a multi-agent large language model based on high-order collaboration includes the following steps: S1. Obtain the input of the task to be processed, as well as the task organization hypergraph, hyperedge dependency graph, and mapping relationship between the hyperedge and task unit corresponding to the input of the task to be processed; S2. Extract task signals for the task units corresponding to each hyperedge, and determine the cooperation mode corresponding to each hyperedge based on the task signals. S3. Based on the determined cooperation mode, assign cooperation roles to the multiple agents participating in the cooperation within each hyperedge, and generate the corresponding cooperation protocol. S4. Based on the cooperation protocol, control multiple agents within each hyperedge to perform collaborative reasoning around the corresponding task unit to obtain the corresponding intermediate reasoning results; S5. Standardize and encapsulate each intermediate reasoning result to generate a structured result record; S6. According to the dependency order represented by the hyperedge dependency graph, perform dependency-consistent aggregation on the structured result record of the current hyperedge and the structured result record of its predecessor hyperedge to obtain the aggregation result record corresponding to each hyperedge. S7. Based on the aggregation result record, hyperedge dependency graph, hyperedge weight, and global state corresponding to each hyperedge, generate the final inference result corresponding to the input of the task to be processed.

[0009] Furthermore, in S1, the task organization hypergraph H is represented as: H=(V,E,w) Where V is the set of nodes, E is the set of hyperedges, and w is the hyperedge weight function; The hyperedge dependency graph G_E is represented as: G_E=(E,F) in, This represents the set of dependencies between hyperedges.

[0010] Further, S2 includes: Extract the task signal vector xk for the task unit corresponding to the k-th hyperedge: xk=φ(τk,wk,χk) And determine the target cooperation mode mk based on the task signal vector: mk=π_mode(xk),mk∈M Where φ(·) represents the task vector encoding function, τk represents the task unit corresponding to the kth hyperedge, wk represents the weight of the kth hyperedge, χk represents the context state information corresponding to the kth hyperedge, M represents the preset collaboration mode set, including one or more of consensus collaboration mode, sequential collaboration mode, brainstorming collaboration mode, leadership decision-making collaboration mode and supervision and review collaboration mode, and π_mode(·) represents the collaboration mode routing.

[0011] Further, S3 includes: Let Ak be the set of agents inside the k-th hyperedge: Ak={a(k,k1),a(k,k2),…,a(k,kn)} Where a(k,ki) represents the i-th agent of the k-th hyperedge; Let the set of characters be R: R = {r1, r2, ..., rP} Where ri represents different expert attributes, and P represents the total number of expert attributes; Roles are assigned to the agent set Ak based on the target cooperation pattern: π_role(k):Ak→R Where π_role(k) represents the role assignment result, and → represents the function mapping; And generate the collaboration protocol Π_k based on the target collaboration mode mk and the role assignment result π_role(k): Π_k=Γ(mk,π_role(k)) Here, Γ(·) represents the integration function of collaboration mode and role routing.

[0012] Further, S4 includes: Initialize the local state set and shared global state of multiple agents within each hyperedge; And perform hyperedge-internal collaborative reasoning according to the collaboration protocol, satisfying the following relationship: yk,{s+_(k,i)}(i=1..n_k),S_glob+=F_e(k)^Π_k({s(k,i)}(i=1..nk),τk,S_glob) Where F_e(k)^Π_k represents the inference function of the k-th hyperedge under the cooperative protocol representation, yk represents the intermediate inference result generated by the k-th hyperedge, s(k,i) represents the local state of the i-th agent in the k-th hyperedge before the update, s+_(k,i) represents the local state of the i-th agent in the k-th hyperedge after the update, S_glob represents the global state before the update, S_glob+ represents the global state after the update, and τk represents the task unit corresponding to the k-th hyperedge.

[0013] Further, S5 includes: The intermediate inference results generated by the k-th hyperedge are standardized and encapsulated to obtain the structured result rk record: rk=Φ_pack(yk,τk,S_glob+) Wherein, Φ_pack represents the standardized encapsulation function, yk represents the intermediate reasoning result generated by the kth hyperedge, and the structured result record rk includes one or more of the following: conclusion claim, satisfied constraint information, unsatisfied constraint information, evidence information, open question information, and source identification information.

