A Real-Time Adaptive Guidance Method and System for Singing Based on a Multimodal Large Model

Abstract

The application discloses a real-time singing self-adaptive guidance method and system based on a multi-modal large model, which comprises the following steps: S1. processing the learner's singing audio, extracting acoustic evidence including the fundamental frequency trajectory, pitch stability index and vocal range boundary, and generating an initial evidence summary; updating the initial evidence summary based on the structured evaluation results; S2. inputting the learner's singing audio into a multi-modal large model, an audio large model, an audio-text multi-modal large model or a fine-tuned model thereof, combining a pre-defined output format, and generating a structured evaluation vector containing dimension score values and explanation texts; S3. providing the training target specification, the current evidence summary, the bounded interaction history and the current learner input as inputs to a large language model to generate stage control instructions in real time; and S4. outputting practice prompts, immediate feedback, review instructions or stage reports matching the current training stage to the learner according to the stage control instructions.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and multimodal learning, specifically relating to a real-time singing adaptive guidance method and system based on a multimodal large model. Background Technology

[0002] Singing practice is an important form of training in music education, typically involving multiple training goals simultaneously, including pitch accuracy, vocal technique, emotional expression, and breath control. The purpose of singing practice is to help learners gradually develop comprehensive control over pitch, breath, and phrasing through repeated, multi-dimensional training. Because singing can be performed without external instruments, its low barrier to entry makes it the preferred way for most beginners to engage with music, thus making singing coaching techniques widely applicable. Through systematic singing coaching, learners can promptly identify their weaknesses in different ability dimensions and correct them in a targeted manner, avoiding inefficient and repetitive practice. Singing coaching techniques also help teachers monitor learners' performance at each training stage, identify practice bottlenecks, and adjust teaching arrangements accordingly. By providing personalized practice content and feedback guidance to learners of different ability levels, singing coaching techniques help lower the professional barrier to music learning, enabling more beginners lacking professional guidance to receive effective training support.

[0003] The shortcomings and deficiencies of existing singing coaching methods: 1) The lack of modeling and process organization for the stages of singing training makes it difficult to match the dynamic evolution of goals and feedback.

[0004] Existing singing assessment and coaching programs often treat singing training as a single-stage task, mainly focusing on one-time assessments or single-round feedback, without explicitly depicting the fact that practice goals and feedback needs change as the learning process progresses. Due to the lack of a phased organization and granular switching mechanism, the system struggles to continuously provide coherent guidance that matches the learner's current stage, thus affecting the feasibility and effectiveness of multi-stage training for beginners.

[0005] 2) It focuses on result evaluation or isolated skill tips, lacks a closed-loop coaching system that runs through the practice process, and the assessment explanations are difficult to reliably support subsequent guidance.

[0006] While existing systems can output multi-dimensional scores or descriptive comments, they mostly remain at the level of result evaluation or single-point skill feedback, lacking a continuous coaching loop that includes breaking down practice, repeated practice, and phased progress and regression correction. Even when large language models or structured outputs are introduced, they are often only used for single-round comment generation, failing to effectively combine model decisions with explicit acoustic evidence and phased control mechanisms, thus making it difficult to stably support subsequent feedback generation and process control.

[0007] 3) The existing multi-dimensional assessment scheme is disconnected from the teaching scheduling process, and the assessment results cannot directly drive phased teaching decisions.

[0008] Existing multi-dimensional singing assessment schemes (including multi-dimensional scoring systems based on machine learning and descriptive assessment frameworks for singing based on multimodal large models) all end with score output or comment generation. They lack machine-consumable decision interfaces that directly map assessment conclusions into instructions for advancement, repetition, or regression. There is no formal feedback path between the assessment module and the teaching scheduling process, which cannot support adaptive closed-loop guidance. Summary of the Invention

[0009] To address the problems of existing technologies, this invention provides a real-time adaptive singing guidance method and system based on a multimodal large model. Using learning objectives as constraints and acoustic evidence and structured evaluation results as the basis, it combines audio modality and text modality information to select practice content in real time and adaptively control the transition between practice stages during multiple rounds of interaction, thereby achieving continuous guidance for learners. It can be implemented as a software system running locally on a terminal device, as a server-side service, or a collaborative deployment of both.

[0010] To achieve the above objectives, the present invention provides the following solution: A real-time adaptive singing guidance method based on a multimodal large model, the method comprising: S1. Process the learner's singing audio, extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, and generate an initial evidence summary; after obtaining the structured evaluation results, update the initial evidence summary based on the structured evaluation results; S2. Input the learner's singing audio into the multimodal large model, audio large model, audio-text multimodal large model or its fine-tuned model, and combine it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text; S3. The training objective reduction, current evidence summary, bounded interaction history and current learner input are provided as input to the large language model to generate stage control instructions in real time. S4. Based on the stage control instructions, output practice prompts, immediate feedback, repetition instructions, or stage reports that match the current training stage to the learner.

[0011] Preferably, in step S1, the method for processing the learner's singing audio to extract acoustic evidence including fundamental frequency trajectory, pitch stability index, and vocal range boundaries, and generating an initial evidence summary includes: in, This represents the set of acoustic evidence signals or its encoded acoustic evidence vector; the two zeros represent default placeholders corresponding to the score vector and explanatory information, respectively, before a structured evaluation result is obtained. After obtaining the structured evaluation result, the score vector and explanatory text are further aggregated with the initial evidence summary to form an updated evidence summary.

