Where multimodal AI papers come with homework done

What it does

This repository is a curated index of papers on multimodal large language models, LLMs, and diffusion models, maintained by a final-year Ph.D. student at the Chinese Academy of Sciences. Unlike a typical awesome-list, most entries link directly to the author’s detailed reading notes—often described as ten-thousand-word essays—hosted on Zhihu. It doubles as a project page for his own work, including benchmarks like MME-RealWorld and models such as SliME.

The interesting bit

The notes dissect technical specifics like reward-model tricks and RL training recipes rather than offering surface-level summaries. The curation is also aggressively current, with May 2025 updates already tracking agentic RL and “think with image” papers from Seed, DeepSeek, and DeepMind.

Key highlights

• Most papers link to the author’s personal reading notes rather than leaving you with a bare citation.
• Covers active subfields: agentic RL for MLLMs, multimodal reward models, unified understanding/generation, and long-context video models.
• Doubles as a project page for the author’s own research, including the MME-RealWorld benchmark and the SliME model.
• Updated frequently; the May 2025 batch adds recent RL and agentic reasoning studies.
• Organized into thematic threads like “Think with Images” and “Alignment With Human Preference.”

Caveats

• Nearly all reading notes are in Chinese and hosted on Zhihu, so non-Chinese readers will need translation.
• The scope and tone reflect a single researcher’s interests rather than a community-curated consensus.

Verdict

Worth bookmarking if you are trying to catch up on recent MLLM and agentic reasoning trends and can navigate Chinese text (or trust browser translate). Skip it if you need executable code or an English-only resource.