Attributable Post-Rationalization in RAG Citations: A Controlled Reproduction and an RLVR Comparison
Mehedi Khan, Md. Shariful Islam Bhuyan
cs.CL · cs.IR
Quick verdict
Claims answer-only RLVR leaves RAG citation faithfulness unchanged; controlled reproduction is persuasive but lacks human validation and an SFT control.
Summary
This reproduction study asks whether reinforcement learning from verifiable rewards (RLVR) teaches search agents to cite honestly. The authors adapt the planting probe of Wallat et al. (2025): a four-word span of the model's own answer is appended to an uncited retrieved passage, and the model is scored on whether it then cites the doctored passage. Their methodological contribution is a no-planting control, derived analytically to be 0% for the main REL-UNCITED variant and 100% for CITED-OTHER, and measured (~0.2% over 7,625 records) for RANDOM. Using Qwen2.5-7B-Instruct as baseline against three RLVR agents (Search-R1, ReSearch, R-Search) on four QA datasets (NQ, QAMPARI, HAGRID, RAGTruth) with fixed BM25 top-5 retrieval, one citation-bearing prompt, and a deterministic scorer (36,690 records), they find baseline post-rationalization of 2.9-14.8%, i.e., roughly one in seven cited sentences on Wikipedia-based datasets. No RLVR agent improves on its base: all difference-in-differences intervals cover zero except R-Search on RAGTruth (+3.0 pp, Holm p = 0.002). Key limitations: single-turn fixed-context evaluation diverges from the agents' multi-turn search training format (abstentions 15-52%), no SFT-only control isolating RL, no manual validation of flagged cases, and fixed span placement at passage ends.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under the planting probe with a no-planting control, the instruction-tuned baseline Qwen2.5-7B-Instruct post-rationalizes at attributable rates of 2.9% (RAGTruth) to 14.8% (HAGRID), i.e., roughly one cited sentence in seven on the three Wikipedia-based datasets.
- Three RLVR agents (Search-R1, ReSearch, R-Search) trained from the same base show no improvement in citation faithfulness under fixed-context inference: all DiD confidence intervals cover zero except R-Search on RAGTruth (+3.0 pp, Holm-adjusted p = 0.002), which is significantly worse.
- Raw planted rates without a control can be misleading or inverted: for the CITED-OTHER variant the control is 100% by construction and attributable rates are negative (-10.4 to -21.1 pp) despite raw rates of 79-90%.
- Answer-only verifiable rewards never inspect the sentence-to-passage link, so citation faithfulness must be directly rewarded and measured; the controlled probe is proposed as a cheap reward term and cross-model evaluation metric.
Strengths
The no-planting control is a genuinely valuable methodological repair
Section 3.3 derives the control analytically for two variants — 0% for REL-UNCITED (removing the appended span restores the clean context byte-for-byte, verified on all 10,720 records) and 100% for CITED-OTHER — and measures ~0.2% over 7,625 records for RANDOM. Table 3 shows raw CITED-OTHER rates of 78.9-89.6% collapse to attributable rates of -10.4 to -21.1 pp, a concrete demonstration that uncontrolled planted rates can be pure sampling artifacts.
Clean controlled design isolating training as the only difference
All four models share the Qwen2.5-7B-Instruct base (Table 2); retrieval corpus, prompt (byte-identical, Appendix A), scorer, and greedy decoding are held fixed, with agent retrieval deliberately disabled — a defensible choice explicitly justified in Section 4 against the confound of differing search policies.
Statistical hygiene well above field norm
Transparent, boundary-aware reporting of the negative result and the R-Search exception
Radical accessibility and reproducibility scaffolding
Weaknesses
Evaluation format mismatch threatens construct validity of the RLVR comparison
No SFT-only control arm, so the 'RLVR' attribution is associational
The scorer is never validated against human judgment, and a competing interpretation is acknowledged but untested
Positional confound in the probe
Modest power and a single-base, single-prompt scope
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper (arXiv 2609.23053) reproduces and repairs the post-rationalization probe of Wallat et al. (2025) and applies it to a question the original study could not answer: does RLVR training teach search agents to cite honestly? The setting is deliberately narrow — one base model (Qwen2.5-7B-Instruct), three agents trained from it (Search-R1/PPO, ReSearch/GRPO, R-Search/GRPO), four QA datasets (NQ 499, QAMPARI/HAGRID/RAGTruth 500 each), fixed BM25 top-5 retrieval, one prompt, greedy decoding. Three claims anchor the paper: (1) baseline post-rationalization is common (2.9–14.8% attributable by dataset; roughly one in seven on Wikipedia-based data, Table 4); (2) RLVR leaves it untouched, with the sole significant exception being R-Search *worse* on RAGTruth (+3.0 pp, Holm p = 0.002, Table 5); (3) raw planted rates without a control can be artifacts, so attribution should be measured as planted-minus-control (Eq. 1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is post-rationalization in RAG citations?
It is when a retrieval-augmented model writes an answer first (often from its own knowledge) and then attaches a citation to a retrieved passage that looks plausible but was not actually used. The answer can be correct and the citation well-formed while the provenance is false, so a reader who opens the cited passage and sees similar wording may be reassured incorrectly.
Does RLVR training make search agents cite more faithfully?
In this paper, no. Three RLVR-trained search agents (Search-R1, ReSearch, R-Search), all built from Qwen2.5-7B-Instruct, showed citation faithfulness statistically indistinguishable from their base model on four QA datasets under fixed-context evaluation. The only significant difference was negative: R-Search post-rationalized about 3 percentage points more than its base on RAGTruth (Holm-adjusted p = 0.002).
How common are unfaithful citations in RAG according to this study?
The baseline Qwen2.5-7B-Instruct followed an experimentally planted (unfaithful) citation in 2.9% to 14.8% of claims depending on the dataset. On the three Wikipedia-based datasets (NQ, QAMPARI, HAGRID) this means roughly one in seven cited sentences pointed at a passage whose only connection to the claim was four appended words. On the web-text RAGTruth dataset the rate was much lower (2.9%), suggesting the behavior is domain-dependent.
What is the no-planting control this paper introduces?
It is a baseline run in which the context is edited (or not) without the actual plant, so the raw planted rate can be corrected by subtraction. For the main variant (appending four words to an uncited passage), the control is provably 0% because removing the appended words restores the original context exactly. For another variant the control is 100%, showing its raw planted rate of 79-90% is a pure sampling artifact — a warning against reporting uncontrolled planted rates.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (e.g., Figure 1's planted-passage example) are available only as caption text; visual layout could not be verified.
- Appendices were partially extracted (e.g., QAMPARI claims counts 329/352/362 in Table 7 vs. 319/332/352 in Table 4), suggesting minor extraction inconsistencies; conclusions were cross-checked against the main tables.
- The 10,720-record verification and per-cell control rates (Appendix D) are summarized in text rather than shown as full tables.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.