Video large language model illusion relieving method and system

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a video large language model illusion relieving method and system. The method comprises the following steps: inputting a video and a text into a video big language model, and starting generation of a plurality of candidate replies in parallel; in the process of generating each candidate reply, monitoring a visual attention degeneration point; upon detecting the visual attention recession point, stopping generation of the corresponding candidate reply, and calculating a timing attention collapse value of the corresponding candidate reply; and selecting the candidate reply with the maximum time sequence attention collapse value to continuously generate until the candidate reply is finished. According to the method, the video large language model can give more attention to the global content of the video, perception illusion is reduced, so that correct reply is made, additional training of the model is not needed, and the calculation cost is low.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method and system for alleviating hallucinations using a large language model in video. Background Technology

[0002] Currently, video big language models have made significant progress in tasks such as video content understanding, event reasoning, and cross-modal generation, and are gradually being applied to scenarios such as autonomous driving, robot control, and multimedia retrieval. However, existing video big language models generally suffer from perceptual illusions, meaning that the text content generated by the model is inconsistent with the actual video content. For example, the model may fabricate objects that do not exist in the video, incorrectly infer the order of actions, or ignore important semantic information in keyframes. This severely hinders their practical applications. Existing illusion mitigation techniques mainly fall into two categories: methods based on model fine-tuning and methods based on contrastive decoding.

[0003] Model fine-tuning methods train the model using additional labeled video data, making it more attentive to fine-grained visual cues. However, these methods are costly, heavily reliant on data, and prone to causing a "catastrophic forgetting" of the model's original reasoning abilities, leading to a decline in overall model performance and significantly impacting the model's reasoning and understanding capabilities.

[0004] Contrast-based decoding methods identify potential hallucinations by perturbing the input video and comparing the differences before and after. However, these methods have two drawbacks: first, they require separate computations for the original and perturbed videos, resulting in extremely high computational costs; second, when the perturbing strategy is inappropriate, it may mistakenly identify normal output as a hallucination, thereby compromising the model's comprehension ability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a video large language model hallucination relief method and system with strong comprehension capabilities and low computational cost.

[0006] The video large language model hallucination mitigation method provided by this invention includes the following steps:

[0007] (a) Input the video and text into the video large language model and start generating multiple candidate responses in parallel;

[0008] (ii) During the generation of each candidate response, monitor visual attention decay points;

[0009] (iii) Once the visual attention decay point is detected, the generation of the corresponding candidate response is stopped, and the temporal attention collapse value of the corresponding candidate response is calculated;

[0010] (iv) Select the candidate response with the largest temporal attention collapse value and continue generating until completion.

[0011] This invention allows video language models to pay more attention to the overall content of the video, reducing perceptual illusions and thus making correct responses. Moreover, this invention does not require additional training of the model and has low computational cost.

[0012] Optionally, the step of inputting the video and text into the video large language model and generating multiple candidate responses in parallel includes the following sub-steps:

[0013] (1) For video input, the video large language model uses a visual encoder to convert each frame of video into a set of visual lexical units, which are directly represented as a sequence of visual lexical units:

[0014] (2) For text input, the video big language model uses a word segmenter to convert the text into a sequence of text words;

[0015] (3) The video big language model concatenates the visual word sequence with the text word sequence to form the overall input sequence;

[0016] (4) Based on the overall input sequence, the video big language model generates multiple candidate responses in an autoregressive manner.

[0017] Optionally, monitoring visual attention decay points during the generation of each candidate response includes the following sub-steps:

[0018] When generating the kth word of the candidate response, the video big language model uses the overall input sequence and the previously generated k-1 words to sample the next word according to the conditional probability.

[0019] For each candidate response, calculate the cumulative visual attention of the k-th word to all the visual words, and simultaneously calculate the cumulative text attention to the previously generated text words.

[0020] Based on the sum of visual attention and the sum of text attention, if the k-th word satisfies the visual decay condition, then the k-th word is the visual decay point of the corresponding candidate response.

[0021] Optionally, once the visual attention decay point is detected, stopping the generation of the corresponding candidate response and calculating the temporal attention collapse value of the corresponding candidate response includes the following sub-steps:

[0022] If the k-th lexical unit is not the visual decay point, then continue generating the next lexical unit;

[0023] If the k-th lexical unit is the visual decay point, then stop generating the corresponding candidate response and calculate the temporal attention collapse value of the corresponding candidate response.

