An adaptive inference method for multi-round question and answer of streaming video

Abstract

The application provides a kind of self-adapting inference method for stream video multi-round question and answer, it is related to artificial intelligence technical field, including: the current video segment of stream video, current round problem and historical question and answer context are input into pre-trained video understanding model, the relationship label for current round problem, initial prediction answer and the confidence of initial prediction answer that video understanding model output is obtained;According to the size comparison result of confidence and preset threshold value, determine the target answer result of current round problem, and merge relationship label into the historical question and answer context of next round.The application introduces the quantitative evaluation of the degree of certainty of the model's preliminary prediction result, can adaptively judge according to the certainty state of the model to different problems and determine the target answer result, thereby effectively improving the rationality and adaptive ability of the answer result determination in stream video multi-round question and answer system.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an adaptive reasoning method for multi-turn question answering of streaming video. Background Technology

[0002] With the rapid development of multimodal large-scale models, video understanding and question answering technologies have been widely applied in scenarios such as autonomous driving and intelligent monitoring. However, in practical applications, models often deal with continuously input streaming video, requiring multi-round question-and-answer interactions. Most existing video question answering methods are trained based on single-round offline videos, lacking the ability to explicitly model the inherent connections and historical context information between consecutive questions, which can easily lead to inconsistencies in the understanding of preceding and following content when handling cross-question reasoning.

[0003] To address the context consistency issue in multi-turn question answering of streaming videos, existing video multi-turn question answering methods introduce explicit historical information reference mechanisms during the inference phase. A common approach is to directly input historical video clips and previous question-and-answer results as contextual input to the model when answering the current question, or to have the model generate a detailed analysis process for each question, thereby improving the model's consistency in answering complex questions.

[0004] However, when determining the final answer to the current round of questions, existing video question answering methods typically output the initial answer generated by the model directly, lacking a mechanism for quantitatively evaluating the reliability or certainty of the generated answer. This fixed and singular method of answer determination cannot make differentiated decisions based on the model's level of confidence in different questions, resulting in a lack of adaptive judgment ability for different states of certainty in the question answering system, thus affecting the rationality and reliability of the final answer. Therefore, how to provide a streaming video multi-round question answering method with answer reliability assessment and adaptive determination capabilities has become an urgent problem to be solved in this field. Summary of the Invention

[0005] This invention provides an adaptive reasoning method for multi-turn question answering in streaming video, which addresses the technical problem of how to provide a multi-turn question answering method for streaming video with the ability to assess the reliability of answers and make adaptive determinations.

[0006] This invention provides an adaptive reasoning method for multi-turn question answering in streaming video, comprising: The current video segment of the streaming video, the current round of questions, and the historical question-answering context are input into a pre-trained video understanding model to obtain the relational label, initial predicted answer, and confidence score of the initial predicted answer output by the video understanding model for the current round of questions. Based on the comparison between the confidence level and the preset threshold, the target answer result for the current round of questions is determined, and the relationship label is merged into the historical question-and-answer context of the next round.

[0007] According to the adaptive reasoning method for multi-turn question answering of streaming video provided by the present invention, the video understanding model is trained through the following steps: Multiple-choice data is generated based on raw streaming video question-and-answer samples; Structured multi-round inference data is constructed based on the multiple-choice data; The baseline model is trained under supervised supervision based on the structured multi-round inference data. The video understanding model is obtained by combining multiple reward constraints to train the baseline model after supervised fine-tuning training through reinforcement learning.

[0008] An adaptive reasoning method for multi-turn question answering in streaming video, provided by the present invention, generates multiple-choice data based on original streaming video question answering samples, including: Based on the video content, corresponding questions, and real answers in the original streaming video question-and-answer sample, multiple interference options are generated to obtain a multi-choice sample to be filtered, which includes the real answers and the interference options; The multiple-choice data is obtained based on the sample of multiple-choice options to be screened.

[0009] An adaptive reasoning method for multi-turn question answering in streaming video, provided by the present invention, obtains the multiple-choice data based on the multiple-choice sample to be filtered, including: The model was validated by independently performing multiple option prediction operations on the multiple-choice sample to be screened. Calculate the probability of correct prediction for the multiple option prediction operations; If the probability of a correct prediction is greater than a set threshold, multiple interference options are regenerated until the probability of a correct prediction is not greater than the set threshold. If the probability of a correct prediction is not greater than the set threshold, the sample to be screened for multiple selection is used as the multiple selection data.

[0010] An adaptive inference method for multi-turn question answering in streaming video, provided by the present invention, constructs structured multi-turn inference data based on the multiple-choice data, including: Perform multiple independent inference operations on the samples to be processed in the multiple-choice data, and count the corresponding prediction frequency results; The initial answer and prediction confidence of the sample to be processed are determined based on the prediction frequency results. Based on the historical question-and-answer information of the sample to be processed, sample relationship labels are generated, and based on the sample relationship labels, reasoning operations are performed again to determine the thought chain process and verify the answer; The sample relationship labels, the initial sample answer, the sample prediction confidence, the thought process, and the verification answer are combined in a specific format to obtain the structured multi-round reasoning data.

[0011] According to the adaptive inference method for multi-turn question answering in streaming video provided by the present invention, reinforcement learning training is performed on the baseline model after supervised fine-tuning training by combining multiple reward constraints, including: The baseline model is used to generate candidate output data for the input samples. Determine the various reward constraints of the candidate output data; The total reward is obtained by weighted summation of the various reward constraints. The parameters of the baseline model are updated based on the total reward using a policy optimization algorithm.

[0012] According to the present invention, an adaptive reasoning method for multi-turn question answering of streaming video is provided, wherein the multiple reward constraints include multiple types of format reward, accuracy reward, relational reward and confidence reward; The format reward is used to constrain the video understanding model to strictly follow a predefined structured output format; The accuracy reward is used to compare and verify whether the answer matches the actual answer; The relation reward is used to measure the contribution of the generated relation labels to subsequent problem reasoning; The confidence reward is used to constrain the confidence level of the explicit output of the video understanding model to be consistent with the quality of the response.

[0013] According to an adaptive inference method for multi-turn question answering in streaming video provided by the present invention, the confidence reward is determined by: Determine the first correct state of the initial prediction portion of the candidate output data relative to the true answer, and the second correct state of the final inference portion relative to the true answer; Based on the second correct state, the output confidence in the candidate output data is calculated by offset product based on the center value to obtain the basic penalty term; The correction term is determined based on the state difference between the second correct state and the first correct state; The confidence reward is obtained by adding the basic penalty term to the correction term.