[0014] Further, S6 includes: Perform a topological sort on the hyperedge dependency graph to obtain the aggregate sequence; For the current hyperedge ek, determine its set of predecessor hyperedges Pred(ek): Pred(ek)={ej∈E|(ej,ek)∈F} Where ej represents the j-th hyperedge, and E represents the set of hyperedges. Represents the set of dependencies between hyperedges; And perform dependency consistent aggregation: r~k=Φ_agg(rk,{r~j|ej∈Pred(ek)},wk,S_glob+) Where Φ_agg represents the aggregation function, r~k represents the aggregation result record corresponding to the kth hyperedge, and wk represents the weight of the kth hyperedge.

[0015] Furthermore, S6 also includes: During dependency-consistent aggregation, only the result record corresponding to the predecessor of the current superedge is allowed to be introduced into the current aggregation process, and the result record of the successor superedge is not introduced.

[0016] Further, S7 includes: Get the aggregated result set R of all hyperedges: R = {r1, r2, ..., rM} r~k represents the aggregation result record corresponding to the kth hyperedge, and M represents the set of preset collaboration modes; The final result is generated based on the aggregation result set R~, the hyperedge dependency graph G_E, the hyperedge weights, and the global state S_glob+. :

[0017] In the formula, Φ_dec(·) represents the overall encoding function, and wk represents the weight of the k-th hyperedge.

[0018] This invention also provides a multi-agent large language model reasoning system based on high-order collaboration, used to implement the multi-agent large language model reasoning method based on high-order collaboration as described above, including: The task acquisition module is used to acquire the input of the task to be processed, as well as the task organization hypergraph, hyperedge dependency graph and mapping relationship corresponding to the input of the task to be processed; The pattern routing module is used to extract task signals for each hyperedge corresponding to the task unit and determine the target cooperation mode based on the task signals. The role assignment module is used to assign collaborative roles to multiple agents within each hyperedge based on the target collaboration mode. The collaboration protocol generation module is used to generate a collaboration protocol based on the target collaboration mode and the assignment of collaboration roles. The collaborative execution module is used to control multiple agents within each hyperedge to perform collaborative reasoning around the corresponding task unit and generate hyperedge-level intermediate results; The result encapsulation module is used to perform standardized encapsulation on the intermediate results generated by each hyperedge, and generate structured result records. The dependency aggregation module is used to perform consistent dependency aggregation according to the topological order of the hyperedge dependency graph; and The global decision module is used to generate the final result by combining the aggregation results of each hyperedge, dependency structure, hyperedge weights, and global state.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1) It can improve the matching degree between collaboration methods and task unit characteristics. Target collaboration modes are selected based on task unit characteristics, and collaboration roles are further assigned, so that different task units can adopt more suitable processing methods.

[0020] 2) It can improve the organization of the collaboration process within the hyperedge. By generating collaboration protocols through collaboration modes and role assignments, multiple agents within the hyperedge can complete processes such as candidate generation, constraint checking, criticism correction, integration adjudication, or review correction according to predetermined responsibilities and interaction order.

[0021] 3) It can improve the standardization and reusability of intermediate results. By standardizing and encapsulating the hyperedge-level intermediate results, result records with a unified field structure are generated, enabling subsequent modules to directly call and process the predecessor results.

[0022] 4) It can improve the consistency of result integration across task units. Based on the hyperedge dependency relationship, it performs dependency-consistent aggregation, allowing only the current hyperedge to call the result record of its predecessor hyperedge, thereby keeping the result integration process consistent with the task dependency direction.

[0023] 5) It can reduce the risk of information leakage and enhance the traceability of global results. By limiting the scope of results introduced during the aggregation process, it prevents information from subsequent task units from being fed back into the aggregation process of preceding task units.

[0024] 6) It can improve the overall collaborative reasoning quality in complex task scenarios. By combining the internal collaboration mechanism of the hyperedge with the cross-hyperedge aggregation mechanism, the multi-agent system can not only produce high-quality interim results at the local task unit level, but also produce a more consistent final output at the global level. Attached Figure Description

[0025] Figure 1 This is a flowchart of a multi-agent large language model reasoning method based on high-order collaboration according to the present invention.

[0026] Figure 2 This is a block diagram of a multi-agent large language model inference system based on high-order collaboration according to the present invention.

