Method for collaborating multiple models based on instance awareness router

Abstract

The invention discloses a method for collaborating multiple models based on an instance aware router. The method comprises the following steps: firstly, constructing a sample-level performance distribution diagram of multiple language models in a target natural language understanding task, and carrying out statistics and analysis on prediction effects of different models on different input samples so as to describe performance differences of the models in different scenes of semantic complexity and reasoning requirements; and forming an'intensity domain 'mapping relation reflecting the model capability boundary. And secondly, on the basis of the sample-level performance distribution characteristics, designing an instance-aware routing instruction tuning mechanism, and automatically learning a corresponding relation between input sample characteristics and model performance through an instruction fine tuning training routing model, thereby realizing intelligent selection and distribution from the sample to an optimal large language model. And then, a multi-model cooperative reasoning mechanism is introduced, different large language models are dynamically called for reasoning aiming at samples with different complexity and feature types according to a decision result of the routing model, and multi-model output is fused through result aggregation strategies such as intersection or union, so that the accuracy and robustness of an overall reasoning result are improved. And finally, a unified input and output template and an anonymization processing strategy are introduced, and the multi-model cooperation process is normalized, so that the generalization ability and fairness of the method in a cross-task and cross-domain application scene are further enhanced.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically to the field of large language model (LLM) collaboration and inference optimization technology, and more specifically, to provide a multi-LLM collaborative inference method based on instance-aware routing. Background Technology

[0002] The rapid development of LLMs has enabled them to demonstrate outstanding capabilities in numerous Natural Language Processing (NLP) tasks, greatly advancing information understanding and generation technologies. As LLMs become core components of various application systems, maximizing their performance while overcoming their limitations has become a key research direction in the current technological field.

[0003] Despite the excellent general capabilities of LLMs, a core challenge in existing research and practical applications is that no single model can consistently maintain optimal performance across all types of inputs. Specifically, different LLMs often exhibit unique strengths and weaknesses when faced with diverse input samples. Our observed "LLM capability specialization phenomenon" indicates that even models with similar architectures or sizes often demonstrate unique processing capabilities for certain specific input samples; that is, only a particular model can correctly handle that sample. This phenomenon reveals the variability of LLM performance at the sample level.

[0004] To address this issue, researchers have proposed existing methods to enhance LLM performance in recent years, such as fine-tuning, prompting engineering, or hybrid expert (MoE) frameworks. These methods typically focus on: (1) customizing and enhancing a single model: These strategies are tailored to specific models and tasks, failing to fully utilize the increasingly rich and diverse available LLM resources. (2) lacking sample-level fine-grained control: Existing routing systems, whether non-predictive routing based on output selection or predictive routing based on pre-inference decisions, primarily operate at the model or system level, failing to consider sample-level routing granularity. In other words, they focus on selecting a model for the entire task or query, rather than customizing routing decisions based on the specific challenges of individual samples.

[0005] This neglect of sample-level performance differences makes it difficult for existing methods to achieve complementary collaboration among multiple LLMs, limiting the system's adaptability and overall performance when handling complex or heterogeneous inputs. In particular, there is a lack of a mechanism to: (1) accurately obtain the mapping relationship between samples and the optimal model: it is difficult to label or learn which LLM performs best on a given input sample in a low-cost and efficient manner. (2) explicitly model the capability distribution of LLMs in the input space: it is difficult to tell the router which model is good at handling which type of input feature.

