Skip to content
EventPast event

Study: RAG citations often originate from parametric memory rather than retrieval

2 reports2 reporting sourcesLatest report

Start here

AI overview

A study published on November 12, 2026 (arXiv:2611.17161) shows that RAG systems often attach citations to retrieved documents that did not participate in generating the corresponding statements. Researchers at COMSATS University Islamabad and Middlesex University London measured causal provenance using a counterfactual value substitution method: replacing numbers in retrieved passages and titles with fake values; if the model outputs the fake value, it indicates reliance on retrieval, while restating the original value indicates the statement came from parametric memory. Evaluating Qwen2.5-Instruct (0.5B–14B) and Llama-3.1-8B-Instruct on the ASQA benchmark with GTR retrieval, 1.2%–15.7% of citable statements actually originated from parametric memory; for the Qwen series this rose with parameter scale from 1.2% at 0.5B to 13.3% at 14B, and Llama-3.1-8B reached 15.7%, with models still attaching citations to 89%–100% of parametric statements. Under the most conservative substitution distance, this rate was 4.6% and 8.5% on the two tested models, respectively. Natural language inference (NLI) verification, the standard method, performed no better than random, even when restricted to cited passages. The authors attribute the cause to structural factors and note that detectors built on this basis remain insufficiently accurate, measuring what the model knows rather than what it actually used.

Generated by AI from reports · Updated 4 days ago

Attention and developments

Attention, hourly

Now 40, 0% in 24 hours·Peak 45 (Nov 16 12:00)

Nov 17Nov 18Nov 19Nov 20Now 40

Progress timeline

Nov 16
  1. CASRAISelected
    Study finds RAG citations often reflect model parametric memory rather than retrieval sources

    A study published on November 12, 2026 shows that RAG systems often attach citations to retrieved documents that did not participate in generating the corresponding statements (arXiv:2611.17161). Researchers at COMSATS University Islamabad and Middlesex University London used a counterfactual value substitution method to measure causal provenance, replacing numbers in retrieved passages and titles with fake values; if the model outputs the fake value, it indicates reliance on retrieval, while restating the original value indicates the statement came from parametric memory. On the ASQA benchmark with GTR retrieval, they evaluated Qwen2.5-Instruct (0.5B–14B) and Llama-3.1-8B-Instruct; 1.2%–15.7% of citable statements actually originated from parametric memory, with the Qwen series rising with parameter scale from 1.2% at 0.5B to 13.3% at 14B, and Llama-3.1-8B reaching 15.7%, and the models still attached citations to 89%–100% of parametric statements.

Nov 13
  1. arXiv
    Cited but not used: parametric memory sources in retrieval-augmented generation

    The paper measures the source authenticity of citations in retrieval-augmented generation (RAG), using a counterfactual value substitution method to replace facts in retrieved passages with false variants and then re-asking the question; it finds that 1.2% to 15.7% of citable claims actually come from parametric memory rather than retrieved content, covering five sizes of one model family and one comparison family. At the most conservative substitution distance, this rate was 4.6% and 8.5% on the two tested models, and within the same family the rate rises with model scale. Natural language inference (NLI) verification results, as a standard method, were no different from random, even when restricted to cited passages. The authors attribute this to structural factors: a model answering from memory will select citations consistent with its output; after comparing ten attribution methods on the same claims, the difference in the model's output distribution with and without context best distinguished the source, but the authors note that a detector built on this is still not accurate enough, measuring what the model knows rather than what it actually used, and distinguishing the two remains an open problem.

Related discussion

Total 3 items

Attention trend

New sources per day

4 in 6 days·most 3 (Nov 15)·today 0

  • Nov 15: 3 new sources (0 outlets, 3 accounts)
  • Nov 16: 1 new sources (1 outlets, 0 accounts)
  • Nov 17: 0 new sources (0 outlets, 0 accounts)
  • Nov 18: 0 new sources (0 outlets, 0 accounts)
  • Nov 19: 0 new sources (0 outlets, 0 accounts)
  • Nov 20: 0 new sources (0 outlets, 0 accounts)

Hourly attention

Current attention 40·Peak in the comparable period 45 (Nov 16 12:00)·Change over the comparable past 24 hours 0%

0204060Nov 1611:00Nov 1716:00Nov 1822:00Nov 2003:00