[0014] According to an adaptive reasoning method for multi-turn question answering in streaming video provided by the present invention, determining the relation reward includes: Extract the cross-question relationship tags contained in the candidate output data; Determine the first accuracy obtained by using the cross-question relationship labels to perform reasoning verification on the next round of questions; The second average accuracy is obtained by performing reasoning verification on the next round of questions using the same set of comparative samples. The first accuracy rate is compared with the second average accuracy rate, and the corresponding relationship reward is determined based on the comparison result.

[0015] An adaptive inference method for multi-turn question answering in streaming video provided by the present invention determines the target answer result for the current round question based on a comparison result between the confidence level and a preset threshold, including: If the confidence level is greater than or equal to the preset threshold, the initial predicted answer is extracted as the target answer result; If the confidence level is less than the preset threshold, the video understanding model continues to generate the thought chain reasoning process and the final reasoning answer for the current round of questions, and the final reasoning answer is used as the target answer result.

[0016] The present invention also provides an adaptive inference device for multi-turn question answering in streaming video, comprising: The input module is used to input the current video segment of the streaming video, the current round of questions, and the historical question-answering context into a pre-trained video understanding model, and to obtain the relational label, initial predicted answer, and confidence score of the initial predicted answer output by the video understanding model for the current round of questions. The determination module is used to determine the target answer result of the current round of questions based on the comparison result of the confidence level and the preset threshold, and to merge the relationship label into the historical question-and-answer context of the next round.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement an adaptive reasoning method for multi-turn question answering of streaming video as described above.

[0018] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an adaptive reasoning method for multi-turn question answering of streaming video as described above.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements an adaptive reasoning method for multi-turn question answering of streaming video as described above.

[0020] The adaptive reasoning method for multi-turn question answering in streaming video provided by this invention obtains relational labels for the current question by inputting the current segment of the streaming video, the current round question, and the historical question-answering context into the model. These relational labels are then merged into the historical question-answering context of the next round, achieving explicit modeling and transfer of historical information between consecutive questions. This effectively overcomes the shortcomings of existing technologies in terms of inconsistent understanding of content during cross-question reasoning, ensuring the coherence of multi-turn question answering. Simultaneously, the model outputs an initial predicted answer and its confidence level. The target answer for the current round question is determined based on a comparison between the confidence level and a preset threshold. This introduces a quantitative assessment of the confidence level of the model's initial prediction, addressing the lack of a mechanism for quantitatively assessing the reliability of generated answers and the inability to make differentiated decisions based on confidence levels in existing video question answering methods. This method can adaptively judge the certainty state of different questions based on the model and determine the target answer, thereby effectively improving the rationality and adaptability of answer determination in streaming video multi-turn question answering systems. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the adaptive reasoning method for multi-turn question answering in streaming video provided by this invention.

[0023] Figure 2 This is a schematic diagram of the architecture of the adaptive reasoning method for multi-turn question answering of streaming video provided by the present invention.

[0024] Figure 3 This is an example schematic diagram of the structured supervised fine-tuning training set provided by the present invention.

[0025] Figure 4 This is a schematic diagram comparing the present invention with other fixed strategy methods.

[0026] Figure 5 This is a schematic diagram of the adaptive reasoning device for multi-turn question answering of streaming video provided by the present invention.

[0027] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] With the rapid development of multimodal large-scale models, they have demonstrated strong capabilities in video understanding and video question answering tasks. However, in real-world scenarios such as autonomous driving, intelligent surveillance, and embodied intelligence, these models deal with continuously inputting streaming video, rather than static or offline video. Unlike traditional video understanding, multi-turn question answering of streaming video requires models to not only understand the current content but also to perform continuous reasoning by combining historical video clips and previous question-and-answer results. Most existing technologies are trained based on single-turn offline video question answering, lacking explicit modeling of the inherent relationships between questions. Therefore, inconsistencies in the understanding of different content can easily arise in online multi-turn question answering scenarios.

[0030] Furthermore, existing streaming video question answering methods still have shortcomings in their reasoning methods. One type of method directly outputs the answer based on the current video content. While this approach is relatively fast, it often struggles to guarantee accuracy for complex questions involving historical context, cross-question relationship analysis, or multi-step reasoning. Another type of method improves its ability to handle complex questions by introducing explicit reasoning processes, but it typically employs fixed reasoning strategies, performing the same depth of reasoning for questions of varying complexity. This approach not only introduces redundant computations and reduces response efficiency for simple questions but also easily leads to overthinking, making it difficult to simultaneously achieve both accuracy and real-time performance.

[0031] Existing video understanding models for multi-turn question answering tasks in streaming video are mostly trained based on single-turn offline video question answering methods. They lack the ability to explicitly model the inherent relationships and historical context information between consecutive questions in multi-turn question answering of streaming video. Therefore, when dealing with cross-question reasoning, inconsistencies in the understanding of the preceding and following content are prone to occur, which affects the accuracy and consistency of the answers.

[0032] Existing video question answering methods often lack flexibility in their reasoning mechanisms. On the one hand, direct answers are difficult to handle complex questions involving historical information and multi-step analysis. On the other hand, explicit reasoning methods often adopt fixed reasoning strategies, which can easily lead to computational redundancy for simple questions and make it difficult to balance overall response efficiency and reasoning quality.

[0033] The following is combined with Figures 1 to 6 This invention describes an adaptive reasoning method for multi-turn question answering in streaming video.

[0034] Figure 1This is a flowchart illustrating the adaptive reasoning method for multi-turn question answering in streaming video provided by the present invention, as shown below. Figure 1 As shown, the method includes, but is not limited to, steps S1 and S2.

[0035] Step S1: Input the current video segment of the streaming video, the current round question, and the historical question-answering context into the pre-trained video understanding model to obtain the relationship label, initial predicted answer, and confidence score of the initial predicted answer output by the video understanding model for the current round question.

[0036] Streaming video is video data that is transmitted and continuously input in the form of a continuous data stream. It can include continuous video recordings of autonomous driving vehicles, real-time video streams of intelligent monitoring systems, online video streams with embody intelligent perspectives, or live web streaming video.

[0037] The current video segment is a portion of the video content extracted from the streaming video that is relevant to the current moment or the current question. It can be a sequence of keyframes extracted through a dynamic frame sampling strategy, a combination of consecutive video frame segments, or a video feature vector after feature encoding.

[0038] The current round question is a text query request made by the user or system for the video content in the current round of interaction. It can include questions about the attribute analysis of objects in the picture, action recognition, event causal reasoning, or time series prediction.