[0006] In summary, there is an urgent need in this field to design a framework based on instance-aware routing to achieve dynamic, sample-level collaboration among LLMs, thereby solving the problem that a single LLM cannot maintain optimal performance on all samples and significantly improving the overall inference result quality of the system. Summary of the Invention

[0007] In view of this, the present invention provides an instance-aware routing (IA-Router) framework, which aims to solve specific challenges for specific input samples through complementary collaboration of LLMs and dynamically select the most suitable LLM for inference to improve overall performance. First, the present invention proposes and empirically analyzes the phenomenon of LLM capability specialization for the first time, that is, it finds that even LLMs of the same origin or similar size have complementarity at the sample level - each model has its own unique subset of specific samples that can be correctly processed. This discovery lays the theoretical foundation for solving the performance bottleneck of a single model and realizing sample-aware LLM routing. Second, the present invention constructs a new instance-aware routing framework (IA-Router), which supports multi-LLM collaborative inference through a functionally defined two-stage design, solving the challenge of difficulty in obtaining the mapping relationship between samples and their optimal models. The framework mainly includes: (1) Instance-aware routing instruction fine-tuning: Automatically constructing a dataset based on the optimal data model, and using this labeled data to fine-tune the routing model to achieve fine-grained instance-level model selection. (2) Multi-LLM Complementary Inference: Combining static feature filtering and dynamic routing decision-making, efficient inference is achieved, ensuring that the most suitable model is selected for each sample, reducing inference costs while maintaining accuracy. By performing intelligent routing among multiple LLMs, this invention can select the most suitable model for each input sample, thereby improving overall performance and the quality of inference results.

[0008] Definition 1: Relation Extraction Task. Given text and predefined set of relation types ,in express The number of lexical units in the text yes The number of relation types in the data. The relation extraction task aims to extract from... Extract the set of triples from the middle ,in The number of triples extracted. Represents head entity With tail entity The relationship between them.

[0009] Definition 2: Named Entity Recognition Task. Given text and predefined entity type set Named entity recognition task aims to... The mentioned fragment was detected in the middle. And assign a corresponding entity type to each extracted fragment. (e.g., location, etc.)

[0010] Definition 3: Aspect-level sentiment analysis task. Given text A set of predefined aspect categories (For example: ), and the set of emotional polarities { , , U}, where yes The number of lexical units in the text It refers to the number of aspect categories. Aspect-level sentiment analysis tasks aim to analyze the number of aspects. Extraction aspect - Emotional Triad Set ,in The number of aspect-emotional triads, express Aspect terms in Is with Related aspect categories, yes Emotional polarity.

[0011] Definition 4: Strength of a large language model. The strength of a large language model is defined as the set of samples that can be answered correctly independently and only by the model. This strength is independent of the model's size or overall capability.

[0012] The technical approach of this invention is as follows: First, an instance-aware router (IA-Router) framework is proposed, which achieves complementary collaboration among multiple LLMs through a two-stage design, including instance-aware routing instruction tuning and complementary inference among multiple LLMs. Second, in the first stage (instance-aware routing instruction tuning), a routing training data template is constructed based on a sample set. Instance-level analysis is performed on the input samples, and the most suitable LLM model is selected for instruction tuning, thereby training the routing model to dynamically select the optimal LLM. Then, in the second stage (complementary inference among multiple LLMs), the routing model is used to perform static feature extraction and routing decisions on the input text. Complementary outputs are generated by combining multiple LLMs, and the results are aggregated through intersection or union to improve overall performance. Finally, an optimized routing system is constructed, comprehensively evaluating sample-level variability and model expertise to generate the final structured results for the NLU task.

[0013] This invention takes multi-language model collaborative reasoning as an example, and the specific steps are as follows:

[0014] S1. Propose an instance-aware router (IA-Router) framework, which achieves complementary cooperation among multiple LLMs through a two-stage design, including instance-aware routing instruction tuning and complementary inference among multiple LLMs, to solve the performance limitations of a single LLM in terms of sample-level variability.

[0015] S2. In the first stage (instance-aware routing instruction tuning), a routing training data template is constructed based on the sample set (such as LLM mini, DeepSeek-R1, Qwen3-30B and InternLM2.5-7B). Instance-level analysis is performed on the input samples, and the most suitable LLM model is selected for instruction tuning, thereby training the routing model to dynamically select the optimal LLM.

[0016] S3. Based on S2, in the second stage (multi-LLM complementary inference), the routing model is used to perform static feature extraction and routing decision on the input text (such as NER and RE tasks), and multi-LLM (such as Qwen3-8B and MiniCPM-4B) is combined to generate complementary outputs, and the results are aggregated by intersection or union to improve the overall performance.

