{"id":"fd1f18d0-4b5b-4335-9a95-7238ee4185b9","arxiv_id":"2412.18004","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Citation correctness is not enough: a RAG model may cite documents that merely contain the answer's wording while actually answering from memory, a behavior the paper calls post-rationalization.","lead":"This paper separates two ways a citation can be good: it can correctly support the answer, or it can reflect how the model actually used the document when writing the answer. The authors show that a RAG model often attaches citations to documents it appears not to have relied on, by inserting the answer's key words into unrelated documents and watching the citations follow.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's headline 'up to 57%' depends on an unverified invariance assumption: citing a document after the answer statement is inserted into it may be genuine context use, not post-rationalization; a contradiction control is needed.","rationale":"The paper makes a valuable conceptual distinction between citation correctness and citation faithfulness, and its proposed necessary-condition test is a reasonable first probe. My main concern is not a logical contradiction internal to the paper but an unverified empirical assumption. Because the paper itself flags the internal-process invariance assumption in Section 6, the test's results cannot yet support the specific 'post-rationalization' mechanism or the precise 'up to 57%' prevalence. A contradiction control is the cheapest decisive check: it directly tests whether citation follows semantic support or surface overlap. If the model cites a document that contradicts its answer, the inference of post-rationalization is solid; if not, the high adversarial-citation rate may reflect the model legitimately using the newly inserted supporting content. I also note the prevalence computation should be based on the full set of cases or clearly state the conditioning on recovered statements and the 'relevant but not cited' subset. None of this changes the overall verdict: the conceptual claim that citation correctness is insufficient is well argued and worth conditional acceptance; the magnitude claim needs strengthening before it can be reported as a stable 'up to 57% of citations' statistic.","tokens_in":14130,"tokens_out":8266,"duration_ms":83026,"concrete_test":"Add a contradiction-control arm to the existing probe, using the 'cited for other reason' subset: instead of appending the original statement, append a lexically overlapping but false statement (e.g., for 'What is the capital of Germany?', append 'The capital of Germany is Bonn' to the otherwise relevant document). If Command-R+ continues to output Berlin and cites the false document, post-rationalization is confirmed and the original 55–57% interpretation survives. If the model instead avoids citing the false document or changes its answer, the original adversarial test conflates genuine context use with post-rationalization and the prevalence claim must be re-estimated with the full denominator (290/829 in the cited-for-other-reason condition).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that up to 57% of citations are post-rationalized—is inferred from the observation that Command-R+ cites an 'adversarial' document after the previously generated statement (e.g., 'Emperor penguin') is appended to it. The inferential step is the weak link: because the input context changes, the model's decision process can legitimately change with it. A RAG-optimized model that is asked to ground its answers may cite a document that now contains the exact claim, not because it is post-rationalizing from parametric memory, but because the document has become a genuine and relevant source. The paper explicitly concedes this in Section 6: the test 'implicitly assumes that the model internals ... do not change through the insertion of additional irrelevant documents' and that internal-state investigation would be needed. Additionally, the quantitative headline is fragile: the 57% figure comes from the 'relevant but not cited' condition (273/476), not from the set of originally issued citations; in the 'cited for other reason' condition, using the full 829-case denominator gives 35% (290/829), not 55–57%. The probe tests only necessary condition (2) from Section 4.1 and cannot establish the causal-dependence condition in the definition of citation faithfulness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that citation correctness in RAG is insufficient for trustworthy attribution, and introduces the notion of citation faithfulness, defined as a causal dependence of the generated statement on the cited document (Section 4.1). The authors propose a set of desiderata and a necessary-condition test, then report an experiment on Cohere's Command-R+ over Natural Questions in which short statements from previously generated answers are appended to other documents. The model cites these adversarial documents in a large fraction of cases where the original statement is recovered, which the authors interpret as post-rationalization. They conclude that current attributed answers often lack citation faithfulness, with 