[0012] Preferably, in step S2, the method for inputting the learner's singing audio into a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof, and combining it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text, includes: Input the audio clip, evaluation prompt template, and optional text context into the model to obtain the raw output: ; in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context; Represents the inference function for a multimodal large model, an audio large model, or an audio-text multimodal large model; This represents the raw text output by the model, containing multiple predefined token segments, each token segment corresponding to a set of dimensions. One dimension of singing ability; Perform deterministic analysis on the model's original output: ; in, This represents a deterministic analytic function that decomposes the original output into a score vector and explanatory text by locating predefined marked segments. The score vector and explanatory text are as follows: ; ; in, Representing dimensions The rating value is an integer ranging from 1 to 5; Representing dimensions The explanatory text describes the specific performance of this dimension and provides suggestions for improvement.

[0013] Preferably, in step S3, the method of providing the training objective reduction, current evidence summary, bounded interaction history, and current learner input as input to the large language model, and generating stage control instructions in real time, includes two stages: The first stage is response generation and preliminary routing: receiving the stage control cue template, bounded history, training objective reduction, and current learner input, while simultaneously generating learner-oriented text responses and preliminary stage control actions. ; in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This represents the input from the current learner; This represents the generated text response; Indicates the initial routing action; The second phase is evidence-driven route update: the routing decision is re-evaluated based on the updated evidence summary and the latest bounded history. ; in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . ,in, To move to the next stage, To repeat practice at the current stage, This is to revert to the previous stage, thereby achieving adaptive stage transition.

[0014] The present invention also provides a real-time singing adaptive guidance system based on a multimodal large model. The system is used to implement the aforementioned method and includes: a target module, a scheduling module, a data acquisition module, an evidence module, an evaluation module, a profiling module, and a presentation module. The target module is used to acquire and maintain the learner's training target specification, which includes at least target practice items, target ability thresholds, and target difficulty. The acquisition module is used to acquire learner audio input and provide audio data to the evidence module and the evaluation module; The evidence module is used to process learners' singing audio, extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, generate an initial evidence summary, and update the initial evidence summary after obtaining structured evaluation results. The evaluation module is used to input learners' singing audio into a multimodal large model, an audio large model, an audio-text multimodal large model or its fine-tuned model, and combine it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text. The profile module is used to store learner profiles and interaction history, and to provide a bounded interaction history for the scheduling module; The scheduling module is used to provide the training objective reduction, current evidence summary, bounded interaction history and current learner input as input to the large language model, and generate stage control instructions in real time. The presentation module is used to output practice prompts, real-time feedback, repetition instructions, or stage reports that match the current training stage to the learner, based on the stage control instructions.

[0015] Preferably, in the evidence module, the process of processing the learner's singing audio to extract acoustic evidence including fundamental frequency trajectory, pitch stability index, and vocal range boundary, and generating an initial evidence summary includes: in, This represents the set of acoustic evidence signals or its encoded acoustic evidence vector; the two zeros represent default placeholders corresponding to the score vector and explanatory information, respectively, before a structured evaluation result is obtained. After obtaining the structured evaluation result, the score vector and explanatory text are further aggregated with the initial evidence summary to form an updated evidence summary.

[0016] Preferably, in the evaluation module, the process of inputting the learner's singing audio into a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof, and combining it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text includes: Input the audio clip, evaluation prompt template, and optional text context into the model to obtain the raw output: ; in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context; Represents the inference function for a multimodal large model, an audio large model, or an audio-text multimodal large model; This represents the raw text output by the model, containing multiple predefined token segments, each token segment corresponding to a set of dimensions. One dimension of singing ability; Perform deterministic analysis on the model's original output: ; in, This represents a deterministic analytic function that decomposes the original output into a score vector and explanatory text by locating predefined marked segments. The score vector and explanatory text are as follows: ; ; in, Representing dimensions The rating value is an integer ranging from 1 to 5; Representing dimensions The explanatory text describes the specific performance of this dimension and provides suggestions for improvement.

[0017] Preferably, in the scheduling module, the process of providing the training objective reduction, current evidence summary, bounded interaction history, and current learner input as input to the large language model, and generating stage control instructions in real time, includes two stages: The first stage is response generation and preliminary routing: receiving the stage control cue template, bounded history, training objective reduction, and current learner input, while simultaneously generating learner-oriented text responses and preliminary stage control actions. ; in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This represents the input from the current learner; This represents the generated text response; Indicates the initial routing action; The second phase is evidence-driven route update: the routing decision is re-evaluated based on the updated evidence summary and the latest bounded history. ; in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . ,in, To move to the next stage, To repeat practice at the current stage, This is to revert to the previous stage, thereby achieving adaptive stage transition.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention can utilize a large language model to jointly model learning objectives, interaction history, and evidence summaries, and generate in real time advance, repetition, or regress instructions that match the current training stage, thereby improving the adaptability and coherence of singing instruction.

[0019] 2. This invention designs the multi-dimensional evaluation output as a structured evaluation vector with a deterministic parsing interface, and incorporates the scoring vector and explanatory text into the evidence summary update process, so that the evaluation module output can be directly transformed into a stage decision signal that can be consumed by the scheduling module, thereby embedding the multi-dimensional evaluation capability into the staged scheduling closed loop of singing training.