[0017] S4. Build an optimized routing system, perform system optimization and performance verification, comprehensively evaluate sample-level variability and model expertise, and generate the final structured results of the NLU task.

[0018] Further, step S1 includes:

[0019] S11. For natural language understanding tasks, construct a multi-task set including Named Entity Recognition (NER), Relation Extraction (RE), and Aspect-Based Sentiment Analysis (ABSA). The input text for each task is denoted as follows: Its corresponding output is a target structured set, such as an entity set, a triple set, or a sentiment pair set.

[0020] S12. Define language model strength, which measures the ability of a single language model to provide independent correct answers at the sample level. Let the sample set be... The language model set is Then the language model The strength is defined as:

[0021]

[0022] This set represents only those models A set of samples that make independent and correct predictions.

[0023] S13. Conduct experimental evaluations on models of different sizes from the same supplier and models of the same size from different suppliers, and calculate the following three types of sample indicators: Dominant sample: only one model answers correctly; Valid sample: at least one model answers correctly; Invalid sample: all models answer correctly.

[0024] S14. Based on the experimental results, the key conclusion is that language models from different sources or of different sizes possess unique "dominance domains" at the sample level, meaning that different models exhibit varying degrees of differentiation across different sample types. This observation provides a theoretical basis for subsequent sample-level routing mechanisms.

[0025] Further, step S2 includes:

[0026] S21, Constructing a system containing This is a foundational library for heterogeneous large language models, performing NER, RE, ABSA, and other task inference tasks under a unified few-shot prompt template. The template consists of four parts: role definition, task description, output format specification, and three context examples.

[0027] S22, For each sample Calculate the F1 score for the predictions across all models. If all models have the same score, the sample is considered invalid and removed. If there are differences, the model with the highest F1 score is marked as the optimal model for that sample. .

[0028] S23. Construct an instance-aware labeled dataset using "sample-optimal model" pairs as labels. The router model is trained on it. Each training sample contains the following structure: routing task definition; model selection criteria (balancing accuracy and recall); sample input text; candidate model output results and anonymization labels (such as "Model A", "Model B").

[0029] S24. The training objective is to minimize the route prediction error: The training objective is to minimize the route prediction error.

[0030]

[0031] This optimizes the router, enabling it to learn the mapping relationship between different sample features and the optimal model selection.

[0032] Further, step S3 includes:

[0033] S31. To reduce the computational cost of inference, a static feature pre-screening mechanism is introduced during the inference phase. Three types of static features are defined: text length... Information entropy Topic Vectors ;

[0034] S32. Based on the feature-performance distribution statistically analyzed during the training phase, calculate the threshold interval for each feature and select the candidate model set with the best Top-K performance. .

[0035] S33, For input samples Only The model performs inference, integrating the outputs of all candidate models into a routing input structure:

[0036]

[0037] The data is then input into the optimized IA-Router, which outputs the final selected model. and corresponding results .

[0038] S34. Optionally, through cross-validation, if the scores of multiple models are close to a threshold, the router can trigger a multi-model fusion strategy (such as weighted voting or confidence interval averaging).

[0039] Further, step S4 includes:

[0040] S41. Combine S2 and S3 to form a two-stage pipeline structure: Stage 1: Router instruction optimization (training stage); Stage 2: Multi-model complementary inference (inference stage).

[0041] S42. In the post-training evaluation phase, the invalid sample removal and static feature screening mechanism can reduce invalid computational overhead by about 75% while maintaining performance comparable to multi-model integration.

[0042] S43. By using a unified instruction template, the routing mechanism can achieve task-independent generalization capabilities, making it stable and applicable to different tasks such as NER, RE, and ABSA.

[0043] S44. Throughout the process, the model identity is anonymized, effectively eliminating semantic biases caused by factors such as model brand and scale, and achieving fair and objective routing decisions.

