🔍 Information Retrieval & RAG¶

🧪 ICML2026 · 2 paper notes

Hierarchical Abstract Tree for Cross-Document Retrieval-Augmented Generation: Ψ-RAG replaces RAPTOR's k-means with a "merge–collapse" hierarchical clustering to construct a cross-document abstract tree, paired with a retrieval-answering Agent capable of multi-turn rewriting and a hybrid sparse BM25 index. This enables Tree-RAG, for the first time, to match or even surpass Graph-RAG on corpus-level, cross-document multi-hop QA, achieving an average F1 25.9% higher than RAPTOR and 7.4% higher than HippoRAG 2.
Very Efficient Listwise Multimodal Reranking for Long Documents: ZipRerank simultaneously eliminates the two main bottlenecks of VLM-based listwise reranking—"overly long visual token sequences" and "autoregressive decoding with per-token ranking output"—by employing query-aware token pruning and single-logit sorting. On MMDocIR, it reduces LLM inference latency by an order of magnitude while matching or surpassing the current SOTA MM-R5.