Best Paper Award @ GLOW WWW '26
Best Paper Award @ GLOW WWW '26
RAG methodologies are often used to ground LLM responses, with knowledge graphs like Wikidata being popular options for information sources. Using traditional graph RAG methods for incorporating highly connected nodes can quickly result in heavy token consumption and filling up context. ConceptFormer solves this by learning fixed-length concepts that represent 1-hop neighborhoods of nodes. This drastically reduces token consumption while matching performance of traditional methodologies.
ConceptFormer started as a Master's Thesis at the University of Zurich and got awarded the Best Paper award on the 1st Workshop on Graph-enhanced LLMs for trustworthy Web data management (GLOW) at The ACM Web Conference (WWW 2026). Congratulations Joel Barmettler, Abraham Bernstein, and Luca Rossetto! The paper ConceptFormer: Towards Graph-Native Grounding of Large Language Models via Latent Concept Injection is available here