Reproduce “DB3 Team's Solution For Meta KDD Cup' 25” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yikuan Xia, Jiazun Chen, Yirui Zhan, Suifeng Zhao, Weipeng Jiang, Chaorui Zhang, Wei Han, Bo Bai and 1 more
The db3 team developed a multi-modal retrieval-augmented generation (RAG) framework to address the Meta CRAG-MM Challenge 2025, which involved complex multi-turn factual question answering with ego-centric images. Their solution secured top rankings across all three tasks and the grand prize by integrating tailored retrieval pipelines with LLM tuning for factual accuracy and controlled refusal.
Key insights
- Tailored retrieval pipelines and advanced query rewriting (e.g., "merged text query") are crucial for effectively processing diverse multi-modal inputs and bridging them to different knowledge sources.
- Reinforcement learning algorithms like DPO and GRPO are effective for training LLMs to control hallucination and judiciously refuse answers in competitive scenarios with penalties for incorrect responses.
- Ego-centric images pose unique challenges, necessitating specialized visual grounding techniques (like Grounding-DINO) to accurately interpret entities within dynamic, first-person contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.