[0039] Historical question-answering context is the historical information of multiple rounds of interactions that have been completed before the current round. It can include the question text of the previous round, the corresponding historical answer text, the cross-question relationship information generated previously, and the feature information of the historical segments.

[0040] Video understanding models are deep learning models capable of processing multimodal data (video, image, text) and understanding temporal relationships. They can be large visual language models or multimodal large language models trained with specific supervised fine-tuning and reinforcement learning, such as the Streaming-o1 model trained based on Qwen2.5-VL-7B-Instruct and other base models.

[0041] Relationship labels are text summaries or conclusions that explicitly indicate the connection between the current segment or question and its historical context. The initial predicted answer is a preliminary judgment or option prediction given by the model directly to the current question without in-depth multi-step reasoning. Confidence level is the model's degree of certainty or probability estimate of its initial predicted answer, ranging from 0 to 1.

[0042] For example, in a multi-turn question-and-answer scenario for streaming video in autonomous driving monitoring, the historical question-and-answer context records "Previous question: What is ahead? Previous answer: Traffic light". The current video segment is a series of frames showing the traffic light changing from red to green. The current question is "What should the vehicle do now?". After serializing this information and inputting it into the video understanding model through a local API interface, the model outputs the relationship label "related to the previous traffic light state, the state has changed", and the initial predicted answer is "start the vehicle" with a confidence level of 0.95.

[0043] Step S1 introduces historical question-and-answer context and relational labels, enabling the model to explicitly model the inherent temporal relationship between the current question and historical information, overcoming the limitation of understanding a single question independently; at the same time, the model outputs a preliminary answer and its confidence level, providing an evaluation basis for subsequent steps to determine whether in-depth reasoning is needed.

[0044] Step S2: Based on the comparison between the confidence level and the preset threshold, determine the target answer to the current round of questions and merge the relationship label into the historical question-and-answer context of the next round.

[0045] The preset threshold can be 0.8, 0.85, or 0.9, etc. The target answer is the final definitive answer that the system ultimately displays to the user or outputs as a control command. It can be the directly extracted initial predicted answer or the final reasoned answer after in-depth thought chain reasoning.

[0046] For example, if the preset threshold is set to 0.8, and the confidence score for step S1 is 0.95, which is greater than 0.8, it indicates that the model is very confident in the initial answer "start the vehicle". At this point, "start the vehicle" is determined as the target answer and output directly. Simultaneously, the relationship label "related to the previous traffic light state, state has changed" is added to the history storage list for use in the next question.

[0047] As described above, this invention obtains relational labels for the current question by inputting the current segment of the streaming video, the current round of questions, and the historical question-and-answer context into the model. These relational labels are then merged into the historical question-and-answer context of the next round, enabling explicit modeling and transfer of historical information between consecutive questions. This effectively overcomes the shortcomings of existing technologies in terms of inconsistent understanding of content during cross-question reasoning, ensuring the coherence of multi-round question-and-answer sessions. Simultaneously, the model outputs an initial predicted answer and its confidence level. The target answer for the current round of questions is determined based on a comparison between the confidence level and a preset threshold. This introduces a quantitative assessment of the confidence level of the model's initial prediction, addressing the lack of a mechanism for quantitatively assessing the reliability of generated answers and the inability to make differentiated decisions based on confidence levels in existing video question-and-answer methods. The invention allows for adaptive judgment based on the model's certainty state for different questions, thereby determining the target answer and effectively improving the rationality and adaptability of answer determination in streaming video multi-round question-and-answer systems.

[0048] In one embodiment, such as Figure 2 As shown, the video understanding model of the present invention can be trained through the following steps: Multiple-choice data is generated based on raw streaming video question-and-answer samples; Construct structured multi-round inference data based on multiple-choice data; Supervised fine-tuning training of the baseline model is performed based on structured multi-round inference data; By combining multiple reward constraints, reinforcement learning is applied to the baseline model after supervised fine-tuning to obtain a video understanding model.

[0049] The raw streaming video question-answering samples are basic image-text video association data obtained from open-source datasets or actual acquisition scenarios without special formatting. These samples can include open-ended question-answering data with temporal correlation and contextual dependence. In specific implementations, this invention collects multiple publicly available streaming video question-answering datasets, including StreamingBench, StreamBench, SVBench, and OVBench. Since this invention targets streaming video question-answering scenarios with cross-question correlations, only samples that cannot be answered independently of their historical context are retained, including the action and event subsets in OVBench and the Sequential Question Answering section in StreamingBench. Because SVBench and StreamBench themselves have strong temporal correlations, all samples are retained. Simultaneously, the Video-R1 dataset, which includes thought chain annotations and image-video hybrid processing, is introduced to enhance the model's perception and basic reasoning capabilities.

[0050] Since some samples are in open-ended question-and-answer format, they need to be converted into multiple-choice question-and-answer format to obtain stable supervision signals. Multiple-choice data is generated by converting open-ended questions into a data format containing one true answer and multiple incorrect candidates (e.g., a four-choice format). Standardizing the variable open-ended questions into uniformly formatted multiple-choice questions can obtain stable and easily quantifiable supervision training signals, eliminating evaluation bias caused by differences in the surface form of open-ended answers (such as sentence length and synonym expression).

[0051] Structured multi-turn reasoning data consists of training samples with a unified predefined field format, which can include structured labels for representing relationships, initial answers, confidence levels, thought processes, and final answers. Structured multi-turn reasoning data can provide a standard data template for subsequent supervised fine-tuning, explicitly representing the relationship between the current question and its historical context, as well as the multi-step logical reasoning process.

[0052] Based on the structured dataset constructed above, this invention first performs supervised fine-tuning training on the baseline model to initialize the basic visual language model, enabling it to follow a predefined structured output format and perform preliminary multi-turn inference. Since relying solely on supervised fine-tuning, the model's optimization objective mainly focuses on local generation quality, making it difficult to fully improve cross-question content understanding consistency, relationship modeling ability, and confidence calibration ability. Therefore, after supervised fine-tuning, reinforcement learning is further introduced for joint optimization to improve the model's overall performance in streaming video multi-turn question answering scenarios.

[0053] The baseline model can be a general-purpose open-source large model such as Qwen2.5-VL-7B-Instruct. Supervised fine-tuning training uses labeled pairwise data (input hints - expected structured output) to update model parameters through the backpropagation algorithm. This allows the baseline model to initially learn the preset structured output template format and possess preliminary cross-question content understanding and multi-turn reasoning ability.