[0020] 3. This invention combines the ability to generate multimodal large models with explicit acoustic evidence constraints, so that real-time feedback generation, stage scheduling and subsequent task orchestration are all based on verifiable acoustic evidence, thereby improving the reliability, interpretability and personalization of the guidance results. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the system structure of the singing training and tutoring system in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the phased singing training and guidance method in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the acoustic evidence generation method in an embodiment of the present invention; The modules are as follows: 1—Learner singing audio; 2—Audio stream; 3—Audio segment; 4—Initial evidence summary; 5—Vocal range field; 6—Scoring vector; 7—Structured evaluation result; 8—Training objective specification; 9—Objective archive; 10—Bounded history; 11—Text response and stage instructions; 12—Control instructions; 13—Practice prompts; 100—Objective module; 120—Scheduling module; 130—Acquisition module; 140—Evidence module; 150—Evaluation module; 160—Profiling module; 170—Presentation module; 180—Template library module. Detailed Implementation

[0023] 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.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 This invention provides a real-time adaptive singing guidance method based on a multimodal large model, the method comprising: S1. Acoustic evidence generation: The learner's singing audio 1 is processed to extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, and an initial evidence summary 4 is generated; after obtaining the structured evaluation result 7, the initial evidence summary 4 is updated based on the structured evaluation result 7. S2. Structured Multidimensional Assessment: Input the learner's singing audio 1 into a multimodal large model, an audio large model, an audio-text multimodal large model or its fine-tuned model, and combine it with a predefined output format to generate a structured assessment vector containing score values ​​for each dimension and explanatory text.

[0026] S3. LLM-based stage scheduling decision: The training objective reduction, current evidence summary, bounded interaction history, and current learner input are provided as input to the large language model, and values ​​are generated in real time. (Progress to the next stage) (Repeat the practice at the current stage) or (Revert to the previous stage) One of the stage control instructions 12; S4. Task arrangement and feedback presentation: Based on the stage control instructions 12, output practice prompts 13, instant feedback, repetition instructions or stage reports that match the current training stage to the learner.

[0027] In this embodiment, the method for processing the learner's singing audio 1 in step S1 to extract acoustic evidence including fundamental frequency trajectory, pitch stability index, and vocal range boundary to generate initial evidence summary 4 includes: in, Indicates the first The acoustic evidence vector is output and encoded by the evidence module 140; the two zeros represent the default placeholders corresponding to the score vector 6 and the explanatory information when the structured evaluation result 7 has not yet been obtained. After obtaining the structured evaluation result 7, the score vector 6 and the explanatory text are further aggregated with the initial evidence summary 4 to form an updated evidence summary.

[0028] In this embodiment, step S2 involves inputting the learner's singing audio 1 into a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof, and combining it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text. Input the audio clip 3, the evaluation prompt template, and the optional text context into the model to obtain the raw output: ; in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context; Represents the inference function for a multimodal large model, an audio large model, or an audio-text multimodal large model; This represents the raw text output by the model, containing multiple predefined token segments, each token segment corresponding to a set of dimensions. One dimension of singing ability; Perform deterministic analysis on the model's original output: ; in, This represents a deterministic analytic function that decomposes the original output into a score vector 6 and explanatory text by locating predefined marked segments. The score vector 6 and the explanatory text are as follows: ; ; in, Representing dimensions The rating value is an integer ranging from 1 to 5; Representing dimensions The explanatory text describes the specific performance of this dimension and provides suggestions for improvement.

[0029] In this embodiment, step S3, which involves providing the training objective reduction, current evidence summary, bounded interaction history, and current learner input as input to the large language model, and generating the stage control instruction 12 in real time, includes two stages: The first stage is response generation and preliminary routing: receiving the stage control prompt template, bounded history 10, training objective reduction, and current learner input, while simultaneously generating learner-oriented text responses and preliminary stage control actions: ; in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This represents the input from the current learner; This represents the generated text response; Indicates the initial routing action; The second phase is evidence-driven route update: the route decision is re-evaluated based on the updated evidence summary and the latest bounded history. ; in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . ,in, To move to the next stage, To repeat practice at the current stage, This is to revert to the previous stage, thereby achieving adaptive stage transition.

[0030] Example 2 The present invention also provides a real-time singing adaptive guidance system based on a multimodal large model. The system is used to implement the method described in Embodiment 1. The system includes: a target module 100, a scheduling module 120, a data acquisition module 130, an evidence module 140, an evaluation module 150, a profiling module 160, and a presentation module 170. The target module 100 is used to acquire and maintain the learner's training target specification, which includes at least target practice items, target ability thresholds, and target difficulty. The acquisition module 130 is used to acquire the learner's audio input and provide audio data to the evidence module 140 and the evaluation module 150; The evidence module 140 is used to process the learner's singing audio 1, extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, generate initial evidence summary 4, and update the initial evidence summary 4 after obtaining structured evaluation result 7. The evaluation module 150 is used to input the learner's singing audio 1 into a multimodal large model, an audio large model, an audio-text multimodal large model or its fine-tuned model, and combine it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text. The profiling module 160 is used to store learner profiles and interaction history, and to provide bounded interaction history for the scheduling module 120; The scheduling module 120 is used to provide the training objective reduction, current evidence summary, bounded interaction history and current learner input as input to the large language model, and generate stage control instructions 12 in real time. The presentation module 170 is used to output practice prompts 13, instant feedback, repetition instructions, or stage reports that match the current training stage to the learner according to the stage control instructions 12.

[0031] In this embodiment, the process of processing the learner's singing audio 1 in the evidence module 140 to extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, and generating initial evidence summary 4 includes: in, Indicates the first The acoustic evidence vector is output and encoded by the evidence module 140; the two 0s represent the default placeholders corresponding to the scoring vector 6 and the explanatory information when the structured evaluation result 7 has not yet been obtained.

