Multi-modal large model illusion relieving method based on multi-view multi-path reasoning

Abstract

The invention discloses a multi-modal large model illusion relieving method based on multi-view multi-path reasoning. The method comprises the following steps: acquiring an input image and a text query of a user; the image is sent into a visual encoder of a visual language model LVLM to be processed, and a visual mark is obtained; processing the text query into a text mark through a word segmentation device of the LVLM; obtaining multi-view information; converting the multi-view description into a mark, and combining the mark with an original visual mark and a text mark to form an enhanced input sequence; 5) inputting the enhanced input sequence into the LVLM for decoding, activating a multi-path reasoning mechanism, and distributing an independent decoding path for information of each view angle; 6, the overall deterministic scores of all the potential answers are ranked, and the answer with the highest deterministic score is selected as the final output. According to the LVLM text output generation method, through cooperative work of the two strategies of multi-view information acquisition and multi-path reasoning, illusion of the LVLM during text output generation can be effectively reduced.

Description

Technical Field

[0001] This invention relates to artificial intelligence technology, and more particularly to a multimodal large model hallucination mitigation method based on multi-view, multi-path reasoning. Background Technology

[0002] Currently, large-scale visual language models (LVLMs) have demonstrated powerful capabilities in multimodal context understanding, but they also face the "illusion" problem, where the model generates text output inconsistent with the image content. To address this issue, researchers have proposed several solutions, which can be broadly categorized into three types: The first category is retraining-based approaches. These approaches primarily rely on supervised fine-tuning of the SFT or reinforcement learning from human feedback, retraining the model by constructing a dataset related to hallucinations. While they have achieved some success in mitigating hallucinations, they require a large number of training samples, are time-consuming and labor-intensive, demand significant computational resources, and carry the risk of overfitting.

[0003] The second category is based on external tools. These approaches explore training-free paradigms to mitigate hallucinations, typically relying on external tools to assist the model's output. For example, some models identify and correct hallucinations from generated text, or utilize classifier-free guidance to incorporate additional object-grounding features to improve the accuracy of visual language model generation. While these methods do not require retraining the model, their heavy reliance on complex external tools increases system complexity and introduces a dependency on external tools, limiting the model's independence and applicability.

[0004] The third category is based on single-viewpoint information extraction schemes. These schemes focus on extracting information from a single viewpoint of the input image, such as using models like CLIP to obtain a global image representation. While these methods can capture overall information about the image, they may overlook complex, object-specific details and background components, leading to a partial understanding of the input image and thus creating illusions when asked about details not captured by the video encoder. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multimodal large model illusion mitigation method based on multi-view and multi-path reasoning, which addresses the shortcomings of the existing technology.

[0006] The technical solution adopted by this invention to solve its technical problem is: a multimodal large-model hallucination mitigation method based on multi-view and multi-path reasoning, comprising the following steps: 1) Obtain the input image and the user's text query; 2) The image is fed into the visual encoder built into the Visual Language Model (LVLM) for processing to obtain visual tags; the text query is processed into text tags by the LVLM's tokenizer. 3) Multi-perspective information acquisition The system generates multi-perspective descriptions using designed prompts; perspectives include: bottom-up, regular, and top-down. 4) Convert multi-perspective descriptions into tags and combine them with the original visual and text tags to form an enhanced input sequence; 5) The enhanced input sequence is fed into the LVLM for decoding, activating the multi-path inference mechanism and assigning an independent decoding path to the information from each perspective, as follows: Multiple decoding paths are introduced for each information perspective; for each perspective, the model considers the top K most likely labels and generates multiple decoding paths based on these labels, thereby obtaining multiple possible answers; Each decoding path corresponds to a potential answer, and each answer has a deterministic score; Then, for the same answer, the model aggregates the deterministic scores from multiple decoding paths to obtain the overall deterministic score for that answer.

[0007] 6) Sort all potential answers by their overall certainty score and select the answer with the highest certainty score as the final output.

[0008] According to the above scheme, in step 3), a multi-perspective description is generated through the designed prompts, as follows: The bottom-up perspective provides pixel-level detail of the image, guiding LVLM to generate detailed descriptions including shape, color, and texture through prompts; A standard perspective provides a basic descriptive summary of the image, while prompts guide LVLM to generate a core content description of the image; The top-down perspective provides the scene background, environmental factors, and relevant visual cues of the image, guiding LVLM to generate a contextual description through prompts.

[0009] According to the above scheme, in step 5), the deterministic score is calculated by quantifying the probability difference between the first two most likely tags in each decoding step.

[0010] The beneficial effects of this invention are: This invention fully utilizes the inherent capabilities of LVLM without using additional external tools or introducing additional training costs. At the same time, by working together with two strategies, multi-view information acquisition and multi-path reasoning, it can effectively reduce the illusions that LVLM generates when producing text output.

[0011] This invention comprehensively perceives images from three different perspectives: "bottom-up," "conventional," and "top-down," compensating for fine details and background information that visual encoders may miss. A multi-path reasoning and deterministic-driven decoding mechanism quantifies the deterministic score of each potential answer, thereby selecting the answer with the highest confidence. Attached Figure Description

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0014] like Figure 1 As shown, a multimodal large-model hallucination mitigation method based on multi-view, multi-path reasoning includes the following steps: 1) Obtain the input image and the user's text query; 2) The image is fed into the visual encoder (e.g., CLIP or BLIP) built into the visual language model LVLM for processing to obtain visual tags; the text query is processed into text tags by the LVLM tokenizer. 3) Multi-perspective information acquisition The system generates multi-perspective descriptions using designed prompts; perspectives include: bottom-up, regular, and top-down. 4) Convert multi-perspective descriptions into tags and combine them with the original visual and text tags to form an enhanced input sequence; 5) The enhanced input sequence is fed into the LVLM for decoding, activating the multi-path inference mechanism and assigning an independent decoding path to the information from each perspective, as follows: Multiple decoding paths are introduced for each information perspective; for each perspective, the model considers the top K most likely labels and generates multiple decoding paths based on these labels, thereby obtaining multiple possible answers; Each decoding path corresponds to a potential answer, and each answer has a deterministic score, which is calculated by quantifying the probability difference between the first two most likely labels in each decoding step. Then, for the same answer, the model aggregates the deterministic scores from multiple decoding paths to obtain the overall deterministic score for that answer.

