arXiv AI
3d ago

Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation

Retrieval‑augmented generation systems attach citations to claims, but these citations often misstate provenance. By replacing facts in retrieved passages with false variants and re‑querying the model, the study finds that 1.2% to 15.7% of cited claims actually come from the model’s parametric memory rather than the retrieved context, with larger models showing higher rates. Standard natural‑language‑inference verification fails to detect this, and even the best attribution methods only distinguish provenance by comparing the model’s output distribution with and without context, leaving the problem unresolved.

By Tanzeel Abbas, Tehseen Zia, David Windridge
Hugging Face Trending Papers
4d ago

Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation

Retrieval‑augmented generation systems attach citations to claims, but these citations often misstate provenance. By substituting facts in retrieved passages with false variants, the study shows that 1.2% to 15.7% of cited claims actually come from the model’s parametric memory rather than the retrieved context, with larger models exhibiting higher rates. Standard natural‑language‑inference verification fails to detect this, and even the best attribution method only distinguishes provenance by comparing the model’s output distribution with and without context, leaving accurate provenance detection unresolved.

arXiv Computation and Language
Oct 16

Beyond Retrieval Failure: Understanding Why LLMs Miss Evidence in Multi-Hop RAG

The paper investigates why large language models (LLMs) miss evidence in multi‑hop retrieval‑augmented generation (RAG). It introduces ExecutionRAG, a training‑free register that tracks open sub‑questions and their supporting evidence, allowing the authors to distinguish between true retrieval failures and search drift. Using this method on 1,000 MuSiQue questions, the study shows a significant reduction in missed evidence but also reveals that other error types, such as premature resolutions and abstentions, increase, shifting the bottleneck from search guidance to evidence verification.

By Jiaqi Wei, Wenxuan Huang, Chengdong Xu, Xiang Zhang, Wanli Ouyang