Interleaved visual-text chain-of-thought reasoning, generating intermediate visual artifacts to support deduction, is a promising paradigm for complex multimodal tasks. However, models suffer a representation tax causing systematic performance degradation when translating abstract reasoning into structured spatial visualizations. We attribute this to alignment bias in conventional supervised fine-tuning, which optimizes labels while neglecting synchronization between visual semantics and textual reasoning trajectories. To address this, we propose a pedagogical distillation framework grounded in the Explain-on-Blackboard (EoB) paradigm. Guided by multimedia learning and social constructivism, our method requires teacher models to generate temporally and semantically aligned teaching trajectories comprising a discourse channel for logical explanation and a canvas channel for incremental HTML/SVG construction. A dynamic-rubric quality assurance mechanism with multi-stage filtering retains only high-fidelity trajectories for distillation. To validate our approach, we construct Cas-Bench, a 1,000-problem multidisciplinary protocol. Experiments show an 8B-parameter model fine-tuned on 4,000 distilled trajectories achieves a 69.76% relative improvement on Cas-Bench over baselines across MMMU, MMMU-Pro, MathVision, and MathVista benchmarks.
Sentiment text classification is a natural language processing technique for recognizing and extracting subjective information in text, such as emotions, attitudes, and opinions. Traditional models such as BERT and XLNet have achieved brilliant results in text classification problems. With the rise of large language model technology, using generative large models to improve text classification accuracy has become a new research direction. However, compared with traditional classification models, large language models have a slight disadvantage in language classification tasks. In this paper, we propose a prompt-based approach to enhance the accuracy of large language models for text classification using image prompt information on multimodal datasets. First, we illustrate the principle of consistency between image and textual information. Second, we propose a multimodal framework for sentiment analysis of images and text, which realizes the prediction of sentiment tendency for both image and textual data by injecting the image prompt information into the text and into the Large Language Model. Finally, we designed experiments and evaluated them using real multimodal datasets to verify the effectiveness and accuracy of the framework.
Cancer, as one of the widely discussed contemporary topics, has attracted much attention. Its occurrence is closely related to economic development, lifestyle choice and environmental factors. Due to factors such as incidence, routine data set processing cannot be applied to data sets with low incidence. As a hot spot in the field of artificial intelligence, the performance of large-scale models has been greatly improved compared with its predecessors. As a lightweight model in the field of large models, Chatglm-6b not only shows some performance characteristics of large models, but also imposes a low resource demand on training..In the training stage, the lora fine-tuned training method is adopted, and the parallel model and Chatglm-6b share Chatglm-6b own tokenizer for serial training. In the inference phase, the parallel model will be run in parallel with Chatglm-6b, and the tensor generated by the Chatglm-6b model will be weighted fused with the same size generated by the parallel model before generating the output results. The model is intended to solve the problem of bias in the prediction results due to data imbalance during the small-sample disease risk prediction task . The model was validated on the real dataset of postoperative lung cancer complications provided by a hospital in Beijing and the public dataset of breast cancer prediction in kaggle.
Online learning is more and more popular because it is not limited by time and space. How to choose a suitable course from thousands of online courses is a great challenge faced by online learners, and online course recommendation came into being. The personalized recommendation algorithm analyzes the user’s preferences by collecting some previous historical records of the user and other information, and generates recommendations for the user. Since Wide&Deep was proposed, due to its inherent ease of implementation, adaptability, and versatility, this approach has gained significant traction across various industry sectors. But its feature intersection method is not efficient. Sufficient feature engineering is required to provide informative features that can effectively distinguish objects. In this paper, the WD-FM model is proposed by combining Wide&Deep and factorization machine, and good results have been achieved through experimental demonstration.
In the era of short videos, the task of sensitive video detection faces new challenges with the increasing diversity and quantity of videos in the network. Aiming at the problem that existing research methods are constrained by missing video comments and user subjectivity, a novel text enhancement method is proposed for sensitive video detection. Firstly, based on the CLIP pre-training model, image caption are generated for video frames. Then, through the integration of external common sense knowledge, the method extracts deep contextual information from the generated captions, including the underlying intentions and purposes conveyed by the text. Besides, considering the complementarity and redundancy between different sources of information, a multi-source data collaborative encoding mechanism and a multi-modal feature fusion mechanism are designed to achieve semantic feature alignment. Finally, state-of-the-art was achieved on the two public datasets NPDI-800 and Pornography-2k, and a large number of detailed comparison and ablation experiments were performed to verify the effectiveness of the method.