[0024] Optionally, the step of stopping the generation of the corresponding candidate response and calculating the temporal attention collapse value of the corresponding candidate response if the k-th lexical unit is the visual decay point includes the following sub-steps:

[0025] Based on the total visual attention of the nth candidate response up to the visual decay point, calculate its average cumulative attention value for each frame in the video;

[0026] The average cumulative attention value is normalized to obtain the frame-level attention distribution, and then the frame-level attention distribution is used to calculate the video frame attention collapse value.

[0027] The video is segmented into multiple video segments based on the semantic and action similarity between consecutive video frames;

[0028] The average cumulative attention value is used to calculate the cumulative attention value for each video segment;

[0029] The accumulated attention value is normalized to obtain the segment-level attention distribution, and then the segment-level attention distribution is used to calculate the attention collapse value of the video segment.

[0030] The attention collapse values ​​of the video frame and the video segment are added together to obtain the temporal attention collapse value of the corresponding candidate response.

[0031] Optionally, the attention collapse value of the video frame satisfies the following relationship:

[0032] ; in, When generating the nth candidate response for the video large language model, the first k-1 lexical units generated; The attention collapse value of the video frame; T is the number of video frames; Let be the frame-level attention distribution of the nth candidate response in the tth frame.

[0033] Optionally, the attention collapse value of the video segment satisfies the following relationship:

[0034] ; in, When generating the nth candidate response for the video large language model, the first k-1 lexical units generated; R is the attention collapse value of the video segment; R is the number of video segments; The segment-level attention distribution is the video segment of the r-th video segment in the n-th candidate response.

[0035] The present invention also provides a video large language model illusion relief system based on the above method, comprising: a data acquisition device, a data output device, a processor, and a storage device; the storage device includes a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by the processor, cause the processor to implement the video large language model illusion relief method provided by the present invention.

[0036] The present invention has at least the following beneficial effects:

[0037] 1. By generating multiple candidate responses in parallel, the model can explore more potential inference paths, which can alleviate the problem that existing video language models are easily limited by greedy decoding during the generation process and cannot escape local optima, resulting in frequent hallucinations.

[0038] 2. By calculating the temporal attention collapse value, we measure whether the video large language model focuses too much attention on local segments in the video when generating a candidate response, thereby determining the degree to which the output content contains hallucinations. Then, we select the response with the least hallucination from multiple candidate responses as the final response, reducing perceived hallucinations and thus making the correct response.

[0039] 3. By detecting visual decay points during the generation process and stopping the generation of corresponding candidate responses at these points, and directly calculating their temporal attention collapse index before generation is completed, the problem of drastically increasing computational costs caused by generating multiple candidate responses is solved by avoiding the continued generation of low-quality candidate responses.

[0040] 4. No additional training is required for the model, resulting in low computational cost.

[0041] 5. A system adapted to the method is provided, which can improve the practicality of the method and facilitate its promotion. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating a video-based large language model hallucination mitigation method according to the present invention.

[0043] Figure 2 This is a 16-frame video from an embodiment of the present invention.

[0044] Figure 3 This is an attention graph of Qwen2.5-VL under default settings and after using this solution, according to an embodiment of the present invention.

[0045] Figure 4This is a schematic diagram of the framework of a video large language model hallucination relief system according to an embodiment of the present invention. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0047] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0048] It should be noted in advance that, in one alternative embodiment, except for independent descriptions, the same symbols or letters appearing in all formulas have the same meaning and value.

[0049] In one optional embodiment, please refer to Figure 1 This invention provides a method for alleviating hallucinations using a large language model in video, the specific steps of which are as follows:

[0050] S1. Input the video and text into the video large language model to generate multiple candidate responses in parallel. This includes the following sub-steps:

[0051] S11. For video input, the video large language model uses a visual encoder to convert each frame of video into a set of visual lexical units and directly represents it as a sequence of visual lexical units.

[0052] Specifically, in this embodiment, the following is used: Let t be the video frame t, and let T be the input video frame t. The video big language model uses a visual encoder to encode each frame of video into M visual tokens, which are then directly represented as a sequence of visual tokens. ,Right now:

[0053]

[0054] in, is the visual word obtained by encoding the m-th region of the t-th frame of the video.

[0055] S12. For text input, the video large language model uses a word segmenter to convert the text into a sequence of text words.