'up to 57% of the citations' affected.","tokens_in":14405,"tokens_out":4690,"duration_ms":44109,"significance":"The conceptual contribution is useful and timely: separating whether a citation is supported by a document from whether the model actually used that document is an important distinction for trustworthy RAG. The paper is also transparent about its limitations, releases code, and proposes a concrete perturbation-based probe that others can build on. However, the headline quantitative claim is not yet solidly established: it rests on an invariance assumption the authors themselves flag, uses a conditional denominator without a reported baseline, tests only a necessary condition, and covers a single model with no statistical uncertainty. If the empirical evidence is strengthened, the paper could be a valuable agenda-setting contribution for evaluation of attribution faithfulness.","major_comments":[{"comment":"The test treats 'the model cites a document that contains the previously generated statement' as evidence of post-rationalization, but this inference requires the invariance assumption the authors themselves flag in Section 6, namely that inserting the statement does not change where the model looks or how it generates the answer. Because the input context is changed, citing the adversarial document could be genuine context use rather than post-rationalization. A contradiction control (e.g., appending a conflicting statement and observing whether the answer changes) is needed to distinguish these possibilities; without it, the experiment cannot establish the causal dependence required by the definition in Section 4.1.","section":"Section 5, 'Experiments'; Section 6"},{"comment":"The headline 'up to 57 percent of the citations' is computed conditionally on the cases in which the original statement was recovered in the new answer (273/476 = 57%, 290/525 = 55%), not from the full set of adversarial documents (290/829 = 35% in the 'cited for other reason' condition, or 273/702 = 39% in the 'relevant but not cited' condition). The abstract and introduction therefore overstate the finding unless the conditional denominator and the relevant baseline citation rates are reported prominently.","section":"Section 5, 'Results'; Figure 3"},{"comment":"The experimental manipulation does not instantiate necessary condition (2), because a document containing the exact answer statement (e.g., 'Emperor penguin') is not an 'irrelevant document' in the sense of that condition; it is highly relevant by token match. There is also no control condition in which an irrelevant document without the inserted statement is presented, so the paper does not measure whether the insertion itself changes citation behavior relative to a baseline. As a result, the experiment tests only a heuristic for token-matching behavior, and it cannot by itself establish that current attributed answers 'often lack citation faithfulness'.","section":"Section 4.1, condition (2); Section 5, 'Setup'"},{"comment":"All quantitative claims come from a single model (Command-R+), a single dataset (Natural Questions), and a single run, with no confidence intervals or repeated sampling; the paper's own Section 7 acknowledges the limited scope. For the empirical claim to be load-bearing, the authors should add at least one additional model family, report variance across runs, and state whether decoding was greedy or sampled.","section":"Section 5, 'Setup'; Section 7"}],"minor_comments":[{"comment":"There are several typos that should be corrected: 'correction' for 'correctness' in Section 1, 'casual' for 'causal' in Section 4, 'Comprehnsiveness' in Table 1, and 'Emperror' in Example 1.","section":"Throughout"},{"comment":"Figure 3 would be easier to interpret if the bars were labeled with the exact percentages in addition to counts, and if the denominator for each percentage were stated in the caption.","section":"Figure 3"},{"comment":"The statement-recovery rate (63–70%) is important for interpreting all downstream percentages; the paper should state the matching criterion (exact token match, substring, or paraphrase) used to decide that an old statement was recovered.","section":"Section 5, 'Results'"}],"recommendation":"major_revision","confidential_remarks":"The 'up to 57%' claim is likely to be quoted out of context; I suggest the editor require the revised version to present raw counts, baselines, and the conditional nature of the number. The conceptual distinction is solid and could anchor a useful line of work, but the empirical evidence is currently exploratory rather than conclusive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing worth knowing: this paper separates citation correctness from citation faithfulness and names post-rationalization. That distinction is real and overdue in RAG evaluation. The probe is simple—append a previously generated statement to other documents and see if the model cites them—and the paper is honest about its limits. Read it for the framing, not for the 57%.