[0054] Multiple reward constraints are used to evaluate the quality of model generation from different dimensions, including format validity scores, logical correctness scores, cross-problem coherence scores, and confidence alignment scores. Reinforcement learning training enables the model to optimize its behavioral policies as an agent through trial and error exploration and maximizing cumulative rewards in the environment. Reinforcement learning algorithms such as proximal policy optimization, group relative policy optimization, or direct preference optimization can be used to calculate policy gradients using multiple reward constraints, thereby guiding the update of the model's network weight parameters. This can compensate for problems such as local optima or format collapse that can easily occur with supervised fine-tuning alone, and jointly optimize the model's output structure stability, historical consistency, and confidence self-calibration ability from a global, multi-dimensional perspective.

[0055] This invention employs a progressive training process involving generating multiple-choice data, constructing structured multi-round inference data, supervised fine-tuning training based on structured data, and reinforcement learning training with multiple reward constraints. This not only ensures that the final video understanding model strictly adheres to a predefined structured output format that facilitates subsequent parsing, but also effectively improves the model's stability in reasoning during continuous multi-round complex interactions and the accuracy of its self-confidence assessment through multi-objective reinforcement learning constraints. This comprehensively enhances the model's basic visual logic understanding and response reliability.

[0056] In one embodiment, generating multiple-choice data based on raw streaming video question-answering samples may further include: Based on the video content, corresponding questions, and real answers in the original streaming video question-and-answer sample, multiple interference options are generated to obtain a multi-choice sample to be filtered, which includes real answers and interference options; Multiple-choice data is obtained from the sample of multiple-choice questions to be screened.

[0057] Distractor options are answer choices that have some semantic similarity or visual confusion to the real answer, but whose descriptions are incorrect. These can include incorrect descriptions of actions within the video, incorrect judgments of the state of objects within the video, or descriptions with reversed time sequences. The multiple-choice sample to be filtered includes the question, the real answer, and multiple distractor options.

[0058] The sampled frames of the current video segment, the corresponding question, and the true answer can be combined and assembled into a prompt, which is then input into a general multimodal large language model, such as GPT-4o. Its generation capabilities are then used to generate multiple distractor options. For example, for an open sample without candidate options, the original question is "What is the person in the video doing?", and the open-ended true answer is "cutting an apple." The GPT-4o model is called to generate three distractor options that are semantically close to the true answer and highly deceptive, while maintaining consistency in style and length among the candidate options. Assuming the distractor options generated by GPT-4o are "cutting a pear," "washing an apple," and "peeling an apple," these are combined with the true correct option to obtain a four-choice sample to be filtered.

[0059] Multiple-choice data refers to the qualified multiple-choice data format that has undergone a screening mechanism and is ultimately used for subsequent training and processing. An auxiliary model can be used to score and filter the options' discriminative power to eliminate unqualified samples.

[0060] This invention combines the original video content, the corresponding question, and the real answer to generate multiple interference options, forming a complete sample of multiple-choice questions to be screened. This effectively ensures the deceptiveness and consistency of the interference options in terms of contextual semantics and expression style, avoiding the superficial form differences caused by simple random sampling. Thus, it constructs a structurally unified and challenging multiple-choice sample, providing a more stable and high-quality comparative learning object for video understanding models.

[0061] In one embodiment, obtaining multiple-choice data based on the samples to be screened may further include: The model was validated by independently performing multiple option prediction operations on the multiple-choice sample to be screened. Calculate the probability of correct prediction for multiple-option prediction operations; If the probability of a correct prediction is greater than a set threshold, multiple interference options are regenerated until the probability of a correct prediction is no greater than the set threshold. If the probability of a correct prediction is not greater than a set threshold, the sample to be screened for multiple-choice questions will be used as multiple-choice data.

[0062] To avoid generating overly simplistic distractors, this invention introduces a difficulty filtering mechanism based on a validation model. The validation model is an auxiliary visual language model specifically designed to evaluate the difficulty of answering questions and the discriminative power of distractors. Lightweight and fast-responding multimodal large models such as Gemini 3 Flash can be used. Multiple option prediction operations involve having the validation model repeatedly select answers under the same input conditions or with slight perturbations. For example, given a sample to be filtered (containing video clip sampling frames, the question, and four candidate options), the Gemini 3 Flash validation model independently performs six forward inference judgment tasks involving four options each.

[0063] The probability of a correct prediction is the ratio of the frequency at which the validation model correctly answers the true answer, or its accuracy rate. It can be calculated by summing the number of times the validation model correctly selects the true answer and then dividing by the total number of executions. For example, in the six independent four-choice judgments executed above, the Gemini 3 Flash correctly selected the true answer four times, resulting in a probability of 4 / 6 ≈ 66.7%.

[0064] The threshold can be set empirically, such as 50% or 60%. For example, the threshold can be set to 50%. Since the statistically correct probability of 66.7% is greater than 50%, it is determined that the candidate group is consistently identified with high accuracy, but the discrimination is insufficient. Therefore, it is necessary to regenerate the interference options. The generator model GPT-4o is automatically sent with the instruction "Please increase the deceptiveness" to regenerate 3 new interference options. The validation model Gemini 3 Flash is then asked to make 6 predictions again. If the correct probability of the new prediction is 33.3%, which is not greater than 50%, and the preset difficulty requirement is met, the loop stops.

[0065] For the sample whose accuracy dropped to 33.3% after being regenerated, since 33.3% < 50%, it indicates that the current distractor options are highly deceptive. Therefore, the sample combination (corresponding to the sampling frame, question, true answer option, and newly generated distractor option) is retained and written into the final usable multiple choice dataset.

[0066] This invention introduces a validation model that independently performs multiple option predictions on the multiple-choice sample to be screened and calculates the probability of correct prediction. Then, when the probability is greater than a set threshold, interference options are regenerated until the probability reaches the threshold. This effectively avoids the problem that the generated interference options are too simple and lose their discriminatory power, thus ensuring that the final multiple-choice data has sufficient discrimination difficulty and provides effective discrimination supervision for subsequent model training.

[0067] In one embodiment, constructing structured multi-turn inference data based on multiple-choice data may further include: Perform multiple independent inference operations on the samples to be processed in the multiple-choice data, and count the corresponding prediction frequency results; The initial answer and prediction confidence of the sample to be processed are determined based on the prediction frequency results. Based on the historical question-and-answer information of the sample to be processed, sample relationship labels are generated, and reasoning operations are performed again based on the sample relationship labels to determine the thought process and verify the answer; By combining the sample relationship labels, initial sample answers, sample prediction confidence, thought process, and verification answers in a specific format, structured multi-round reasoning data is obtained.