[0056] Specifically, in this embodiment, for text input, the video large language model uses a word segmenter to convert the text into a sequence of text words. ,Right now:

[0057]

[0058] in, For the first L is the total number of text words.

[0059] S13. The video big language model concatenates the visual word sequence with the text word sequence to form a whole input sequence.

[0060] Specifically, in this embodiment, the overall input sequence F can be represented as:

[0061]

[0062] S14. Based on the overall input sequence, the video big language model generates multiple candidate responses in an autoregressive manner.

[0063] S2. During the generation of each candidate response, visual attention decay points are monitored, specifically including the following sub-steps:

[0064] S21. When generating the kth word of the candidate response, the video big language model uses the overall input sequence and the previously generated k-1 words to sample the next word according to the conditional probability.

[0065] Specifically, in this embodiment, when generating the kth word of the nth candidate response, the video big language model uses the overall input sequence F and the previously generated k-1 words. Based on conditional probability Sample the k-th word , It is a conditional probability mass function (PMF).

[0066] S22. For each candidate response, calculate the sum of visual attention of the k-th word to all visual words, and simultaneously calculate the sum of text attention to previously generated text words.

[0067] Specifically, in this embodiment, for the nth candidate response, the sum of visual attention of its kth word element to all visual word elements is... The following relationship must be satisfied:

[0068] ; in, When generating the kth word of the nth candidate response, the visual attention value of that word to the mth visual word from the tth frame.

[0069] For the nth candidate response, the sum of text attention of its kth word to the previously generated text words. The following relationship must be satisfied:

[0070] ; in, When generating the k-th term in the n-th candidate response, this term corresponds to the k-th term. The text attention value of each generated lexical unit.

[0071] S23. Based on the sum of visual attention and the sum of text attention, if the kth word satisfies the visual decay condition, then the kth word is the visual decay point of the corresponding candidate response.

[0072] Specifically, in this embodiment, based on the calculated visual attention sum and text attention sum, if the k-th word satisfies the visual decay condition, then the k-th word is the visual decay point of the corresponding candidate response, denoted as . ,Right now The aforementioned visual decay condition is:

[0073] ; in, When generating the nth candidate response for a video large language model, the first u-1 lexical units generated; For the nth candidate response, the sum of visual attention of its uth word element to all visual word elements; For the nth candidate response, the sum of text attention to the uth word of the previously generated text words is calculated. As the threshold, and ; Let be the step size, and .

[0074] S3. Once the visual attention decay point is detected, the generation of the corresponding candidate response is stopped, and the temporal attention collapse value of the corresponding candidate response is calculated. This specifically includes the following sub-steps:

[0075] S31. If the k-th lexical unit is not the visual decay point, then continue generating the next lexical unit.

[0076] Specifically, in this embodiment, if the kth word is not a visual decay point, then the (k+1)th word of the corresponding candidate response is generated.

[0077] S32. If the kth lexical unit is the visual decay point, then stop generating the corresponding candidate response and calculate the temporal attention collapse value of the corresponding candidate response.

[0078] Specifically, step S32 includes the following steps:

[0079] S321. Calculate the average cumulative attention value of the nth candidate response for each frame in the video based on all visual attention received by the candidate response up to the point of visual decay.

[0080] Specifically, in this embodiment, the average cumulative attention value of the nth candidate response to the tth frame of the video is... The following relationship must be satisfied:

[0081] ; in, When generating the nth candidate response for a video large language model, the generated first... Each word element; To generate the nth candidate response When there are m visual words, the visual attention value of that word to the m-th visual word from the t-th frame.

[0082] S322. Normalize the average cumulative attention value to obtain the frame-level attention distribution, and then use the frame-level attention distribution to calculate the video frame attention collapse value.

[0083] Specifically, in this embodiment, with Corresponding frame-level attention distribution The following relationship must be satisfied:

[0084] ;

[0085] After obtaining the frame-level attention distribution, the video frame attention collapse value is calculated using the frame-level attention distribution. The attention collapse value of a video frame satisfies the following relationship:

[0086]

[0087] like The smaller the value, the more the attention of the video's large language model is concentrated on a few frames, and the more severe the frame-level attention collapse phenomenon is.

[0088] S323. Based on the semantic and action similarity between consecutive video frames, the video is divided into multiple video segments.