[0015] 6) Sort all potential answers by their overall certainty score and select the answer with the highest certainty score as the final output.

[0016] Experimental evaluation For evaluating the multi-view, multi-path reasoning (MVP) framework of this invention, we used two main benchmark datasets: POPE (Polling-based Object Probing Evaluation) and MME (Multimodal Model Evaluation). The POPE benchmark primarily evaluates the illusion problem of large visual language models (LVLMs) regarding object presence, while MME is a more comprehensive benchmark used to evaluate the performance of LVLMs across multiple dimensions, including perceptual and cognitive abilities, with a particular focus on object-level and attribute-level illusions. The POPE benchmark datasets are derived from three different sources: MSCOCO (Microsoft Common Objects in Context), A-OKVQA (A-OK Visual Question Answering), and GQA (Grounded Question Answering). Each dataset contains 500 images and employs three different settings for constructing negative samples: random, popular, and adversarial. The random setting randomly selects negative samples, the popular setting selects common negative samples, and the adversarial setting selects indistinguishable negative samples. These settings aim to comprehensively evaluate the model's performance in different scenarios. The MME benchmark includes 10 perception-related subtasks and 4 cognition-related subtasks. Furthermore, MME specifically focuses on evaluating object-level illusions (including presence and counting subtasks) and attribute-level illusions (including location and color subtasks). In terms of experimental settings, the MVP framework was tested on four state-of-the-art LVLM models: LLaVA1.5, Qwen-VL, and InstructBLIP. For fair comparison, MVP was compared with the original LVLM models and two recently proposed training-free methods: VCD (Visual Contrastive Decoding) and OPERA (Over-trust Penalty and Retrospection-Allocation). The experimental results are shown in Table 1. In the POPE benchmark, MVP significantly outperformed the original models on all LVLM models. For example, on the LLaVA1.5 model, MVP achieved an average accuracy improvement of 15.9 and an average F1 score improvement of 21.84 in random, popular, and adversarial settings. Furthermore, MVP achieved better results in most cases compared to VCD and Opera.On the illusion subset of the MME benchmark, MVP significantly improves the performance of all LVLM models in both object-level and attribute-level illusions, outperforming VCD and Opera. On the full MME dataset, models using MVP outperform the original models on most evaluation subsets. Experimental analysis shows that the multi-view information acquisition strategy enables the model to perceive images from different dimensions, thus better understanding image content. For example, the LLaVA-1.5 model identifies an average of 16.43 objects when using only the regular viewpoint, while it identifies an average of 36.66 objects when using three viewpoints. Furthermore, the multi-path reasoning mechanism effectively improves the model's accuracy during the decoding process. By quantifying and aggregating the deterministic scores of each potential answer, the model can select the answer with the highest confidence, thereby reducing illusions caused by uncertainty in answer labeling.

[0017] Table 1 Comparison of different methods on MME

[0018] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A multimodal large-scale hallucination mitigation method based on multi-view, multi-path reasoning, characterized in that, Includes the following steps: 1) Obtain the input image and the user's text query; 2) The image is fed into the visual encoder built into the Visual Language Model (LVLM) for processing to obtain visual tags; The text query is processed into text tokens using the LVLM tokenizer; 3) Acquisition of information from multiple perspectives; Generate multi-perspective descriptions through designed prompts; Perspectives include: bottom-up, conventional, and top-down; 4) Convert multi-perspective descriptions into tags and combine them with the original visual and text tags to form an enhanced input sequence; 5) The enhanced input sequence is fed into the LVLM for decoding, activating the multi-path inference mechanism and assigning an independent decoding path to the information from each perspective, as follows: Multiple decoding paths are introduced for each information perspective; for each perspective, the model considers the top K most likely labels and generates multiple decoding paths based on these labels, thereby obtaining multiple possible answers; Each decoding path corresponds to a potential answer, and each answer has a deterministic score; Then, for the same answer, the model aggregates the deterministic scores from multiple decoding paths to obtain the overall deterministic score for that answer; 6) Sort all potential answers by their overall certainty score and select the answer with the highest certainty score as the final output.

2. The multimodal large-model hallucination mitigation method based on multi-view, multi-path reasoning according to claim 1, characterized in that, In step 3), a multi-perspective description is generated through the designed prompts, as follows: The bottom-up perspective provides pixel-level detail of the image, guiding LVLM to generate detailed descriptions including shape, color, and texture through prompts; A standard perspective provides a basic descriptive summary of the image, while prompts guide LVLM to generate a core content description of the image; The top-down perspective provides the scene background, environmental factors, and relevant visual cues of the image, guiding LVLM to generate a scene-based description through prompts.

3. The multimodal large-model hallucination mitigation method based on multi-view, multi-path reasoning according to claim 1, characterized in that, In step 5), the deterministic score is calculated by quantifying the probability difference between the first two most likely tags in each decoding step.

4. An electronic device, characterized in that, include: One or more processors; as well as Storage device for storing one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors perform the method according to any one of claims 1 to 3.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 3.

Images