[0044] After adopting the above strategy, the positive effects of the present invention are:

[0045] (1) In view of the problem that there are differences in the "intensity domain" of different large language models in terms of reasoning ability and knowledge mastery, this invention proposes an innovative sample-level large language model routing framework. By automatically identifying the features of input samples, the optimal matching between models is achieved, which significantly improves the intelligent scheduling and dynamic adaptation capabilities of multi-model systems.

[0046] (2) In view of the problems of redundant computation and insufficient consistency in multi-model collaborative reasoning, this invention designs a two-stage structure that combines instruction tuning and multi-model complementary reasoning, which significantly reduces the model calling cost and improves the overall reasoning efficiency and accuracy while maintaining the generation quality.

[0047] (3) In view of the problem of insufficient model bias and generalization ability in cross-task transfer, this invention proposes a unified template and anonymization strategy. Through task-independent structured input, the model achieves fairness and broad generalization ability in multi-domain and multi-task scenarios, ensuring stable adaptation and consistent performance in tasks such as medical information extraction, question answering system, and text generation. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the algorithm flow of multi-LLM complementary cooperation for instance-aware routing proposed in this invention.

[0049] Figure 2 This is a schematic diagram of the overall structure of the sample-level large language model routing framework proposed in this invention. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0051] Please see Figure 1 , Figure 1This is a schematic diagram illustrating the "intensity domain" phenomenon of large language models, as proposed in this invention. In this diagram, the area of ​​the ellipse represents the number of correctly answered samples of the corresponding large language model; a larger area indicates stronger model performance. The position of the ellipse in the sample set reflects the distribution of correct answers from different large language models at the sample level. It is evident that different large language models exhibit significant performance differences and complementarity when facing different sample types, forming an "intensity domain" distribution phenomenon at the sample dimension. Based on this phenomenon, this invention proposes a sample-level model dynamic routing mechanism to achieve precise scheduling and optimal matching of multi-model systems.

[0052] Please see Figure 2 , Figure 2 This is a schematic diagram of the overall structure of the instance-aware router, a large language model routing framework proposed in this invention:

[0053] S1. Propose an instance-aware router (IA-Router) framework, which achieves complementary cooperation among multiple LLMs through a two-stage design, including instance-aware routing instruction tuning and complementary inference among multiple LLMs, to solve the performance limitations of a single LLM in terms of sample-level variability.

[0054] S11. For natural language understanding tasks, construct a multi-task set including Named Entity Recognition (NER), Relation Extraction (RE), and Aspect-Based Sentiment Analysis (ABSA). The input text for each task is denoted as follows: Its corresponding output is a target structured set, such as an entity set, a triple set, or a sentiment pair set.

[0055] S12. Define language model strength, which measures the ability of a single language model to provide independent correct answers at the sample level. Let the sample set be... The language model set is Then the language model The strength is defined as:

[0056]

[0057] This set represents only those models A set of samples that make independent and correct predictions.

[0058] S13. Conduct experimental evaluations on models of different sizes from the same supplier and models of the same size from different suppliers, and calculate the following three types of sample indicators: Dominant sample: only one model answers correctly; Valid sample: at least one model answers correctly; Invalid sample: all models answer correctly.

[0059] S14. Based on the experimental results, the key conclusion is that language models from different sources or of different sizes possess unique "dominance domains" at the sample level, meaning that different models exhibit varying degrees of differentiation across different sample types. This observation provides a theoretical basis for subsequent sample-level routing mechanisms.

[0060] S2. In the first stage (instance-aware routing instruction tuning), a routing training data template is constructed based on the sample set (such as LLM mini, DeepSeek-R1, Qwen3-30B and InternLM2.5-7B). Instance-level analysis is performed on the input samples, and the most suitable LLM model is selected for instruction tuning, thereby training the routing model to dynamically select the optimal LLM.

[0061] S21, Constructing a system containing This is a foundational library for heterogeneous large language models, performing NER, RE, ABSA, and other task inference tasks under a unified few-shot prompt template. The template consists of four parts: role definition, task description, output format specification, and three context examples.