[0068] The samples to be processed are those currently extracted from the multiple-choice dataset, ready for two-stage formatted data construction. They can contain frame sequences from single-round or multi-round video question-and-answer sessions, along with text for multiple-choice questions with four options. The prediction frequency result is the distribution characteristic of the number of times each candidate answer appears in multiple samples. A specific instruction can be used to fine-tune the inference model, such as the Qwen2.5-7B-Instruct model. For a given input question, candidate options, and corresponding sample frames, K (e.g., K=8) independent inferences are performed, and the number of times each candidate answer 'a' is predicted is counted, denoted as N(a). For example, using the Qwen2.5-7B-Instruct model to perform 8 independent inferences on the input sample, the statistical result is that option A appears 5 times, option B appears 2 times, and option C appears 1 time; this is the prediction frequency result.

[0069] The initial sample answer is the optimal choice tendency extracted from the statistical frequency results; it can be the most frequent option letter or its corresponding text content. The sample prediction confidence score, denoted as c, quantifies the reliability of the initial answer and ranges from 0 to 1. The candidate option with the highest prediction frequency (i.e., maxN(a)) can be selected as the initial sample answer; and its relative frequency in the total number of predictions is calculated as the confidence estimate, using the following formula: Where K is the total number of independent reasoning operations. For example, given the statistical result that option A appears 5 times, K=8, we determine that A, which has the highest frequency, is the initial answer for the sample, and at the same time, we calculate the sample prediction confidence level c=5 / 8=0.625.

[0070] Sample relationship labels are used to summarize the dynamic temporal relationship between the current input and historical information. The thought process is a detailed logical analysis text that involves multiple steps of step-by-step deduction. The verification answer is the final answer obtained after in-depth thought process analysis.

[0071] The designed prompts, historical question-answer pairs, current round questions, and corresponding sampling frames can be input into a given large-scale parameter model, such as the Qwen3-VL-30B-A3B-Instruct model, to obtain relation labels. If it is the first round and does not rely on historical information, preset text (such as Single Query) is directly generated as relation labels.

[0072] After obtaining the relation labels, the relation labels, the current question, candidate options, and sampling frames are used as input to execute the large model K=8 times again. The output format is as follows: <think>< / think> <answer>< / answer> The independent reasoning is represented by the `think` field, which indicates the thought process, and the `answer` field, which indicates the verification of the answer.

[0073] If there exists an answer obtained from a certain reasoning (denoted as the i-th time). With the real answer If they are consistent, then the thought process involves verifying the correct answer. The corresponding thought chain and answer If the predictions are inconsistent in any instance, select the thought chain and answer corresponding to the answer with the highest prediction frequency (i.e., satisfying the condition). of The decision to retain this information is based on the following formula: Structured multi-turn inference data can be represented as: <relation> ...< / relation> <answer> ...< / answer> <confidence> ...< / confidence> <think> ...< / think> <answer> ...< / answer> ); Among them, the relation field represents the sample relationship label between the current question and the historical context; the first answer field represents the model's initial sample answer to the current question; the confidence field represents the sample prediction confidence; the think field represents the thought process; and the second answer field represents the verification answer after reasoning.

[0074] For data containing only single-round questions and already provided with available thought chains and final answer annotations, the existing thought chains and final answers can be directly reused, and the initial predicted answers and their confidence levels can be constructed according to the first-stage process described above. After automated construction, human annotators can review, revise, and verify the retained thought chains to improve their quality in terms of correctness and consistency. Based on whether the final answer is correct, the samples can be further divided into supervised fine-tuning datasets and reinforcement learning datasets. The supervised fine-tuning dataset retains only high-quality samples with correct final answers, while the reinforcement learning dataset contains all samples. An example of the supervised fine-tuning dataset of this invention is shown below. Figure 3 As shown.

[0075] This invention obtains initial answers and confidence levels by first performing multiple independent inferences and statistical predictions on samples, then performing inferences again based on relational labels to select the optimal thought chain and verify the answer, and finally combining the information in various formats. This results in structured multi-round inference data that not only contains high-quality logical thought chains that reveal the derivation process, but also integrates quantitative information on confidence levels that accurately reflect the inherent uncertainty of the model. This provides a rich and hierarchical learning dimension for subsequent model fine-tuning, significantly improving the model's ability to imitate and generalize inference links and confidence outputs.

[0076] In one embodiment, training the supervised fine-tuned baseline model with reinforcement learning by combining multiple reward constraints may further include: Candidate output data are generated from the input samples using a baseline model; Determine multiple reward constraints for candidate output data; The total reward is obtained by weighted summation of various reward constraints. The parameters of the baseline model are updated based on the total reward using a policy optimization algorithm.

[0077] The input sample, denoted as x, is a multimodal training batch data object with a cue prefix used in the reinforcement learning training phase. The candidate output data consists of complete sequence response texts with predefined format labels actually generated by the baseline model under the current policy based on this input environment. Specifically, it manifests as multiple sets of text outputs containing different reasoning exploration paths or answer results. The current policy baseline model (denoted as x) can be utilized. (where θ is the model parameter) A random sampling and decoding strategy is used for the same input x, for example, setting the sampling temperature to 1.0 to encourage large-scale exploration, generating G different exploratory candidate outputs, denoted as . .

[0078] Then output the i-th candidate in the group. Multiple independent evaluation metrics (such as legality and accuracy) are calculated and scored separately. For example, for the generated candidate output, if its format label is found to be completely closed and compliant, the format reward constraint is 1.0; if the last answer is extracted and compared with the real answer and the option is incorrect, the accuracy reward constraint is -1.0.

[0079] Unlike single-turn video question answering, multi-turn streaming video question answering requires the model to not only ensure the correctness of the answer in the current turn, but also maintain consistent understanding of cross-question content across continuously changing video streams and multiple turns of interaction. Furthermore, the model should maintain a stable output structure, generate relational information beneficial for subsequent reasoning, and provide confidence levels consistent with the reliability of the judgments. Optimizing only a single objective can easily lead to the model overemphasizing local answers, resulting in imbalances in cross-question relationships and confidence level calibration. Therefore, this invention designs the total reward as a weighted combination of multiple reward items.

[0080] The policy optimization algorithm can be the Group Relative Policy Optimization (GRPO) algorithm. Let the old policy before the optimization update be denoted as . Based on the total reward, a relative comparison within the group can be performed to calculate the normalized relative advantage of the i-th candidate output within the group, denoted as . The GRPO objective function formula, which includes a clipping term and a KL divergence penalty term, is optimized to calculate the loss value L, and backpropagation is performed to update the baseline model parameters θ: ); in, is the hyperparameter pruning coefficient, used to limit the absolute magnitude of a single-step policy update; KL(.||.) represents the KL divergence between the output distributions of the old and new policies, and β is the weighting coefficient of the KL penalty term; This is a strategy for reference.