[0027] Figure 3 This is a ranking comparison chart of various methods under different judges (with standard deviation). Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Please see Figure 1 A reasoning method based on a high-order collaborative multi-agent large language model includes the following steps: S1. Obtain the input of the task to be processed, as well as the task organization hypergraph, hyperedge dependency graph, and mapping relationship between hyperedge and task unit corresponding to the input of the task to be processed.

[0030] The task organization hypergraph H is represented as: H=(V,E,w) Where V is the set of nodes, E is the set of hyperedges, and w is the hyperedge weight function.

[0031] The hyperedge dependency graph G_E is represented as: G_E=(E,F) in, This represents the set of dependencies between hyperedges.

[0032] S2. Extract task signals for the task units corresponding to each hyperedge, and determine the cooperation mode corresponding to each hyperedge based on the task signals, including: Extract the task signal vector xk for the task unit corresponding to the k-th hyperedge: xk=φ(τk,wk,χk) And determine the target cooperation mode mk based on the task signal vector: mk=π_mode(xk),mk∈M Where φ(·) represents the task vector encoding function, τk represents the task unit corresponding to the k-th hyperedge, wk represents the weight of the k-th hyperedge, and χk represents the context state information corresponding to the k-th hyperedge. M represents a set of preset collaboration modes, including consensus-based collaboration mode, sequential collaboration mode, brainstorming collaboration mode, leadership decision-making collaboration mode, and supervision and review collaboration mode. π_mode(·) represents the collaboration mode routing.

[0033] S3. Based on the determined cooperation mode, assign cooperation roles to the multiple agents participating in the cooperation within each hyperedge, and generate corresponding cooperation protocols, including: Let Ak be the set of agents inside the k-th hyperedge: Ak={a(k,k1),a(k,k2),…,a(k,kn)} Where a(k,ki) represents the i-th agent of the k-th hyperedge.

[0034] Let the set of characters be R: R = {r1, r2, ..., rP} Where ri represents different expert attributes, and P represents the total number of expert attributes.

[0035] Roles are assigned to the agent set Ak based on the target cooperation pattern: π_role(k):Ak→R Where π_role(k) represents the role assignment result, and → represents the function mapping.

[0036] And generate the collaboration protocol Π_k based on the target collaboration mode mk and the role assignment result π_role(k): Π_k=Γ(mk,π_role(k)) Here, Γ(·) represents the integration function of collaboration mode and role routing.

[0037] S4. Based on the cooperation protocol, control multiple agents within each hyperedge to perform collaborative reasoning around the corresponding task unit, and obtain the corresponding intermediate reasoning results, including: Initialize the local state set and shared global state of multiple agents within each hyperedge; And perform hyperedge-internal collaborative reasoning according to the collaboration protocol, satisfying the following relationship: yk,{s+_(k,i)}(i=1..n_k),S_glob+=F_e(k)^Π_k({s(k,i)}(i=1..nk),τk,S_glob) Where F_e(k)^Π_k represents the inference function of the k-th hyperedge under the cooperative protocol representation, yk represents the intermediate inference result generated by the k-th hyperedge, s(k,i) represents the local state of the i-th agent in the k-th hyperedge before the update, s+_(k,i) represents the local state of the i-th agent in the k-th hyperedge after the update, S_glob represents the global state before the update, S_glob+ represents the global state after the update, and τk represents the task unit corresponding to the k-th hyperedge.

[0038] S5. Standardize and encapsulate each intermediate inference result to generate a structured result record, including: The intermediate inference results generated by the k-th hyperedge are standardized and encapsulated to obtain the structured result rk record: rk=Φ_pack(yk,τk,S_glob+) Wherein, Φ_pack represents the standardized encapsulation function, and the structured result record rk includes the conclusion claim, satisfied constraint information, unsatisfied constraint information, evidence information, open question information, and source identification information.

[0039] S6. According to the dependency order represented by the hyperedge dependency graph, perform dependency-consistent aggregation on the structured result record of the current hyperedge and the structured result record of its predecessor hyperedge to obtain the aggregation result record corresponding to each hyperedge, including: Perform a topological sort on the hyperedge dependency graph to obtain the aggregate sequence.