\n\nWhat it does well: the definition of citation faithfulness in Section 4.1 is careful, adapting the causality requirement from interpretability work. The two necessary conditions (alter the source, add irrelevant documents) are concrete and testable. The paper also gives desiderata beyond correctness—appropriateness and comprehensiveness—and acknowledges in Section 6 that the probe assumes model internals do not change when the input changes. That admission is more candid than most. Code is available.\n\nWhere it gets soft: the headline \"up to 57% of citations\" is not what the experiment actually counts. The 57% is 273/476, meaning cases where the original statement was recoverable in the \"relevant but not cited\" condition and the adversarial document was cited. The \"cited for other reason\" condition gives 290/525 = 55%, but using the full 829-case denominator drops it to 35%. The abstract's phrasing overstates the prevalence. Also, the inference from a citation switch to post-rationalization is a leap: inserting the statement into a document can make that document a genuinely relevant source, so citing it is not necessarily evidence of shallow token-matching. The paper flags this, but it means the empirical claim is conditional on an untested invariance. The random-document condition (12%) is a baseline, but there is no pre-insertion citation rate for the same document, no error bars, and only one model and one dataset.\n\nNone of this kills the conceptual contribution. The paper is a useful corrective to the field's near-exclusive focus on citation correctness, and the probe is a reasonable first step. I would bring it to a reading group and cite it for the definition of post-rationalization. But the abstract and conclusion should be rewritten to say \"in our test, under the invariance assumption, up to 57% of recovered statements in one condition switched citation\"—not \"57% of citations lack faithfulness.\"\n\nFor peer review: yes, send it out. The idea is important and the paper is honest. A good referee will push for a more restrained abstract, additional models, and a control that checks whether the model's internal processing actually changes when the document is inserted. The central claim—that correctness is not faithfulness—holds up; the magnitude is the problem.","headline":"A genuinely useful conceptual distinction between citation correctness and citation faithfulness, but the 57% headline is a loose reading of the experiment.","tokens_in":14878,"tokens_out":2470,"would_cite":true,"duration_ms":25219,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A citation can be factually correct yet unfaithful: RAG models often cite documents they did not actually use to answer.","keywords":["citation faithfulness","post-rationalization","retrieval-augmented generation","RAG attribution","faithfulness evaluation","large language models","trustworthy AI"],"falsifier":"Replace the planted answer snippet in the adversarial document with a factually contradicting sentence (e.g., 'The capital of Germany is Bonn') while keeping the rest of the document unchanged. If the model continues to answer 'Berlin' and continues to cite that document, the citation is not causally dependent on the cited content, confirming post-rationalization. If the answer or citation changes, the causal-use assumption of the test fails.","tokens_in":13926,"feed_emoji":"🧾","tokens_out":6813,"duration_ms":57251,"temperature":0.7,"pith_summary":"Retrieval-augmented generation systems answer questions with citations to source documents, and standard evaluation measures whether each cited document supports the statement made. This paper argues that support is not enough: a faithful citation must also have causally influenced the model's answer, not merely match its wording. The authors define citation faithfulness in terms of causal dependence and design an adversarial test that inserts the answer's own wording into unrelated documents. In a study of a 104B-parameter RAG model, the model went on to cite those planted documents for the same statement in up to 57% of cases, suggesting post-rationalization: the model answers from memory and then finds a plausible-looking source. If this is right, correctness metrics overstate trustworthiness, and users cannot verify an answer by checking that its citations exist and support the claim.","feed_headline":"Up to 57% of RAG citations are post-rationalized","feed_subtitle":"A planted-snippet experiment shows retrieval-augmented models often cite by token matching, not genuine reliance.","key_machinery":"The load-bearing mechanism is the post-rationalization test, an adversarial-document experiment. The test starts from an attributed answer, extracts short statements (e.g., 'Emperor penguin'), appends those statements to documents that have no contextual support for them, and re-runs generation with the doctored documents in the context. If the model cites a doctored document for the same statement, the citation cannot reflect a causal reliance on the document's actual content, because the document only contains the statement's wording. This makes post-rationalization observable without requiring access to model internals. The formal definition of citation faithfulness—document in context, support, and causal impact—frames the test as evaluating one necessary condition of