[0081] This invention generates multiple candidate output data and determines various reward constraints. The total reward is obtained by weighted summation of the reward constraints. Then, the baseline model parameters are updated based on the total reward using a policy optimization algorithm. This allows the model to take into account constraint signals from different dimensions, effectively avoiding the specific capability degradation and imbalance caused by a single optimization objective. It achieves robust and joint improvement of the overall output quality of the baseline model in complex scenarios.

[0082] In one embodiment, multiple reward constraints may include multiple types of format rewards, accuracy rewards, relational rewards, and confidence rewards; Format rewards are used to constrain video understanding models to strictly adhere to a predefined structured output format; Accuracy rewards are used to compare and verify whether the answer matches the true answer; Relationship reward is used to measure the contribution of generated relationship labels to subsequent problem reasoning; The confidence reward is used to constrain the confidence of the explicit output of the video understanding model to be consistent with the quality of the response.

[0083] The format reward constrains the model to strictly adhere to a predefined structured output format, checking the existence of all required fields, the correct order of fields, and the proper closure of labels. A positive reward is only given when the output fully meets the format requirements. The accuracy reward compares the last answer field with the true answer; a positive reward is given if the prediction is correct, and a negative reward is given if the prediction is incorrect. These two rewards respectively ensure the validity of the output format and the correctness of the content.

[0084] Confidence rewards are used to ensure that the confidence level of the model's explicit output aligns with the quality of the response. The model receives a higher reward when it provides the correct answer with higher confidence; conversely, it is penalized more severely for providing an incorrect answer with high confidence, thus encouraging the model to align its explicit confidence level with the quality of its actual response. Furthermore, if the model corrects an initially incorrect answer to the correct one after inference, it receives an additional reward; conversely, if inference corrects an initially correct answer to the incorrect one, the reward is reduced, thereby suppressing unnecessary or even harmful inference.

[0085] Relationship rewards are used to measure the generation of relationships. <relation>The contribution of a field to reasoning for subsequent questions. In multi-turn video question answering, subsequent questions often rely on observations from previous segments, so the model needs to maintain consistency in understanding the context between consecutive questions. <relation>The field is used to summarize the key information and time dynamics of the current segment, providing contextual clues for subsequent reasoning.

[0086] If multiple reward constraints can include format rewards, accuracy rewards, relational rewards, and confidence rewards, then the total reward... It can be represented as: ; in, , , and These represent the format reward, accuracy reward, relational reward, and confidence reward, respectively, with corresponding weights of 1 / 3. , , , .

[0087] This invention uses multiple rewards, including format rewards, accuracy rewards, relational rewards, and confidence rewards, as joint constraints for enhanced optimization. This comprehensively constrains and guides the model, enabling it to continuously evolve and optimize towards a high level of comprehensive excellence in parameter updates, characterized by legal and standardized format structure, correct content inference, coherence across problem contexts, and reliable self-confidence calibration.

[0088] In one embodiment, determining the confidence reward may further include: Determine the first correct state of the initial prediction part relative to the true answer in the candidate output data, and the second correct state of the final inference part relative to the true answer; Based on the second correct state, the output confidence in the candidate output data is calculated by offset product based on the center value to obtain the basic penalty term; The correction term is determined based on the state difference between the second correct state and the first correct state; The confidence reward is obtained by adding the base penalty and the correction.

[0089] The true answer is the objectively correct answer from the original supervised training dataset. The first correct state is denoted as... , is the extracted initial answer (i.e., the first one in the candidate output). <answer>The correctness of the content contained herein. .

[0090] The second correct state is denoted as This is the answer after reasoning through the model's thought process (i.e., the second candidate output). <answer>The correctness of the content contained herein. .

[0091] Output confidence is derived from candidate outputs. <confidence>The confidence score extracted from the field is denoted as c. [0,1]. The basic penalty term is represented as , where 0.5 is the specified center value. This numerical mechanism works as follows: when the final answer is correct ( When the confidence level c is 1, the product simplifies to (c-0.5). The higher the confidence level c, the larger the reward value. When the final answer is incorrect ( When c = 0), the product simplifies to -(c-0.5), and giving an incorrect answer under high confidence will be directly penalized.

[0092] The correction term is represented as When the model's inference successfully corrects the initially incorrect answer to the correct answer, that is... =1、 When the value is 0, a higher positive gain is obtained; if the reasoning changes the originally correct answer to an incorrect one, a deduction penalty will be imposed.

[0093] Confidence reward It can be represented as: ; This invention calculates the basic penalty term by multiplying the second correct state by the offset of the confidence level, and adds it to the correction term determined by the difference between the first and second correct states to obtain the confidence reward. This results in the model being penalized when it gives a high confidence level but actually answers incorrectly, and receiving additional positive gain when it successfully corrects the initial error. This encourages the model to strictly align its internal deterministic perception with its external numerical expression, effectively suppressing the tendency of the system to generate harmful and redundant thought chains, and improving the reliability of the confidence signal.

[0094] In one embodiment, determining the relationship reward may further include: Extract cross-question relationship labels contained in the candidate output data; Determine the first accuracy obtained by using cross-question relationship labels to perform reasoning verification on the next round of questions; The second average accuracy is obtained by performing reasoning verification on the next round of questions using the same set of comparative samples. The first accuracy rate is compared with the second average accuracy rate, and the corresponding relationship reward is determined based on the comparison result.

[0095] Cross-problem relationship labels are key elements contained in the candidate output data used to summarize the current and historical time-series dynamic characteristics. <relation>Field text content.

[0096] The first accuracy rate is denoted as The cross-question relationship label generated for the i-th candidate output sample, when passed as a prior historical context into the question verification environment of the (t+1)-th round, assists in the correct answer score or average probability. For example, if the relationship text "The pedestrian is crossing the center of the zebra crossing" extracted in the previous step is used as historical information and concatenated into the input context of the next round question "Can the vehicle be restarted now?", the evaluation model's answer conclusion is "No," which is completely consistent with the correct answer in the evaluation set, thus determining the first accuracy rate of this answer to be 1.0.

