Yubo Dong / Research portfolio
I study how language models reason, investigate complex questions, and work together, alongside collaborative work on efficient multimodal models. Here I share the projects, ideas, and tools behind that research.
A step-level view of language model reasoning. Cognitive tags and dependency graphs make reasoning traces easier to analyze, with MaxFlow and LCS guiding structure-aware optimization.
516 public examples · 23 cognitive step types · MIT dataset
Evaluating complex research through deep investigation, wide retrieval, and graph-anchored auditing of evidence and conclusions.
A graph-based framework for coordinating complex task dependencies. VillagerBench tests teamwork through building, cooking, and escape-room challenges.
Look at what you'll lose before pruning. A training-free approach that selects visual tokens hierarchically and transfers their information before removing them.
Training-free visual token pruning · Coarse-to-fine selectionBehind the projects
My research connects structured reasoning with autonomous investigation and multi-agent collaboration. My collaborative work also explores efficient vision-language models. WorkShare brings together the motivation, visual explanations, and resources for each project.
Explore the project notes for a closer look at the methods, or visit the dedicated websites for demos and benchmarks.
Follow the work on GitHub ↗