[0032] In this embodiment, the process of inputting the learner's singing audio 1 into a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof, and combining it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text, in the evaluation module 150, includes: Input the audio clip 3, the evaluation prompt template, and the optional text context into the model to obtain the raw output: ; in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context; Represents the inference function for a multimodal large model, an audio large model, or an audio-text multimodal large model; This represents the raw text output by the model, containing multiple predefined token segments, each token segment corresponding to a set of dimensions. One dimension of singing ability; Perform deterministic analysis on the model's original output: ; in, This represents a deterministic analytic function that decomposes the original output into a score vector 6 and explanatory text by locating predefined marked segments. The score vector 6 and the explanatory text are as follows: ; ; in, Representing dimensions The rating value is an integer ranging from 1 to 5; Representing dimensions The explanatory text describes the specific performance of this dimension and provides suggestions for improvement.

[0033] In this embodiment, the scheduling module 120 provides the training objective reduction, current evidence summary, bounded interaction history, and current learner input as input to the large language model, and the process of generating the stage control instruction 12 in real time includes two stages: The first stage is response generation and preliminary routing: receiving the stage control prompt template, bounded history 10, training objective reduction, and current learner input, while simultaneously generating learner-oriented text responses and preliminary stage control actions: ; in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This represents the input from the current learner; This represents the generated text response; Indicates the initial routing action; The second phase is evidence-driven route update: the route decision is re-evaluated based on the updated evidence summary and the latest bounded history. ; in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . ,in, To move to the next stage, To repeat practice at the current stage, This is to revert to the previous stage, thereby achieving adaptive stage transition.

[0034] Example 3 The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0035] like Figure 1 As shown, the tutoring system of the present invention includes at least the following functional modules: First, the target module 100 obtains the training target specification, and the scheduling module 120 drives the acquisition module 130 to acquire the learner's singing audio 1 in real time; then, the evidence module 140 generates acoustic evidence and initial evidence summary 4, and when needed, the evaluation module 150 generates structured evaluation results 7 based on a multimodal large model; the scheduling module 120 then generates stage instructions in real time based on the current evidence summary and bounded history 10, and the presentation module 170 outputs practice prompts 13, instant feedback or stage reports accordingly, thereby forming a real-time singing adaptive guidance closed loop.

[0036] 1. Target Module 100: Used to acquire and maintain the learner's training target specification. The training target specification includes at least target practice items, target ability thresholds, and target difficulty. Target practice items can be any of the following: scale practice, short phrase practice, or whole piece practice. The system writes the training target specification into the profile module 160 and provides a target query interface to the scheduling module 120.

[0037] 2. Scheduling Module 120: Used for real-time stage decision-making and adaptive task orchestration based on the large language model. The scheduling module 120 receives the training objective reduction, the bounded history 10 up to the previous round, the updated evidence summary for the current round, and the current learner input, and outputs stage control instructions 12 and the next practice content; wherein, the current learner input includes at least one of the following: the speech-to-text of the current round, a practice confirmation instruction, a continue instruction, and a retry instruction; the stage control instructions 12 include at least one of the following: advance, repeat, and revert. The scheduling module 120 is also used to maintain the bounded history 10 and map the stage control actions output by the large language model to executable system control actions to update the stage state and current practice index in the profiling module 160.

[0038] 3. Acquisition Module 130: Used to acquire learner audio input and output raw waveform data and sampling rate information, providing audio data to Evidence Module 140 and Evaluation Module 150. Acquisition Module 130 can be connected to a microphone, external sound card, or mobile terminal recording interface.

[0039] 4. Evidence Module 140: Used to generate comparable acoustic evidence from audio, outputting at least signals such as fundamental frequency trajectory, voicing confidence, stability index, and vocal range boundaries, and forming an initial evidence summary 4 for use in stage control. Evidence Module 140 simultaneously writes the vocal range boundary field and the recent stability field to the profiling module 160; wherein the vocal range boundary field includes at least the learner's lowest and highest stable pitches, and the recent stability field includes at least the fundamental frequency fluctuation or normalized stability score of the most recent several rounds of practice, for use in subsequent personalized practice generation and evidence summary updates.

[0040] 5. Evaluation Module 150: Used to perform multi-dimensional real-time evaluation of learner audio based on a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof. Evaluation Module 150 receives audio segment 3, evaluation prompt templates, and optional text context, and outputs structured evaluation results 7. The structured evaluation results 7 can be parsed into scores and explanations for each singing ability dimension, and used to generate immediate feedback and update evidence summaries for stage control by the scheduling module 120.

[0041] 6. Profile Module 160: Used to store learner profiles and interaction history. The profile includes at least the following fields: vocal range boundary field, recent stability field, stage status field, generated practice prompts field, and historical stage evidence sequence field; and provides a bounded memory window for the scheduling module 120. The most recent few rounds of interaction history are extracted to control the length of the inference context.

[0042] 7. Presentation Module 170: Used to output practice content prompts, reference audio, instant feedback, repetition instructions, and progress reports to learners. Presentation Module 170 receives control instructions 12 from Scheduling Module 120 and structured evaluation results 7 from Evaluation Module 150.

[0043] 8. Template Library Module 180: Used to store stage control prompt templates, structured evaluation prompt templates, feedback generation templates, and parsing rules, and is available for use by the scheduling module 120, evaluation module 150, and presentation module 170; wherein, the stage control prompt template is used to generate stage control instructions 12, the structured evaluation prompt template is used to generate structured evaluation results 7, the feedback generation template is used to generate practice prompts 13, immediate feedback, repetition instructions, or stage reports, and the parsing rules are used to parse the structured evaluation results 7 into a scoring vector 6 and explanatory text.

[0044] The data flow between the above modules includes at least: training objective reduction flow, audio flow 2, text prompt flow, acoustic evidence flow, structured evaluation flow, evidence summary flow, and stage control instruction flow.

