BrowseNet: Graph-Based Associative Memory for Contextual Information Retrieval
A research framework, published at ICLR'26, that decomposes multi-hop questions into a query subgraph and traverses a graph of document chunks to retrieve reasoning-aligned context for RAG at a fraction of the LLM cost of prior methods.
BrowseNet is a graph-based associative-memory framework for Retrieval-Augmented Generation that decomposes a multi-hop query into a directed acyclic graph, called a query-subgraph, and traverses a Graph-of-Chunks to retrieve structured context for an LLM to reason over. The accompanying paper reports state-of-the-art retrieval performance on the 2WikiMQA, HotpotQA, and MuSiQue benchmarks while reducing LLM inference cost by roughly 33 times compared to the prior state of the art, HippoRAG-2, without a significant latency trade-off.
The released code includes environment setup through Conda, a dependency on a pre-trained ColBERTv2 checkpoint for semantic keyword matching, and benchmark datasets under a datasets directory. Evaluating BrowseNet on a new dataset requires providing a corpus.json file of passages and a questions.json file of multi-hop queries in a documented format, with optional gold answer IDs and edge lists for retrieval evaluation.