[0062] S22, For each sample Calculate the F1 score for the predictions across all models. If all models have the same score, the sample is considered invalid and removed. If there are differences, the model with the highest F1 score is marked as the optimal model for that sample. .

[0063] S23. Construct an instance-aware labeled dataset using "sample-optimal model" pairs as labels. The router model is trained on it. Each training sample contains the following structure: routing task definition; model selection criteria (balancing accuracy and recall); sample input text; candidate model output results and anonymization labels (such as "Model A", "Model B").

[0064] S24. The training objective is to minimize the route prediction error: The training objective is to minimize the route prediction error.

[0065]

[0066] This optimizes the router, enabling it to learn the mapping relationship between different sample features and the optimal model selection.

[0067] S3. Based on S2, in the second stage (multi-LLM complementary inference), the routing model is used to perform static feature extraction and routing decision on the input text (such as NER and RE tasks), and multi-LLM (such as Qwen3-8B and MiniCPM-4B) is combined to generate complementary outputs, and the results are aggregated by intersection or union to improve the overall performance.

[0068] S31. To reduce the computational cost of inference, a static feature pre-screening mechanism is introduced during the inference phase. Three types of static features are defined: text length... Information entropy Topic Vectors ;

[0069] S32. Based on the feature-performance distribution statistically analyzed during the training phase, calculate the threshold interval for each feature and select the candidate model set with the best Top-K performance. .

[0070] S33, For input samples Only The model performs inference, integrating the outputs of all candidate models into a routing input structure:

[0071]

[0072] The data is then input into the optimized IA-Router, which outputs the final selected model. and corresponding results .

[0073] S34. Optionally, through cross-validation, if the scores of multiple models are close to a threshold, the router can trigger a multi-model fusion strategy (such as weighted voting or confidence interval averaging).

[0074] S4. Build an optimized routing system, perform system optimization and performance verification, comprehensively evaluate sample-level variability and model expertise, and generate the final structured results of the NLU task.

[0075] S41. Combine S2 and S3 to form a two-stage pipeline structure: Stage 1: Router instruction optimization (training stage); Stage 2: Multi-model complementary inference (inference stage).

[0076] S42. In the post-training evaluation phase, the invalid sample removal and static feature screening mechanism can reduce invalid computational overhead by about 75% while maintaining performance comparable to multi-model integration.

[0077] S43. By using a unified instruction template, the routing mechanism can achieve task-independent generalization capabilities, making it stable and applicable to different tasks such as NER, RE, and ABSA.

[0078] S44. Throughout the process, the model identity is anonymized, effectively eliminating semantic biases caused by factors such as model brand and scale, and achieving fair and objective routing decisions.

[0079] Specific embodiments of the present invention have been described above with reference to the accompanying drawings. However, those skilled in the art will understand that various modifications and substitutions can be made to the specific embodiments of the present invention without departing from the spirit and scope of the invention. All such modifications and substitutions fall within the scope defined by the claims of the present invention.

Claims

1. A multi-model collaboration method based on instance-aware routers, characterized in that... Includes the following steps: S1. Propose an instance-aware router (IA-Router) framework, which achieves complementary cooperation among multiple LLMs through a two-stage design, including instance-aware routing instruction tuning and complementary inference among multiple LLMs, to solve the performance limitations of a single LLM in terms of sample-level variability. S2. In the first stage (instance-aware routing instruction tuning), a routing training data template is constructed based on the sample set (such as LLM mini, DeepSeek-R1, Qwen3-30B and InternLM2.5-7B). Instance-level analysis is performed on the input samples, and the most suitable LLM model is selected for instruction tuning, thereby training the routing model to dynamically select the optimal LLM. S3. Based on S2, in the second stage (multi-LLM complementary inference), the routing model is used to perform static feature extraction and routing decision on the input text (such as NER and RE tasks), and multi-LLM (such as Qwen3-8B and MiniCPM-4B) is combined to generate complementary outputs, and the results are aggregated by intersection or union to improve the overall performance. S4. Build an optimized routing system, perform system optimization and performance verification, comprehensively evaluate sample-level variability and model expertise, and generate the final structured results of the NLU task.