[0045] The key data structure is defined as follows: To ensure decision consistency across multiple rounds of interaction, this invention introduces the following data structure (which can be stored in the profiling module 160): Dimension set: Represents the set of dimensions of the singing ability being evaluated, defined as: in, This represents a set of dimensions related to singing ability. Corresponding to the dimension of vocal technology, Corresponding to the dimension of emotional expression, Corresponding to the timbre and sound quality dimensions, This corresponds to the breathing control dimension. The number of elements in the dimension set is not limited to four and can be expanded according to actual training needs.

[0046] Target threshold vector: Represents the expected level threshold for each dimension, defined as: in, Representing dimensions The corresponding expected level threshold is set within a preset scoring range (e.g., an integer or real number from 1 to 5). When the learner's actual scores in each dimension are not lower than the corresponding threshold, the training objective for that stage is considered achieved.

[0047] Interaction history: Represents the history up to the [number]th [number]. The exercise items and the sequence of evidence for each stage are defined as follows: in, Indicates the first The exercise items are labeled on the wheel; Indicates the first The corresponding stage evidence record for each round includes at least one or more of the following: initial evidence summary 4, updated evidence summary, scoring vector 6, and explanatory text. The interaction history is used by the scheduling module 120 to construct a bounded history 10 and supports adaptive stage control based on stage evidence.

[0048] Bounded memory window: Represents the length of the most recent rounds of historical interactions retained when scheduling module 120 calls the large language model. It is used to control the context size and is defined as follows: in, This is the preset memory window length; Indicates the first Only the most recent one is retained during the round. A bounded subset of the history of the wheel interaction record.

[0049] Evidence summary: indicating the first A wheel is used for the set of decisionable signals in phase control. In one implementation, the generation of evidence summaries includes three steps: interpreting text parsing, vectorizing acoustic evidence, and feature aggregation, defined as: in, The acoustic evidence vector output and encoded by the evidence module 140 includes at least the statistical characteristics, stability index and frequency range boundary corresponding to the fundamental frequency trajectory. Represents the score vector for structured assessment; Explanatory text representing each dimension; This represents the explanatory text parsing function, used to parse explanatory text into machine-readable explanatory feature vectors. ; This represents a vector concatenation operation. Therefore, the evidence summary... At the same time, acoustic evidence, scoring results, and explanatory information are retained and can serve as a core component of the current round's stage evidence record for use by the scheduling module 120 for stage control.

[0050] When only acoustic evidence is available and structured assessment results are not yet obtained, Then the initial evidence summary 4 is defined as: System prompt : Represents the large language model prompt template used for decision-making in the output stage of the constraint scheduling module 120.

[0051] Assessment Tips : This indicates a prompt template for the structured evaluation result 7 output by the constraint evaluation module 150.

[0052] Routing action: indicates the first The stage control actions output by the scheduling module 120 are defined as follows: in, This indicates progress to the next stage. This indicates that you should repeat the practice at the current stage. This indicates a regression to the previous stage for supplementary training.

[0053] Among them, an implementation example of acoustic evidence generation (evidence module 140) like Figure 3 As shown, evidence module 140 can generate comparable acoustic evidence in the following manner: 1. Audio Shaping: The original waveform is sequentially subjected to high-pass filtering, band-pass filtering, and peak normalization to suppress low-frequency drift and non-target frequency noise, and to limit peak amplitude to avoid clipping. The audio shaping process is defined as follows: in, Indicates the sampling rate The original singing waveform; Indicates the cutoff frequency as High-pass filters (such as Butterworth filters) are used to suppress low-frequency drift; Indicates passband is A bandpass filter is used to preserve the frequency band where the fundamental frequency of singing is located and suppress out-of-band noise; This indicates a peak normalization operation, which scales the peak amplitude of the signal to the upper limit. To avoid clipping distortion in subsequent processing; This represents the shaped audio signal. The order of the filter... These are configurable parameters.

[0054] 2. Fundamental Frequency Estimation: Calculate the frame-by-frame fundamental frequency trajectory for the shaped audio and output a frame-by-frame voicing confidence sequence for filtering reliable audio frames. The fundamental frequency estimation process is defined as follows: in: Represents the fundamental frequency estimation function; Indicates the first The shaped audio collected by the wheel; Indicates the sampling rate; This represents the set of hyperparameters of the estimated function (including minimum frequency, maximum frequency, etc.). Indicates the first The frame index after the audio is framed; Indicates the first Estimated fundamental frequency of the frame (in Hertz); Indicates the first The confidence level of a frame's voicing is such that the closer the value is to 1, the higher the confidence level that the frame is a sound frame and that the fundamental frequency estimation is reliable.

[0055] 3. Stability Indicator: In the absence of a reference pitch sequence, stability confidence can be calculated on a set of spoken frames. First, a subset of the fundamental frequencies of the spoken frames is defined: in, A preset confidence threshold for turbidity / clarity is used to retain only the fundamental frequency values ​​of audio frames with a confidence level higher than this threshold. Based on this, the stability index is defined as follows: in, and Let represent the arithmetic mean and standard deviation of the subset, respectively. This indicates that the result will be truncated to... Range operations; Indicates the first The pitch stability confidence score is calculated based on the number of notes, with a value closer to 1 indicating greater pitch stability. This stability index can be calculated without referring to a pitch sequence and is suitable for free vocalization or range testing scenarios.

[0056] 4. Vocal Range Boundaries: Robust quantiles are calculated for the fundamental frequency set of audio frames with voicing confidence levels above a threshold, serving as the lower and upper bounds of the vocal range, respectively, and written into the image module 160. The vocal range boundary estimation is defined as: in, Indicates taking the first position of a given set. Quantile operations; and These are the lower bound quantile and the upper bound quantile (e.g. , ), used to exclude extreme frames to improve robustness; and These represent the estimated lower and upper bounds of the pitch range, respectively (unit: Hertz).