[0089] Specifically, in this embodiment, the video is segmented into several video segments based on the semantic and action similarity between consecutive video frames. Specifically, the following cost matrix is ​​used to measure the semantic and action similarity between consecutive video frames:

[0090]

[0091] in, Let be the cost matrix of the j-th image block in the t-th frame of the video and the i-th image block in the (t-1)-th frame of the video. Let be the visual embedding vector of the i-th image patch in the (t-1)-th frame of the video. Let be the visual embedding vector of the j-th image patch in the t-th frame of the video. Let be the normalized spatial coordinates of the j-th image block in the t-th frame of the video. Let be the normalized spatial coordinates of the i-th image block in the (t-1)-th frame of the video.

[0092] Then, based on the obtained cost matrix, and assuming one-to-one matching of image blocks in adjacent video frames, the distance between frame t and frame (t-1) is obtained through optimal matching. ,Right now:

[0093] ;

[0094] ; Where N is the total number of image patches, The image block that minimizes the total matching cost between the i-th image block in the t-th frame of the video and the i-th image block in the (t-1)-th frame of the video is identified. For the t-th frame of the video The cost matrix of each image block and the i-th image block in the (t-1)-th frame of video. This represents all possible matching methods for image blocks in the t-th and (t-1)-th frame of the video.

[0095] like Greater than the distance threshold , If the difference between frame t and frame (t-1) is significant in semantics and action, then frame t is taken as the starting point for a new video segment. This method divides the entire video into R video segments. Furthermore, this similarity calculation considers not only physical similarity but also the degree of dynamic change between adjacent frames, making the segmentation of video segments more accurate.

[0096] S324. Calculate the attention accumulation value for each video segment using the average attention accumulation value.

[0097] Specifically, in this embodiment, when the video big language model generates the nth candidate response, the cumulative attention value for the rth video segment is... The following relationship must be satisfied:

[0098] ;

[0099] S325. Normalize the accumulated attention value to obtain the segment-level attention distribution, and then use the segment-level attention distribution to calculate the attention collapse value of the video segment.

[0100] Specifically, in this embodiment, with Corresponding fragment-level attention distribution The following relationship must be satisfied:

[0101] ;

[0102] After obtaining the segment-level attention distribution, the attention collapse value of the video segment is calculated using the segment-level attention distribution. The attention collapse value of a video clip satisfies the following relationship:

[0103] ;

[0104] like The smaller the value, the more the attention of the video language model is concentrated in a small number of time periods, and the more severe the segment-level attention collapse phenomenon is.

[0105] S326. Add the attention collapse value of the video frame and the attention collapse value of the video segment to obtain the temporal attention collapse value of the corresponding candidate response.

[0106] Specifically, in this embodiment, the attention collapse value of the video frame and the attention collapse value of the video segment of the nth candidate response are added together to obtain its temporal attention collapse value. ,Right now:

[0107] ;

[0108] The larger the value, the more thoroughly the candidate response references the video content, and the lower the probability of hallucination.

[0109] S4. Select the candidate response with the largest temporal attention collapse value and continue generating until completion.

[0110] Specifically, in this embodiment, after calculating the temporal attention collapse value of all candidate responses, the video big language model selects the candidate response with the largest temporal attention collapse value to continue generating until completion, and outputs the generated result as the final result.

[0111] This embodiment compares the performance of the proposed solution with cutting-edge hallucination mitigation methods, including Dinoheal (CVPR'25), TCD (arXiv'25), and VCD (CVPR'24), on three video large language models: Qwen2.5-VL, Video-R1, and LLaVA-NEXT. The proposed solution is systematically evaluated from two perspectives: hallucination suppression capability and understanding and reasoning ability. First, to measure the hallucination suppression capability of the proposed solution compared to existing hallucination mitigation methods on video large language models, the VRIPT-HAL dataset (a video description generation task) and the EventHallusio dataset (based on video question answering) were used for testing video description generation and video question answering. The test results are shown in Table 1. Second, to comprehensively evaluate video understanding and reasoning capabilities, two large-scale datasets containing long videos and complex scenes, namely the Video-MME dataset and the Video-MMMU dataset, were used for testing. The test results are shown in Table 2.

[0112] Table 1. Performance of different hallucination relief methods in terms of hallucination suppression ability.

[0113] .

[0114] Table 2. Performance of different hallucination relief methods on video comprehension and reasoning abilities.

[0115] .

