RAG retrieves the right docs but still answers wrong? Saarland University team diagnoses why | ACL 2026

RAG搜对了却答错？德国萨尔大学找到了真相丨ACL'26

A Saarland University-led team introduced Disco-RAG, adding a 3-step “reading” layer between retrieval and generation, and says the paper was accepted as an ACL 2026 main-conference long paper. The post says it uses RST-based argument trees, cross-passage relation graphs, and outline generation with zero training; it reports gains on Loong, ASQA, and SciNews, but does not fully disclose the exact scores. The key claim is that many RAG failures come from reading and discourse understanding, not retrieval recall.

Why it matters: This is a solid research release with HKR-H, HKR-K, and HKR-R: a strong practical hook, a concrete mechanism, and a pain point RAG builders know well. I keep it at 80, not higher, because the post does not fully disclose benchmark numbers and external replication is still missing