[0057] Here is an example of the implementation of personalized practice prompts (scheduling module 120). When the image module 160 has stored the vocal range boundaries, the scheduling module 120 can construct basic practice prompts based on this. The generation process of the practice prompts 13 is defined as follows: First, calculate the midpoint frequency of the vocal range: in, and These are the lower and upper limits of the vocal range stored in the image module 160, respectively. The geometric mean of the two values ​​is used as the middle pitch of the practice sequence. The reason for using the geometric mean instead of the arithmetic mean is that the human ear's perception of pitch is approximately logarithmically distributed along the frequency axis, and the geometric mean can better reflect the perceived "middle position".

[0058] Secondly, construct the target frequency sequence: in, This represents a practice pitch sequence consisting of five target frequency values, forming a symmetrical "low-middle-high-middle-low" structure to cover the learner's current available vocal range. The sequence length and scale pattern can be adjusted according to training needs.

[0059] Finally, a reference audio file is generated: in, This represents a reference tone synthesis function (e.g., a piano tone synthesizer). This represents a set of synthesis parameters, including configurable items such as timbre type, velocity, and duration of each note; The generated reference audio is written to the image module 160 and then output by the presentation module 170 for learners to imitate.

[0060] Among them, an implementation example of structured evaluation of multimodal large models (evaluation module 150) is provided. The evaluation module 150 can employ a multimodal large model, an audio large model, an audio-text multimodal large model, a singing evaluation model fine-tuned by instructions, or a combination thereof, to output multi-dimensional scores and explanatory text for the input audio and related text context. The evaluation process can be divided into two steps: model generation and structured parsing, defined as follows: First, input the audio clip 3, the evaluation prompt template, and the optional text context into the model to obtain the raw output: in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context.

[0061] Then, for the original output Parsing according to predefined label rules yields structured evaluation result 7: in, This represents the score vector for each dimension of singing ability. This represents the set of explanatory texts corresponding to each dimension of singing ability.

[0062] In one implementation, let the total number of evaluation dimensions be... Then the rating vector 6 can be represented as: The explanatory text set can be represented as: in, Indicates the first The score for each dimension of singing ability. Indicates the first Explanatory text for each dimension of singing ability. The scoring vector... and Explanatory Text Together, they constitute a structured evaluation result that can be used for subsequent evidence summary updates and phase control invocation.

[0063] In one implementation, the evaluation module 150 only receives audio segments. and assessment prompt template In another implementation, the evaluation module 150 may also receive the fundamental frequency trajectory output by the evidence module 140. Stability indicators or range field As This is a component of the model evaluation, designed to enhance the basis of the evaluation and reduce unsubstantiated feedback.

[0064] Among them, an example of real-time adaptive stage control based on a large language model (scheduling module 120) is provided. The scheduling module 120 can implement real-time adaptive stage control using a large language model that follows instructions. The stage control process is divided into two stages: The first stage is response generation and preliminary routing: the scheduling module 120 receives the stage control prompt template, bounded history 10, training objective reduction, and current learner input, and simultaneously generates learner-oriented text responses and preliminary stage control actions, defined as follows: in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This indicates the input from the current learner, such as speech-to-text results or interactive commands; This represents the generated text response; This indicates the initial routing action.

[0065] The second stage is evidence-driven route update: After acquiring new evidence returned by the specialization module and completing the writing of the current round of evidence records, the scheduling module 120 re-evaluates the route decision based on the updated evidence digest of the current round and the updated bounded history 10 after writing the current round of evidence records, defined as: in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . This enables adaptive phase transition.

[0066] In one implementation, the scheduling module 120 receives the updated stage control action. Then, through the mapping function This is converted into executable system control actions; among which, Indicates system control actions, Indicates the current stage state. Indicates the current practice index. This represents the action mapping function. Specifically, when When the system advances the phase state to the next training phase, it updates the current practice index to the corresponding practice item for the next phase; when When RETRY, the system maintains the current stage state and either retains the current practice index or switches to the current stage's repetition item; when When this happens, the system will revert the stage state to the previous training stage and adjust the current practice index to the corresponding previous practice item.

[0067] To quantify the gap between the learner's current performance and the training objective, a target gap metric is defined: in, This represents the target gap indicator for the nth round. For dimension The target threshold; For dimension The actual score. When When this occurs, it indicates that the learner has reached or exceeded the target threshold in all dimensions, and the scheduling module 120 can trigger accordingly. action; The larger the value, the greater the gap between the target and the target. The scheduling module 120 can then determine whether to trigger the appropriate action. or Action. The target gap indicator can serve as a summary of evidence. The components are input into the decision update function of the second stage.

[0068] In one implementation, the scheduling module 120 maintains a bounded history window. Only keep the most recent The interaction content of the wheel is adjusted to control the context length and reduce the impact of redundant information on decision stability.

[0069] In another implementation, the scheduling module 120 can also handle the vocal range boundaries in the evidence digest. Stability indicators Dimensional scoring vector and dimensional explanation text After performing aggregate encoding, the data is then input into a large language model to improve the targeting of stage switching and task orchestration.