[0116] In Tables 1 and 2, "Qwen2.5_Dinoheal" indicates Qwen2.5-VL using Dinoheal, and the meanings of "Qwen2.5_TCD", "Qwen2.5_VCD", and "Qwen2.5_this solution" follow the same logic; "Videor1_Dinoheal" indicates Video-R1 using Dinoheal, and the meanings of "Videor1_TCD", "Videor1_VCD", and "Videor1_this solution" follow the same logic; "LLaVA_Di "noheal" indicates the use of Dinoheal's LLaVA-NEXT; the meanings of "LLaVA_TCD", "LLaVA_VCD", and "LLaVA_this scheme" follow the same logic; R is recall, P is precision, and Acc is accuracy; "long", "medium", and "short" describe the video duration in the Video-MME dataset; "Pere.", "Compe.", and "Adap." are three subsets of the Video-MMMU dataset, used to evaluate the perception, understanding, and reasoning abilities of the video large language model, respectively.

[0117] As shown in Table 1, Dinoheal, TCD, and VCD generally exhibit insufficient cross-model transferability in terms of hallucination suppression. They often only achieve limited improvement on the LLaVA-NEXT model, while frequently experiencing performance degradation on Qwen2.5-VL and Video-R1. In contrast, our proposed solution can suppress hallucinations in both video description generation and video question answering tasks.

[0118] As shown in Table 2, Dinoheal, TCD, and VCD tend to significantly reduce the video understanding and reasoning capabilities of the video big language model. In contrast, our proposed solution achieves a consistent improvement in both video understanding and reasoning capabilities across all three video big language models.

[0119] Therefore, this solution not only effectively suppresses the illusions in the perception process of the video big language model, but also improves the video understanding and reasoning ability of the video big language model.

[0120] Furthermore, Figure 2The video presented 16 frames with the question: What was the weather like when the helicopter took off? Based on the keyframes (frames 14-16) showing the helicopter taking off from the platform, and the overcast background sky, the correct answer is "rainy." However, Qwen2.5-VL's answer is: "...no weather conditions are explicitly shown (such as rain or any specifically mentioned weather conditions). Since the image does not provide evidence related to the actual helicopter takeoff scene, the correct answer cannot be determined solely from the image content...therefore, among the given options, the most appropriate answer is cloudless." Applying this approach to Qwen2.5-VL yields the answer: "The scene in the video shows a helicopter taking off from the platform. The sky appears cloudy and overcast. Based on these observations, the correct answer is: rainy."

[0121] As can be seen, Qwen2.5-VL, under its default settings, overemphasizes the first frame and three nearly identical last frames, thus ignoring key content in the video, leading to illusions and incorrect predictions. This proposed solution, however, allows the model to pay more attention to the overall content of the video, resulting in the correct response.

[0122] Furthermore, Figure 3 Images (a) and (b) show the attention graphs of Qwen2.5-VL under default settings and after using this approach, respectively. Figure 3 In (a) of the model, the purple box indicates that the model is overly focused on a single frame and may ignore other keyframes; the red box indicates that the model is overly focused on local segments and has not considered the environmental context. Figure 3 In (b) of the above, the attention distribution is obviously more balanced, which can integrate information from multiple frames and achieve more accurate reasoning.

[0123] It should be noted that in some cases, the actions described in the specification can be performed in different orders and still achieve the desired results. In this embodiment, the order of steps is given only to make the embodiment clearer and easier to explain, and not to limit it.

[0124] In one optional embodiment, please refer to Figure 4 To improve the practicality of this method and facilitate its promotion, the present invention also provides a video large language model illusion relief system. The video large language model illusion relief system includes: a data acquisition device 1, a data output device 2, a processor 3, and a storage device 4. The storage device 4 includes a computer-readable storage medium storing a computer program. The computer program includes program instructions, which, when executed by the processor 3, cause the processor 3 to perform the content described in steps S1 to S4.

[0125] In summary, this invention has at least the following beneficial effects: Firstly, by generating multiple candidate responses in parallel, this method explores more potential inference paths for the model, alleviating the problem of frequent hallucinations caused by greedy decoding limitations and the inability to escape local optima in existing video large language models during generation. Secondly, this method measures whether the video large language model focuses excessive attention on local segments in the video when generating a candidate response by calculating the temporal attention collapse value, thereby determining the degree of hallucination in the output content and selecting the response with the least hallucination from multiple candidate responses as the final response, reducing perceived hallucinations and thus making the correct response. Thirdly, this method avoids continuing to generate low-quality candidate responses by detecting visual decay points during generation and stopping the generation of corresponding candidate responses at these points, and directly calculating their temporal attention collapse index even before generation is complete, thus solving the problem of a sharp increase in computational cost caused by generating multiple candidate responses. Fourthly, this method does not require additional model training and has low computational cost. Fifthly, it provides a system adapted to the method, which can improve the practicality of this method and facilitate its promotion.

