Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation
Organizations: COMSATS University Islamabad · Middlesex University London
Abstract
Retrieval-augmented generation systems attach citations to attribute claims to their sources. These citations frequently misstate provenance, and standard verification cannot detect this. We measure provenance by counterfactual value substitution: a fact in the retrieved passages is replaced with a false variant and the model re-queried. We find that across five sizes of one model family and a second comparator family, 1.2% to 15.7% of citable claims derive from parametric memory rather than the retrieved context. At the most conservative substitution distance these figures are 4.6% and 8.5% on the two models we tested at it, with the rate rising with model size within a family. We further find that natural language inference verification, the standard approach, gives results indistinguishable from chance across models, even when restricted to the cited passage. We determine that the cause is structural: a model answering from memory selects a citation that agrees with its output. Comparing ten attribution methods on identical claims, the contrast between the model's distribution with and without context best separates provenance. Acting on that signal does not improve provenance honesty at these base rates. The resulting detector is hence still not accurate enough to act on: it measures what the model knew rather than what it used, and separating those remains the open problem.
Figures & tables
| Model | B | Rate | 95% CI | par./dec. |
|---|---|---|---|---|
| Qwen2.5-0.5B | 0.5 | 1.2 | [0.6, 2.6] | 6 / 498 |
| Qwen2.5-1.5B | 1.5 | 6.5 | [4.5, 9.3] | 27 / 417 |
| Qwen2.5-3B | 3.0 | 8.3 | [6.1, 11.2] | 38 / 456 |
| Qwen2.5-7B | 7.6 | 8.2 | [6.0, 11.2] | 35 / 426 |
| Qwen2.5-14B | 14.8 | 13.3 | [10.4, 16.9] | 56 / 420 |
| Llama-3.1-8B | 8.0 | 15.7 | [12.6, 19.4] | 70 / 446 |
| Qwen2.5-7B | Qwen2.5-14B | |||
|---|---|---|---|---|
| Substitution | Rate | 95% CI | Rate | 95% CI |
| Near | 4.6 | [3.0, 6.8] | 8.5 | [6.3, 11.3] |
| Default | 8.2 | [6.0, 11.2] | 13.3 | [10.4, 16.9] |
| Far | 16.3 | [12.9, 20.3] | 17.6 | [14.2, 21.5] |
| Model | Cases | Share of records |
|---|---|---|
| Qwen2.5-0.5B | 0 | 0.0 |
| Qwen2.5-1.5B | 0 | 0.0 |
| Qwen2.5-3B | 0 | 0.0 |
| Qwen2.5-7B | 1 | 0.2 |
| Qwen2.5-14B | 140 | 25.7 |
| Llama-3.1-8B | 5 | 0.9 |
| Param. | Context. | |||
|---|---|---|---|---|
| Model | cited | % | cited | % |
| Qwen2.5-0.5B | 1 / 6 | 17 | 58 / 491 | 12 |
| Qwen2.5-1.5B | 12 / 27 | 44 | 206 / 388 | 53 |
| Qwen2.5-3B | 36 / 38 | 95 | 392 / 417 | 94 |
| Qwen2.5-7B | 35 / 35 | 100 | 373 / 390 | 96 |
| Qwen2.5-14B | 50 / 56 | 89 | 209 / 224 | 93 |
| Method | 0.5B | 1.5B | 3B | 7B | 14B | Llama | Mean |
|---|---|---|---|---|---|---|---|
| Localised (ours) | 0.877 | 0.833 | 0.802 | 0.779 | 0.743 | 0.697 | 0.788 |
| JSD, no context | 0.900 | 0.844 | 0.820 | 0.756 | 0.744 | 0.671 | 0.789 |
| Sentence-level | 0.698 | 0.725 | 0.669 | 0.596 | 0.683 | 0.643 | 0.669 |
| ConSens | 0.717 | 0.751 | 0.701 | 0.619 | 0.739 | 0.665 | 0.699 |
| Sufficiency | 0.807 | 0.684 | 0.643 | 0.596 | 0.670 | 0.644 | 0.674 |
| LOO max | 0.667 | 0.640 | 0.557 | 0.541 | 0.656 | 0.616 | 0.613 |
| Signal | 0.5B | 1.5B | 3B | 7B | 14B | L-8B |
|---|---|---|---|---|---|---|
| log p(v), no context | 0.678 | 0.811 | 0.811 | 0.731 | 0.733 | 0.673 |
| entropy, no context | 0.607 | 0.833 | 0.818 | 0.778 | 0.736 | 0.683 |
| change in log p | 0.855 | 0.816 | 0.805 | 0.721 | 0.748 | 0.695 |
| change in entropy | 0.877 | 0.833 | 0.802 | 0.776 | 0.743 | 0.697 |
| log p(v), with context | 0.191 | 0.475 | 0.576 | 0.502 | 0.477 | 0.457 |
| entropy, with context | 0.191 | 0.480 | 0.578 | 0.501 | 0.478 | 0.457 |
| Model | Honesty | Gain | Cost red. 3:1 |
|---|---|---|---|
| Qwen2.5-0.5B | 0.988 | 0.0 | |
| Qwen2.5-1.5B | 0.935 | 7.4 | |
| Qwen2.5-3B | 0.917 | 0.0 | |
| Qwen2.5-7B | 0.918 | 0.0 | |
| Qwen2.5-14B | 0.867 | 32.1 | |
| Llama-3.1-8B | 0.843 | 2.9 |
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Condition | Before | After |
|---|---|---|
| ASQA Qwen2.5-1.5B | 12.4% | 6.5% |
| ASQA Qwen2.5-7B | 18.7% | 8.2% |
| ASQA Llama-3.1-8B | 22.3% | 15.7% |
| ELI5 Qwen2.5-7B | 14.7% | 4.5% |
| QAMPARI Qwen2.5-7B | 16.2% | 2.4% |
| Verifier | 0.5B | 1.5B | 3B | 7B | 14B | L-8B |
|---|---|---|---|---|---|---|
| DeBERTa-v3-base | 0.553 | 0.452 | 0.505 | 0.526 | 0.506 | 0.539 |
| DeBERTa-v3-large | 0.650 | 0.415 | 0.476 | 0.537 | 0.548 | 0.487 |
| RoBERTa-base | 0.530 | 0.538 | 0.547 | 0.573 | 0.515 | 0.543 |
| DeBERTa MNLI-F | 0.576 | 0.579 | 0.539 | 0.519 | 0.566 | 0.540 |
| Method | Cost per claim |
|---|---|
| Localised detector (ours) | 2 forward |
| MIRAGE ( Qi et al., 2024 ) | 2 fwd, 1 bwd |
| ARC-JSD ( Li et al., 2026 ) | forward |
| ContextCite ( Cohen-Wang et al., 2024 ) | 32 to 256 forward |
| Counterfactual labelling | 1 generation |