[0070] Among them, alternative implementations To avoid being limited by a single algorithm or deployment method, this invention also includes the following alternative implementations: The fundamental frequency estimation algorithm can be replaced by any algorithm or model that can output the fundamental frequency trajectory and confidence level frame by frame; The vocal range boundary estimation can be replaced by truncated mean, robust statistics, or cluster boundary estimation, as long as it can output the upper and lower bounds and update the profile field. Practice Tip 13 can be replaced with other scale modes, arpeggio modes, or span strategies, as long as they are constrained by the image field and can generate the target frequency sequence and prompt content. The segmentation method can be replaced by segmentation by lyrics, musical phrases, or fixed duration windows, as long as it can form a local closed loop of "segmented data collection - segmented evaluation - real-time feedback"; The stage control decision can be replaced by a large language model, a multimodal large model, or its lightweight, fine-tuned, or distilled variants, as long as it can output three types of control instructions (advance, repetition, and rollback) and update the stage state under the evidence summary drive.

[0071] Example 4 The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0072] like Figure 2 As shown, the method flow of the present invention may include the following steps: S101: Training objective specification. The objective module 100 acquires and records the training objective specification 8, and initializes the stage state and profile record.

[0073] S102: Real-time acquisition control. The scheduling module 120 sends a real-time acquisition control command 12 to the acquisition module 130 to acquire the learner's singing audio 1 and form an audio stream 2.

[0074] S103: Initial evidence summary 4 is generated. Evidence module 140 generates acoustic evidence based on audio stream 2 and forms initial evidence summary 4; at the same time, the sound range field 5 and stability index are written into the profiling module 160.

[0075] S104: Exercise prompt 13 judgment. The scheduling module 120 determines whether a new exercise prompt 13 needs to be output to the learner in the current round based on the current stage status and exercise index in the portrait module 160; if yes, execute S105, otherwise execute S106.

[0076] S105: Practice prompt 13 output. Presentation module 170 outputs practice prompt 13 and returns to S102 to enter the next round of acquisition control.

[0077] S106: Multimodal evaluation trigger judgment. The scheduling module 120 determines whether to call the structured evaluation process based on the multimodal large model, audio large model or audio-text multimodal large model according to the current stage type and exercise item category (e.g., triggered in the short sentence exercise or whole song exercise stage, skipped in the scale exercise or range test stage); if yes, execute S107, otherwise execute S109.

[0078] S107: Audio segment 3 acquisition. The scheduling module 120 determines the segment index, and the acquisition module 130 organizes the acquisition by segment to obtain audio segment 3.

[0079] S108: Structured Multimodal Evaluation. The evaluation module 150 inputs the audio segment 3, the evaluation prompt template, and the optional text context into the multimodal large model, the audio large model, or the audio-text multimodal large model, generates a structured evaluation result 7 containing multiple predefined marked segments, and parses it to obtain the score vector 6 and the explanatory text for each dimension; the presentation module 170 outputs real-time feedback and repetition instructions accordingly.

[0080] S109: Phase 1 response generation and preliminary routing. Scheduling module 120 is based on bounded history up to the previous round. The target threshold vector in the training objective reduction and current learner input Combined with stage control prompt template This forms the input for the first stage model: The input from the first-stage model is provided to the large language model to obtain the raw output of the first stage: Then, the original output of the first stage is parsed according to the predefined label rules to obtain the text response. Control actions in the initial stage : in, The value is , , One of them; the presentation module 170 responds to the text. Output corresponding text prompts, practice prompts, instant feedback, or review instructions.

[0081] S110: Evidence Summary Update. If the current round does not trigger a structured evaluation, the initial evidence summary 4 is written into bounded history 10 as the evidence record for the current round. The initial evidence summary 4 is defined as: If the current round has already generated a structured evaluation result 7, then the scoring vector will be... and explanatory text Compared with the initial evidence summary An updated evidence digest is generated by pooling these evidences and written into bounded history 10 as the current round of evidence records. The updated evidence digest is defined as follows: in, Indicates the first The acoustic evidence vector is output and encoded by the evidence module 140. This represents the score vector for structured assessment. Explanatory text representing structured assessment, This represents the evidence aggregation function.

[0082] S111: Second-stage evidence-driven routing update. After completing the current round of evidence digest update, the scheduling module 120 will send the stage control prompt template. Updated Bounded History and the current round of updated evidence summary The data is then combined to form the input for the second-stage model. The second-stage model input is fed into the large language model to obtain the original output of the second stage: Then, the original output of the second stage is parsed according to the action label rules to obtain the updated stage control action. : in, The value is , , One of them; the scheduling module 120 based on the updated stage control action Output final stage control command 12.

[0083] S112: Control action mapping and state update. The scheduling module 120 will map the updated stage control actions. Mapped to system control actions And update the current stage status in the portrait module 160. With current practice index : when When, switch to the corresponding exercise for the next stage; when When RETRY, maintain the current stage and keep the current practice index, or switch to the current stage's repetition item; when When that happens, switch to the corresponding exercise from the previous stage.

[0084] S113: Termination Judgment. Determine whether to output a stage report; if not, return to S102 to continue the loop; if yes, execute S114.

[0085] S114: Stage report output. The stage report is output by presentation module 170, and the process ends.

[0086] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A real-time adaptive singing guidance method based on a multimodal large model, characterized in that, The method includes: S1. Process the learner's singing audio, extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, and generate an initial evidence summary; after obtaining the structured evaluation results, update the initial evidence summary based on the structured evaluation results; S2. Input the learner's singing audio into the multimodal large model, audio large model, audio-text multimodal large model or its fine-tuned model, and combine it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text; S3. The training objective reduction, current evidence summary, bounded interaction history and current learner input are provided as input to the large language model to generate stage control instructions in real time. S4. Based on the stage control instructions, output practice prompts, immediate feedback, repetition instructions, or stage reports that match the current training stage to the learner.

