Multimodal and Large Language Model Paper List: A Researcher's Daily arXiv Reading Tracking

This article introduces the open-source GitHub repository Multimodal-AND-Large-Language-Models maintained by Yangyi-Chen, which aims to address the reading dilemma caused by the explosion of papers in the AI field. By recording the author's daily reading trajectory on arXiv, it systematically tracks cutting-edge research in the intersection of multimodal and large language models. Its core value lies in its personalized and real-time characteristics, providing efficient literature screening and learning references for researchers, engineers, and beginners.

Papers in the field of artificial intelligence are growing rapidly; the cs.CL and cs.CV sections of arXiv add over a hundred new papers daily. Traditional literature management methods (Google Scholar alerts, email subscriptions, scattered collections) struggle to cope with the information flood, so researchers urgently need systematic tools to organize their reading workflow.

Yangyi-Chen's repository is an open-source GitHub project that records the author's daily reading of papers in the intersection of multimodal and LLM fields on arXiv. Unlike general reviews, it is the reading trajectory of a real researcher, reflecting the evolution of paper screening preferences and focus over time, with the uniqueness of being personalized and real-time.

The repository covers four core directions:

1. Vision-language pre-training models: Classic works like CLIP and ALIGN, as well as techniques such as alignment and contrastive learning;
2. Multimodal large language models: From VisualBERT and ViLBERT to Flamingo, BLIP-2, LLaVA, including vision-language docking and instruction fine-tuning;
3. Cross-modal understanding and generation: Image-text bidirectional generation, application of diffusion models, cross-modal reasoning;
4. Multimodal extensions of LLM: External tool calls (e.g., GPT-4V), visual module embedding, multimodal prompt engineering.

Organization Method: Classified by topic (model architecture, training methods, etc.), recorded in chronological order (preserving temporal development), and concise metadata (title, authors, links, short notes). Target Audience:

• Researchers: Quickly understand the current state of the field and get access to screened papers;
• Industry engineers: Grasp technical trends and evaluate practical directions;
• Beginners: Learn literature screening and organization methods as a starting point for reading. Usage Suggestions: Regularly browse updates to discover papers, observe structural evolution to understand the field's context, and use it as a basis for discussion.

The repository has limitations such as personal preference (reflecting the author's interests, some subfields may be overlooked), timeliness lag (needs to be combined with daily arXiv updates and academic social platforms), and limited depth (short introduction per paper, need to consult the original text). It is recommended to use it as a reference rather than an authoritative guide, and combine it with other resources.

Inspiration for Personal Literature Management: Establish a screening mechanism, share publicly to get feedback, keep it lightweight and easy to maintain, and review regularly to update knowledge structure. Future Trends in the Field: Unified multimodal architecture, efficient training methods, long-context cross-modal reasoning, accelerated transformation of research prototypes into products. Summary: This repository is a practical strategy to deal with information overload, providing researchers with literature resources and learning method demonstrations, helping to improve knowledge management capabilities.