Open access
Oct 2026
The Bottleneck Is Retrieval, Not Reading: Diagnosing Evidence-Based Generation with RAG for Long Context Question Answering
Evidence-based generation (EBG), requiring models to return an answer with line-level citations, and a joint Ans+Evid metric that scores correctness and grounding simultaneously, is introduced, isolating two RAG limitations that long-context models avoid.
Debopriyo Banerjee, Owais Aijaz, Ahmer Tabassum et al.
· Conference on Empirical Meth... · 0 citations