2. The multi-LLM complementary cooperation method for instance-aware routing according to claim 1, characterized in that, Step S1 specifically includes: S11. For natural language understanding tasks, construct a multi-task set including Named Entity Recognition (NER), Relation Extraction (RE), and Aspect-Based Sentiment Analysis (ABSA). The input text for each task is denoted as follows: Its corresponding output is a target structured set, such as an entity set, a triple set, or a sentiment pair set. S12. Define language model strength, which measures the ability of a single language model to provide independent correct answers at the sample level. Let the sample set be... The language model set is Then the language model The strength is defined as: This set represents only those models A set of samples that make independent and correct predictions. S13. Conduct experimental evaluations on models of different sizes from the same supplier and models of the same size from different suppliers, and calculate the following three types of sample indicators: Dominant sample: only one model answers correctly; Valid sample: at least one model answers correctly; Invalid sample: all models answer correctly. S14. Based on the experimental results, the key conclusion is that language models from different sources or of different sizes have unique "dominance domains" at the sample level, meaning that different models exhibit varying degrees of differentiation across different sample types. This observation provides a theoretical basis for subsequent sample-level routing mechanisms.

3. The multi-LLM cooperation method for an instance-aware router according to claim 1, characterized in that, Step S2 specifically includes: S21, Constructing a system containing This is a foundational library for heterogeneous large language models, performing NER, RE, ABSA, and other task inference tasks under a unified few-shot prompt template. The template consists of four parts: role definition, task description, output format specification, and three context examples. S22, For each sample Calculate the F1 score for the predictions across all models. If all models have the same score, the sample is considered invalid and removed. If there are differences, the model with the highest F1 score is marked as the optimal model for that sample. . S23. Construct an instance-aware labeled dataset using "sample-optimal model" pairs as labels. The router model is trained on it. Each training sample contains the following structure: routing task definition; model selection criteria (balancing accuracy and recall); sample input text; candidate model output results and anonymization labels (such as "Model A", "Model B"). S24. The training objective is to minimize the route prediction error: The training objective is to minimize the route prediction error. This optimizes the router, enabling it to learn the mapping relationship between different sample features and the optimal model selection.

4. The multi-LLM complementary cooperation method for instance-aware routing according to claim 1, characterized in that, Step S3 specifically includes: S31. To reduce the computational cost of inference, a static feature pre-screening mechanism is introduced during the inference phase. Three types of static features are defined: text length... Information entropy Topic Vectors ; S32. Based on the feature-performance distribution statistically analyzed during the training phase, calculate the threshold interval for each feature and select the candidate model set with the best Top-K performance. . S33, For input samples Only The model performs inference, integrating the outputs of all candidate models into a routing input structure: The data is then input into the optimized IA-Router, which outputs the final selected model. and corresponding results . S34. Optionally, through cross-validation, if the scores of multiple models are close to a threshold, the router can trigger a multi-model fusion strategy (such as weighted voting or confidence interval averaging).

5. The multi-LLM cooperation method for an instance-aware router according to claim 1, characterized in that, Step S4 specifically includes: S41. Combine S2 and S3 to form a two-stage pipeline structure: Stage 1: Router instruction optimization (training stage); Stage 2: Multi-model complementary inference (inference stage). S42. In the post-training evaluation phase, the invalid sample removal and static feature screening mechanism can reduce invalid computational overhead by about 75% while maintaining performance comparable to multi-model integration. S43. By using a unified instruction template, the routing mechanism can achieve task-independent generalization capabilities, making it stable and applicable to different tasks such as NER, RE, and ABSA. S44. Throughout the process, the model identity is anonymized, effectively eliminating semantic biases caused by factors such as model brand and scale, and achieving fair and objective routing decisions.

Images