[0126] 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for hallucination mitigation of a video large language model, characterized in that, The specific steps are: (1) input video and text into a video large language model, and start generating multiple candidate replies in parallel; (2) during the generation of each candidate reply, monitor the visual attention decay point; (3) once the visual attention decay point is detected, stop the generation of the corresponding candidate reply, and calculate the timing attention collapse value of the corresponding candidate reply; (4) select the candidate reply with the largest timing attention collapse value to continue generating until completion.

2. The video large language model delusion mitigation method of claim 1, wherein, The input of video and text into the video large language model and the start of the generation of multiple candidate replies in parallel include the following sub-steps: (1) for video input, the video large language model uses a visual encoder to convert each frame of video into a set of visual tokens and directly represents it as a visual token sequence: (2) for text input, the video large language model uses a tokenizer to convert the text into a text token sequence; (3) the video large language model concatenates the visual token sequence and the text token sequence to form an overall input sequence; (4) according to the overall input sequence, the video large language model generates multiple candidate replies in a self-recursive manner.

3. The video large language model delusion mitigation method of claim 2, wherein, The monitoring of the visual attention decay point during the generation of each candidate reply includes the following sub-steps: (1) when the kth token of the candidate reply is generated, the video large language model uses the overall input sequence and the previously generated k-1 tokens to sample the next token according to the conditional probability; (2) for each candidate reply, calculate the visual attention cumulative sum of the kth token to all visual tokens, and calculate the text attention cumulative sum to the previously generated text tokens; (3) according to the visual attention cumulative sum and the text attention cumulative sum, if the kth token meets the visual decay condition, the kth token is the visual decay point of the corresponding candidate reply.

4. The video large language model delusion mitigation method of claim 1, wherein, Once the visual attention decay point is detected, the generation of the corresponding candidate reply is stopped, and the timing attention collapse value of the corresponding candidate reply is calculated, including the following sub-steps: (1) if the kth token is not the visual decay point, continue to generate the next token; (2) if the kth token is the visual decay point, stop generating the corresponding candidate reply, and calculate the timing attention collapse value of the corresponding candidate reply.

5. The video large language model delusion mitigation method of claim 4, wherein, If the kth token is the visual decay point, stop generating the corresponding candidate reply, and calculate the timing attention collapse value of the corresponding candidate reply, including the following sub-steps: (1) according to the total visual attention of the nth candidate reply before generating to the visual decay point, calculate its average attention accumulation value to each frame of video; (2) normalize the average attention accumulation value to obtain a frame-level attention distribution, and then use the frame-level attention distribution to calculate the video frame attention collapse value; (3) based on the similarity between consecutive video frames in semantics and actions, the video is divided into multiple video segments; (4) use the average attention accumulation value to calculate the attention accumulation value of each video segment; (5) normalizing the attention accumulation value to obtain a segment-level attention distribution, and then using the segment-level attention distribution to calculate a video segment attention collapse value; (6) adding the video frame attention collapse value and the video segment attention collapse value to obtain a corresponding time sequence attention collapse value of the candidate reply.

6. The video large language model delusion mitigation method of claim 5, wherein, The video frame attention collapse value satisfies the following relationship: ； wherein, the first k-1 wordpieces generated when generating the nth candidate reply of the video large language model; is the video frame attention collapse value; T is the number of video frames; is the frame-level attention distribution of the tth frame in the nth candidate reply.

7. The video large language model delusion mitigation method of claim 5, wherein, The video segment attention collapse value satisfies the following relationship: ； wherein, the first k-1 wordpieces generated when generating the nth candidate reply of the video large language model; is the attention collapse value for the video segment; R is the number of video segments; is the segment-level attention distribution for the rth video segment in the nth candidate reply.

8. A video large language model delusion mitigation system, characterized in that, The video large language model illusion relief system includes a data acquisition device, a data output device, a processor, and a storage. The storage includes a computer readable storage medium, and the computer readable storage medium stores a computer program. The computer program includes program instructions, and the program instructions, when executed by the processor, cause the processor to implement the video large language model illusion relief method of any one of claims 1-7.

Images