[0097] The same set of comparative samples consists of G distinct parallel exploration branch outputs generated by the current baseline model on the same input sample in the same batch and current state. The second average accuracy is denoted as... The second average accuracy is the system's average performance metric calculated after running the next round of tests using all the different relation labels generated in this group. All G relation labels generated in this group can be substituted into the next round of evaluation environment to generate G verification results concurrently. These results are then summed and averaged to obtain the second average accuracy. For example, if the baseline model generates 8 different relation label descriptions for the current image, and these 8 descriptions are substituted into the next round of testing for a specific question, resulting in 5 correct answers and 3 incorrect answers, then the second average accuracy is 5 / 8 = 0.625.

[0098] Relationship rewards can be defined as: By comparing the performance of a given sample on subsequent questions with the group average, it can be determined whether the cross-question relationships generated in the current round contribute positively to the reasoning in the next round. If the performance is higher than the group average, a positive reward is given; if it is lower than the group average, a negative reward is given; if they are equal, the reward is zero.

[0099] This invention extracts cross-question relationship labels and verifies them on the next round of questions to obtain the first accuracy rate. It then compares the first accuracy rate with the second average accuracy rate of the same group of comparative samples to determine the relationship reward. This clearly quantifies the substantial help and advantage that the information extracted in the current round brings to future reasoning. It can effectively guide and incentivize the model to actively capture and extract the temporal clue features that are most beneficial to the continuity of the context when processing continuous multi-round question-and-answer streams.

[0100] In one embodiment, determining the target answer for the current round of questions based on a comparison between the confidence level and a preset threshold may further include: If the confidence level is greater than or equal to a preset threshold, the initial predicted answer is extracted as the target answer result. If the confidence level is less than a preset threshold, the video understanding model continues to generate the thought chain reasoning process and the final reasoning answer for the current round of questions, and uses the final reasoning answer as the target answer result.

[0101] During the inference process, the model generates output sequentially according to a predefined output format. <relation>The first <answer>as well as <confidence> Field. When< / confidence> After generation, the confidence value c is extracted and compared with a preset threshold τ. If c ≥ τ, the initial predicted answer is returned directly to reduce additional inference overhead; if c < τ, subsequent predictions are generated. <think>field and the second <answer>The field will be used to select the second answer as the target response.

[0102] This invention, by comparing the confidence level with a preset threshold, either extracts the initial answer in advance or triggers the model to continue generating the thought chain and the final reasoning answer. Based on the model's perception of the internal certainty of the current problem, it adaptively decides whether to incur additional computational costs. This allows it to maintain the accuracy of rigorous reasoning when dealing with complex problems, while significantly reducing redundant computation in simple question-and-answer scenarios, thereby greatly improving the system's adaptive adjustment capability and overall response efficiency.

[0103] As summarized above, this invention addresses the issues of insufficient consistency in cross-question content understanding and difficulty in adaptively adjusting inference depth in streaming video scenarios. It constructs structured multi-turn inference data, introduces a cross-question relationship modeling mechanism, and employs a confidence-aware adaptive inference strategy. This enables the model to maintain inference coherence throughout continuous video input and multi-turn question interactions, and dynamically selects inference depth based on question complexity, thus balancing answer accuracy and response efficiency. Furthermore, this invention designs a multi-objective reinforcement learning training method combining format constraints, answer correctness, relationship consistency, and confidence calibration to enhance the model's comprehensive inference ability in streaming video understanding tasks. Experimental results demonstrate that, compared to existing technologies, this invention significantly improves inference consistency, answer accuracy, and inference efficiency in multi-turn question-answering scenarios in streaming video, validating the effectiveness and superiority of the proposed method.

[0104] This invention significantly improves the reasoning ability of video understanding models in streaming video multi-turn question answering tasks by introducing structured multi-turn inference data, cross-question relationship modeling mechanisms, multi-objective enhanced supervision fine-tuning, and confidence-driven adaptive inference mechanisms. Compared with existing technologies, this invention effectively solves the problem of inconsistent content understanding between questions in streaming video multi-turn question answering scenarios. Furthermore, this invention can dynamically adjust the inference depth according to question complexity, reducing redundant computation on simple questions while ensuring the quality of inference for complex questions, thereby improving the model's answer accuracy, inference efficiency, and practical application capabilities.

[0105] Comparison of this invention with other fixed-strategy methods Figure 4 As shown. To evaluate the effectiveness of this invention, the Qwen2.5-VL-7B-Instruct was used as the base model for training, and supervised fine-tuning and reinforcement learning training were completed using a full-parameter fine-tuning method. Training was performed on eight 80GB NVIDIA H800 GPUs, generating eight candidate results for each sample, with a temperature set to 1.0. Simultaneously, a dynamic frame sampling strategy was employed for the streaming video data, and format rewards, accuracy rewards, relational rewards, and confidence rewards were used jointly during training. To verify the effectiveness of the method, tests were conducted on several representative benchmarks. SVBench and OVO-Bench were used to evaluate real-time understanding, multi-turn question answering, long-term temporal information integration, and key event tracking capabilities in streaming video scenarios; MathVision and MMMU were used to evaluate mathematical visual reasoning and multidisciplinary multimodal understanding capabilities; MVBench and VideoMME were used to evaluate general video understanding capabilities; and VideoMMMU and MMVU were used to evaluate high-level reasoning capabilities in complex video scenarios.

[0106] Table 1. Comparison of the proposed method with other models on streaming video question answering benchmarks.

[0107] Table 2 Comparison of this method with other methods in general multimodal benchmark testing

[0108] Table 3. Comparison of accuracy and response time of different strategy models in various benchmark tests.

[0109] This invention is compared with several representative models. GPT-4v is a general-purpose visual language model with strong image and text understanding and question answering capabilities; GPT-4o is a multimodal large model that excels in complex visual reasoning and video understanding tasks; Gemini 1.5 Pro boasts strong long-context modeling and video understanding capabilities; Flash-VStream-7B is an online video understanding model optimized for streaming video scenarios; VideoLLM-online-8B is a video language model supporting online video input; Event-VStream is a model designed for event-level streaming video understanding tasks; ET-Instruct-3B is a video understanding model optimized by instructions; Video-R1-7B is a large-scale video model incorporating inference optimization strategies; VideoRFT-SFT-7B is a supervised fine-tuned video inference model; Qwen2.5-VL-7B is a high-performance general-purpose visual language foundation model; VideoChat-R1 is a model optimized for video dialogue and inference tasks; Time-R1 is a video model designed for temporal inference tasks; VidBridge-R1 is a multimodal model that balances video understanding and inference capabilities; InternLM-XC2.5 is a multimodal model with strong visual language understanding capabilities; LLaVA-OneVision is a widely used visual language model that performs well in image and video understanding tasks.