2. The method according to claim 1, characterized in that, In step S1, the method for processing the learner's singing audio and extracting acoustic evidence including fundamental frequency trajectory, pitch stability index, and vocal range boundary to generate an initial evidence summary includes: in, This represents the set of acoustic evidence signals or its encoded acoustic evidence vector; the two zeros represent default placeholders corresponding to the score vector and interpretation information, respectively, when structured evaluation results have not yet been obtained.

3. The method according to claim 1, characterized in that, In step S2, the method for inputting the learner's singing audio into a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof, and combining it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text, includes: Input the audio clip, evaluation prompt template, and optional text context into the model to obtain the raw output: ； in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context; Represents the inference function for a multimodal large model, an audio large model, or an audio-text multimodal large model; This represents the raw text output by the model, containing multiple predefined token segments, each token segment corresponding to a set of dimensions. One dimension of singing ability; Perform deterministic analysis on the model's original output: ； in, This represents a deterministic analytic function that decomposes the original output into a score vector and explanatory text by locating predefined marked segments. The score vector and explanatory text are as follows: ； ； in, Representing dimensions The rating value is an integer ranging from 1 to 5; Representing dimensions The explanatory text describes the specific performance of this dimension and provides suggestions for improvement.

4. The method according to claim 1, characterized in that, In S3, the method of providing the training objective reduction, current evidence summary, bounded interaction history, and current learner input as input to the large language model, and generating stage control instructions in real time, includes two stages: The first stage is response generation and preliminary routing: receiving the stage control cue template, bounded history, training objective reduction, and current learner input, while simultaneously generating learner-oriented text responses and preliminary stage control actions. ； in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This represents the input from the current learner; This represents the generated text response; Indicates the initial routing action; The second phase is evidence-driven route update: the routing decision is re-evaluated based on the updated evidence summary and the latest bounded history. ； in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . ,in, To move to the next stage, To repeat practice at the current stage, This is to revert to the previous stage, thereby achieving adaptive stage transition.

5. A real-time singing adaptive guidance system based on a multimodal large model, the system being used to implement the method described in any one of claims 1-4, characterized in that, The system includes: a target module, a scheduling module, a data acquisition module, an evidence module, an evaluation module, a profiling module, and a presentation module; The target module is used to acquire and maintain the learner's training target specification, which includes at least target practice items, target ability thresholds, and target difficulty. The acquisition module is used to acquire learner audio input and provide audio data to the evidence module and the evaluation module; The evidence module is used to process learners' singing audio, extract acoustic evidence including fundamental frequency trajectory, pitch stability index and vocal range boundary, generate an initial evidence summary, and update the initial evidence summary after obtaining structured evaluation results. The evaluation module is used to input learners' singing audio into a multimodal large model, an audio large model, an audio-text multimodal large model or its fine-tuned model, and combine it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text. The profile module is used to store learner profiles and interaction history, and to provide a bounded interaction history for the scheduling module; The scheduling module is used to provide the training objective reduction, current evidence summary, bounded interaction history and current learner input as input to the large language model, and generate stage control instructions in real time. The presentation module is used to output practice prompts, real-time feedback, repetition instructions, or stage reports that match the current training stage to the learner, based on the stage control instructions.

6. The system according to claim 5, characterized in that, The evidence module involves processing the learner's singing audio to extract acoustic evidence, including fundamental frequency trajectory, pitch stability index, and vocal range boundaries, and generating an initial evidence summary. in, This represents the set of acoustic evidence signals or its encoded acoustic evidence vector; the two zeros represent default placeholders corresponding to the score vector and interpretation information, respectively, when structured evaluation results have not yet been obtained.

7. The system according to claim 5, characterized in that, The evaluation module involves inputting the learner's singing audio into a multimodal large model, an audio large model, an audio-text multimodal large model, or a fine-tuned model thereof, and combining it with a predefined output format to generate a structured evaluation vector containing score values ​​for each dimension and explanatory text. This process includes: Input the audio clip, evaluation prompt template, and optional text context into the model to obtain the raw output: ； in, Indicates the first The audio clip input in the wheel; This is an evaluation prompt template used to define the scoring dimensions, scoring range, output label format, and interpretation granularity. Indicates an optional text context; Represents the inference function for a multimodal large model, an audio large model, or an audio-text multimodal large model; This represents the raw text output by the model, containing multiple predefined token segments, each token segment corresponding to a set of dimensions. One dimension of singing ability; Perform deterministic analysis on the model's original output: ； in, This represents a deterministic analytic function that decomposes the original output into a score vector and explanatory text by locating predefined marked segments. The score vector and explanatory text are as follows: ； ； in, Representing dimensions The rating value is an integer ranging from 1 to 5; Representing dimensions The explanatory text describes the specific performance of this dimension and provides suggestions for improvement.

8. The system according to claim 5, characterized in that, In the scheduling module, the process of providing the training objective reduction, current evidence summary, bounded interaction history, and current learner input as input to the large language model, and generating stage control instructions in real time, includes two stages: The first stage is response generation and preliminary routing: receiving the stage control cue template, bounded history, training objective reduction, and current learner input, while simultaneously generating learner-oriented text responses and preliminary stage control actions. ； in, This indicates that the instruction follows the inference function of the large language model; This represents a stage control prompt template used to constrain the output format and decision logic of a large language model. This indicates the bounded history up to the previous round; Represents the target threshold vector; This represents the input from the current learner; This represents the generated text response; Indicates the initial routing action; The second phase is evidence-driven route update: the routing decision is re-evaluated based on the updated evidence summary and the latest bounded history. ； in, Represents the decision update function of a large language model; Indicates the first Summary of updated evidence for the wheel; This represents the updated stage control action, with a value space of . ,in, To move to the next stage, To repeat practice at the current stage, This is to revert to the previous stage, thereby achieving adaptive stage transition.

Images