[0040] For the current hyperedge ek, determine its set of predecessor hyperedges Pred(ek): Pred(ek)={ej∈E|(ej,ek)∈F} Here, ej represents the j-th superedge.

[0041] And perform dependency consistent aggregation: r~k=Φ_agg(rk,{r~j|ej∈Pred(ek)},wk,S_glob+) Where Φ_agg represents the aggregation function, and r~k represents the aggregation result record corresponding to the kth hyperedge.

[0042] During dependency-consistent aggregation, only the result record corresponding to the predecessor of the current superedge is allowed to be introduced into the current aggregation process, and the result record of the successor superedge is not introduced.

[0043] S7. Based on the aggregation result records corresponding to each hyperedge, the hyperedge dependency graph, the hyperedge weights, and the global state, generate the final inference result corresponding to the input of the task to be processed, including: Get the aggregated result set R of all hyperedges: R = {r1, r2, ..., rM} The final result is generated based on the aggregation result set R~, the hyperedge dependency graph G_E, the hyperedge weights, and the global state S_glob+. :

[0044] In the formula, Φ_dec(·) represents the overall encoding function.

[0045] Please see Figure 2 A high-order collaborative multi-agent large language model reasoning system is provided to implement the high-order collaborative multi-agent large language model reasoning method described above, comprising: The task acquisition module is used to acquire the input of the task to be processed, as well as the task organization hypergraph, hyperedge dependency graph and mapping relationship corresponding to the input of the task to be processed. The operation of the task acquisition module is described in S1 and S2 of the above-mentioned inference method based on high-order collaborative multi-agent large language model.

[0046] The pattern routing module is used to extract task signals for each hyperedge-corresponding task unit and determine the target cooperation mode based on the task signals. The operation of the pattern routing module is described in section S3 of the above-mentioned inference method based on a high-order cooperative multi-agent large language model.

[0047] The role assignment module is used to assign collaborative roles to multiple agents within each hyperedge based on the target collaboration mode. The operation of the role assignment module is described in section S3 of the multi-agent large language model inference method based on high-order collaboration.

[0048] The collaboration protocol generation module is used to generate a collaboration protocol based on the target collaboration mode and the allocation of collaboration roles. The operation of the collaboration protocol generation module is described in section S3 of the above-mentioned inference method based on a high-order collaborative multi-agent large language model.

[0049] The collaborative execution module controls multiple agents within each hyperedge to perform collaborative inference around their corresponding task units, generating hyperedge-level intermediate results. The operation of the collaborative execution module is described in section S4 of the aforementioned multi-agent large language model inference method based on high-order collaboration.

[0050] The result encapsulation module is used to perform standardized encapsulation of the intermediate results generated by each hyperedge, generating structured result records. The operation of the result encapsulation module is described in S5 of the above-mentioned inference method based on a high-order collaborative multi-agent large language model.

[0051] The dependency aggregation module performs consistent dependency aggregation according to the topological order of the hyperedge dependency graph. For details on how the dependency aggregation module works, refer to S6 in the above-mentioned inference method based on a high-order cooperative multi-agent large language model.

[0052] The global decision-making module is used to generate the final result by combining the aggregation results of each hyperedge, dependency structure, hyperedge weights, and global state. The operation of the global decision-making module is described in S7 of the aforementioned inference method based on a high-order collaborative multi-agent large language model.

[0053] The aforementioned task acquisition module, pattern routing module, role allocation module, collaboration protocol generation module, result encapsulation module, dependency aggregation module, and global decision-making module are all computer programs.

[0054] Example: To illustrate the implementation process of this invention in complex long-term task scenarios, this example selects the "Capital Allocation Task under Coupling Constraints" from the CLM Complex Long-Term Multi-Task Dataset as an example task. The CLM dataset contains 50 challenging task scenarios, which typically feature cross-domain coupling and multiple constraints. On average, each task contains 3-4 hard constraints, making it suitable for verifying the technical effectiveness of this invention in pattern routing, role assignment, result encapsulation, and dependency consistency aggregation.

[0055] Table 1. Schematic diagram of the task scenarios in this embodiment.

[0056] In this embodiment, the "capital allocation task under coupling constraints" includes at least the following input information: current business background information, budget size information, minimum core business capital retention requirements, risk-return characteristics of different business directions, and final output requirements. The task requires the system to generate a capital allocation result that balances risk, return, and development flexibility, while meeting budget constraints and core business minimum requirements.