[0110] As shown in Tables 1 to 3, this invention demonstrates strong competitiveness across multiple benchmarks. In streaming video tasks, it significantly outperforms Qwen2.5-VL-7B and Video-R1-7B on OVO-Bench, indicating its clear advantages in cross-round temporal understanding and key attribute tracking. It also achieves high results on SVBench. In general image reasoning, video question answering, and video reasoning tasks, it also outperforms the base model, demonstrating good generalization ability. Furthermore, as shown in Table 3, this invention maintains high accuracy while also considering inference efficiency, demonstrating a good balance between performance and runtime overhead, thus validating the effectiveness of this invention.

[0111] The adaptive inference apparatus for multi-turn question answering of streaming video provided by the present invention will be described below. The adaptive inference apparatus for multi-turn question answering of streaming video described below can be referred to in correspondence with the adaptive inference method for multi-turn question answering of streaming video described above.

[0112] like Figure 5 As shown, the adaptive inference device for multi-turn question answering of streaming video provided by the present invention includes: The input module is used to input the current video segment of the streaming video, the current round question, and the historical question-answering context into the pre-trained video understanding model, and obtain the relationship label, initial predicted answer, and confidence score of the initial predicted answer output by the video understanding model for the current round question. The determination module is used to determine the target answer to the current round of questions based on the comparison between the confidence level and the preset threshold, and to merge the relationship label into the historical question-and-answer context of the next round.

[0113] Figure 6 The diagram illustrates the physical structure of an electronic device, which may include a processor, a communications interface, memory, and a communication bus. The processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute an adaptive reasoning method for multi-turn question answering of streaming video.

[0114] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0115] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the adaptive reasoning method for multi-turn question answering of streaming video provided by the above methods.

[0116] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an adaptive reasoning method for multi-turn question answering of streaming video provided by the methods described above.

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.< / answer> < / think> < / answer> < / relation> < / relation> < / confidence> < / answer> < / answer> < / relation> < / relation>

Claims

1. An adaptive reasoning method for multi-turn question answering in streaming video, characterized in that, include: The current video segment of the streaming video, the current round of questions, and the historical question-answering context are input into a pre-trained video understanding model to obtain the relational label, initial predicted answer, and confidence score of the initial predicted answer output by the video understanding model for the current round of questions. Based on the comparison between the confidence level and the preset threshold, the target answer result for the current round of questions is determined, and the relationship label is merged into the historical question-and-answer context of the next round.

2. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 1, characterized in that, The video understanding model is trained through the following steps: Multiple-choice data is generated based on raw streaming video question-and-answer samples; Structured multi-round inference data is constructed based on the multiple-choice data; The baseline model is trained under supervised supervision based on the structured multi-round inference data. The video understanding model is obtained by combining multiple reward constraints to train the baseline model after supervised fine-tuning training through reinforcement learning.

3. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 2, characterized in that, Multiple-choice data was generated based on the original streaming video question-answering samples, including: Based on the video content, corresponding questions, and real answers in the original streaming video question-and-answer sample, multiple interference options are generated to obtain a multi-choice sample to be filtered, which includes the real answers and the interference options; The multiple-choice data is obtained based on the sample of multiple-choice options to be screened.

4. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 3, characterized in that, The multiple-choice data is obtained based on the sample to be screened, including: The model was validated by independently performing multiple option prediction operations on the multiple-choice sample to be screened. Calculate the probability of correct prediction for the multiple option prediction operations; If the probability of a correct prediction is greater than a set threshold, multiple interference options are regenerated until the probability of a correct prediction is not greater than the set threshold. If the probability of a correct prediction is not greater than the set threshold, the sample to be screened for multiple selection is used as the multiple selection data.

5. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 2, characterized in that, Structured multi-turn inference data is constructed based on the multiple-choice data, including: Perform multiple independent inference operations on the samples to be processed in the multiple-choice data, and count the corresponding prediction frequency results; The initial answer and prediction confidence of the sample to be processed are determined based on the prediction frequency results. Based on the historical question-and-answer information of the sample to be processed, sample relationship labels are generated, and based on the sample relationship labels, reasoning operations are performed again to determine the thought chain process and verify the answer; The sample relationship labels, the initial sample answer, the sample prediction confidence, the thought process, and the verification answer are combined in a specific format to obtain the structured multi-round reasoning data.

6. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 2, characterized in that, The baseline model, after supervised fine-tuning training, is trained using reinforcement learning by combining multiple reward constraints, including: The baseline model is used to generate candidate output data for the input samples. Determine the various reward constraints of the candidate output data; The total reward is obtained by weighted summation of the various reward constraints. The parameters of the baseline model are updated based on the total reward using a policy optimization algorithm.

7. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 6, characterized in that, The various reward constraints include multiple types of format rewards, accuracy rewards, relational rewards, and confidence rewards; The format reward is used to constrain the video understanding model to strictly follow a predefined structured output format; The accuracy reward is used to compare and verify whether the answer matches the actual answer; The relation reward is used to measure the contribution of the generated relation labels to subsequent problem reasoning; The confidence reward is used to constrain the confidence level of the explicit output of the video understanding model to be consistent with the quality of the response.

8. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 7, characterized in that, Determining the confidence reward includes: Determine the first correct state of the initial prediction portion of the candidate output data relative to the true answer, and the second correct state of the final inference portion relative to the true answer; Based on the second correct state, the output confidence in the candidate output data is calculated by offset product based on the center value to obtain the basic penalty term; The correction term is determined based on the state difference between the second correct state and the first correct state; The confidence reward is obtained by adding the basic penalty term to the correction term.

9. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 7, characterized in that, Determining the relationship reward includes: Extract the cross-question relationship tags contained in the candidate output data; Determine the first accuracy obtained by using the cross-question relationship labels to perform reasoning verification on the next round of questions; The second average accuracy is obtained by performing reasoning verification on the next round of questions using the same set of comparative samples. The first accuracy rate is compared with the second average accuracy rate, and the corresponding relationship reward is determined based on the comparison result.

10. The adaptive reasoning method for multi-turn question answering in streaming video according to claim 1, characterized in that, Based on the comparison between the confidence level and a preset threshold, the target answer result for the current round of questions is determined, including: If the confidence level is greater than or equal to the preset threshold, the initial predicted answer is extracted as the target answer result; If the confidence level is less than the preset threshold, the video understanding model continues to generate the thought chain reasoning process and the final reasoning answer for the current round of questions, and the final reasoning answer is used as the target answer result.

Images