[0057] Table 2. Schematic diagram of task unit division in this embodiment.

[0058] Table 3. Collaboration Mode Configuration Diagram

[0059] Table 4. Scheme of Structured Result Recording

[0060] Table 5. Dependency Aggregation Order Diagram

[0061] To verify the effectiveness of this embodiment, the method of the present invention was experimentally evaluated on the CLM dataset. The CLM task output is an open-ended result; a unified output template was used, and comparisons were made according to a consistent evaluation protocol, with the average ranking used as the overall result metric. Figure 3 As shown in Table 6, the experimental results demonstrate that the method of the present invention can achieve better overall performance in complex long-range task scenarios.

[0062] Table 6 Overall Evaluation Results on the CLM Dataset

[0063] Depend on Figure 3 As shown in Table 6, the method of this invention achieved the best average ranking in all four evaluation models on the CLM dataset, indicating that the collaborative reasoning method based on hyperedge collaboration, dependency consistency aggregation and global decision-making can better adapt to complex long-term task scenarios.

[0064] In summary, this embodiment, taking a capital allocation task under coupled constraints as an example in a complex long-term task scenario of CLM, illustrates the complete implementation process of the present invention from task input acquisition, pattern routing, role allocation, collaborative execution, result encapsulation, dependency consistency aggregation to global decision output. Experimental results further demonstrate that the present invention can improve the matching degree between the collaboration method and task requirements, the reusability of intermediate results, and the consistency of the final inference results in complex task scenarios.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A reasoning method based on a high-order collaborative multi-agent large language model, characterized in that, Includes the following steps: S1. Obtain the input of the task to be processed, as well as the task organization hypergraph, hyperedge dependency graph, and mapping relationship between the hyperedge and task unit corresponding to the input of the task to be processed; S2. Extract task signals for the task units corresponding to each hyperedge, and determine the cooperation mode corresponding to each hyperedge based on the task signals. S3. Based on the determined cooperation mode, assign cooperation roles to the multiple agents participating in the cooperation within each hyperedge, and generate the corresponding cooperation protocol. S4. Based on the cooperation protocol, control multiple agents within each hyperedge to perform collaborative reasoning around the corresponding task unit to obtain the corresponding intermediate reasoning results; S5. Standardize and encapsulate each intermediate reasoning result to generate a structured result record; S6. According to the dependency order represented by the hyperedge dependency graph, perform dependency-consistent aggregation on the structured result record of the current hyperedge and the structured result record of its predecessor hyperedge to obtain the aggregation result record corresponding to each hyperedge. S7. Based on the aggregation result record, hyperedge dependency graph, hyperedge weight, and global state corresponding to each hyperedge, generate the final inference result corresponding to the input of the task to be processed.

2. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, In S1, the task organization hypergraph H is represented as: H=(V,E,w) Where V is the set of nodes, E is the set of hyperedges, and w is the hyperedge weight function; The hyperedge dependency graph G_E is represented as: G_E=(E,F) in, This represents the set of dependencies between hyperedges.

3. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, S2 includes: Extract the task signal vector xk for the task unit corresponding to the k-th hyperedge: xk=φ(τk,wk,χk) And determine the target cooperation mode mk based on the task signal vector: mk=π_mode(xk),mk∈M Where φ(·) represents the task vector encoding function, τk represents the task unit corresponding to the kth hyperedge, wk represents the weight of the kth hyperedge, χk represents the context state information corresponding to the kth hyperedge, M represents the preset collaboration mode set, including one or more of consensus collaboration mode, sequential collaboration mode, brainstorming collaboration mode, leadership decision-making collaboration mode and supervision and review collaboration mode, and π_mode(·) represents the collaboration mode routing.

4. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, S3 includes: Let Ak be the set of agents inside the k-th hyperedge: Ak={a(k,k1),a(k,k2),…,a(k,kn)} Where a(k,ki) represents the i-th agent of the k-th hyperedge; Let the set of characters be R: R = {r1, r2, ..., rP} Where ri represents different expert attributes, and P represents the total number of expert attributes; Roles are assigned to the agent set Ak based on the target cooperation pattern: π_role(k):Ak→R Where π_role(k) represents the role assignment result, and → represents the function mapping; And generate the collaboration protocol Π_k based on the target collaboration mode mk and the role assignment result π_role(k): Π_k=Γ(mk,π_role(k)) Here, Γ(·) represents the integration function of collaboration mode and role routing.

5. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, S4 includes: Initialize the local state set and shared global state of multiple agents within each hyperedge; And perform hyperedge-internal collaborative reasoning according to the collaboration protocol, satisfying the following relationship: yk,{s+_(k,i)}(i=1..n_k),S_glob+=F_e(k)^Π_k({s(k,i)}(i=1..nk),τk,S_glob) Where F_e(k)^Π_k represents the inference function of the k-th hyperedge under the cooperative protocol representation, yk represents the intermediate inference result generated by the k-th hyperedge, s(k,i) represents the local state of the i-th agent in the k-th hyperedge before the update, s+_(k,i) represents the local state of the i-th agent in the k-th hyperedge after the update, S_glob represents the global state before the update, S_glob+ represents the global state after the update, and τk represents the task unit corresponding to the k-th hyperedge.

6. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, S5 includes: The intermediate inference results generated by the k-th hyperedge are standardized and encapsulated to obtain the structured result rk record: rk=Φ_pack(yk,τk,S_glob+) Wherein, Φ_pack represents the standardized encapsulation function, yk represents the intermediate reasoning result generated by the kth hyperedge, and the structured result record rk includes one or more of the following: conclusion claim, satisfied constraint information, unsatisfied constraint information, evidence information, open question information, and source identification information.

7. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, S6 includes: Perform a topological sort on the hyperedge dependency graph to obtain the aggregate sequence; For the current hyperedge ek, determine its set of predecessor hyperedges Pred(ek): Pred(ek)={ej∈E|(ej,ek)∈F} Where ej represents the j-th hyperedge, and E represents the set of hyperedges. Represents the set of dependencies between hyperedges; And perform dependency consistent aggregation: r~k=Φ_agg(rk,{r~j|ej∈Pred(ek)},wk,S_glob+) Where Φ_agg represents the aggregation function, r~k represents the aggregation result record corresponding to the kth hyperedge, and wk represents the weight of the kth hyperedge.

8. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 7, characterized in that, S6 further includes: During dependency-consistent aggregation, only the result record corresponding to the predecessor of the current superedge is allowed to be introduced into the current aggregation process, and the result record of the successor superedge is not introduced.

9. The reasoning method based on a high-order collaborative multi-agent large language model according to claim 1, characterized in that, S7 includes: Get the aggregated result set R of all hyperedges: R = {r1, r2, ..., rM} r~k represents the aggregation result record corresponding to the kth hyperedge, and M represents the set of preset collaboration modes; The final result is generated based on the aggregation result set R~, the hyperedge dependency graph G_E, the hyperedge weights, and the global state S_glob+. : In the formula, Φ_dec(·) represents the overall encoding function, and wk represents the weight of the k-th hyperedge.

10. A multi-agent large language model reasoning system based on high-order collaboration, used to implement the multi-agent large language model reasoning method based on any one of claims 1-9, characterized in that, include: The task acquisition module is used to acquire the input of the task to be processed, as well as the task organization hypergraph, hyperedge dependency graph and mapping relationship corresponding to the input of the task to be processed; The pattern routing module is used to extract task signals for each hyperedge corresponding to the task unit and determine the target cooperation mode based on the task signals. The role assignment module is used to assign collaborative roles to multiple agents within each hyperedge based on the target collaboration mode. The collaboration protocol generation module is used to generate a collaboration protocol based on the target collaboration mode and the assignment of collaboration roles. The collaborative execution module is used to control multiple agents within each hyperedge to perform collaborative reasoning around the corresponding task unit and generate hyperedge-level intermediate results; The result encapsulation module is used to perform standardized encapsulation on the intermediate results generated by each hyperedge, and generate structured result records. The dependency aggregation module is used to perform dependency-consistent aggregation according to the topological order of the hyperedge dependency graph; as well as The global decision module is used to generate the final result by combining the aggregation results of each hyperedge, dependency structure, hyperedge weights, and global state.

Images