faithfulness, following the causality-based definition of faithfulness from the explainability literature.","core_discovery":"The central discovery is that citation correctness and citation faithfulness are distinct properties, and current attributed answers frequently satisfy the first while failing the second. The paper defines citation faithfulness as a three-part condition: the cited document is in the retrieved context, the document supports the statement, and the statement is causally impacted by the document. It then operationalizes the causal condition through a post-rationalization test. Short statements (2–4 tokens) from a model's own answers are appended to other documents—random documents, relevant-but-uncited documents, and documents cited for other reasons—and those documents are substituted into the context. When the model cites the adversarial document for the original statement, the citation is taken to be a post-rationalization rather than genuine use. Command-R+, a 104B-parameter RAG-optimized model, cited the planted documents in 12% of random cases, 57% of relevant-but-uncited cases, and 55% of documents cited for other reasons. The authors conclude that correct-but-unfaithful citations are common, so correctness alone is insufficient for trustworthy attribution.","pith_inferences":["A natural extension is to test other RAG models and datasets; if the 57% figure is representative, the problem is systematic rather than specific to one model.","The insertion test could be strengthened into an ablation test: replace the planted statement with a contradicted one and check whether the answer and citation change; if neither changes, the citation's causal role is refuted even more directly.","The paper's framing connects RAG attribution to the broader chain-of-thought faithfulness literature, implying that the same post-rationalization failure likely affects other self-explanations such as step-by-step reasoning.","For users, the practical consequence is that verifying a citation's content is not the same as verifying the answer; interfaces should communicate the difference, or display process-level evidence when available."],"forward_implications":["Evaluating attribution solely by citation correctness will overstate trustworthiness; faithfulness must be measured as a separate dimension.","Post-hoc attribution methods that retrieve evidence after the answer is generated are unfaithful by design and should not be used where users rely on citations for verification.","Models trained for grounded generation also exhibit post-rationalization, so even direct-prompting RAG pipelines need faithfulness checks.","The adversarial-document test provides a cheap, model-agnostic way to detect one form of unfaithful behavior without probing hidden states.","High-stakes applications such as legal and medical question answering should treat a correct citation as a necessary but insufficient condition for trusting an answer."],"supporting_citations":[{"why":"Prior work on attributed question answering that defines citation correctness via NLI support; this is the baseline the paper argues is insufficient.","marker":"[6]"},{"why":"Provides the definition of a faithful explanation as one that reflects the model's decision process, which the paper adapts to citations.","marker":"[22]"},{"why":"Establishes that faithfulness is a causal relationship, the basis for the paper's causal-impact condition.","marker":"[32]"},{"why":"Supplies the evaluation strategy for unfaithful chain-of-thought explanations that the paper adapts to the adversarial-document test.","marker":"[50]"},{"why":"Supplies the NaturalQuestions benchmark of 1,444 questions used to generate attributed answers in the experiment.","marker":"[25]"},{"why":"Supplies the passage-chunking strategy (100-token chunks with titles prepended) used to build the retrieval index.","marker":"[12]"},{"why":"Supplies the ColBERT v2 reranker used to select the top 5 documents passed to the model.","marker":"[44]"},{"why":"Documents the phenomenon of post-hoc reasoning in chain-of-thought explanations, a precursor to the paper's post-rationalization concept.","marker":"[26]"}],"fun_headline_variants":["Correct but unfaithful: RAG citations often post-rationalized","Up to 57% of RAG citations lack citation faithfulness","RAG models cite by token matching, not genuine reliance","Post-rationalization undermines trust in RAG attributions","Citation correctness doesn't guarantee faithfulness in RAG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The test assumes that when a document contains the exact wording of the answer but no supporting context, a citation to it must be post-rationalization rather than genuine causal use of the newly inserted content, and that inserting the document does not alter the model's internal answer-generation process; the paper notes that verifying the latter would require investigating internal states.","fun_headline_variants_meta":{"raw":{"variants":["Correct but unfaithful: RAG citations often post-rationalized","Up to 57% of RAG citations lack citation faithfulness","RAG models cite by token matching, not genuine reliance","Post-rationalization undermines trust in RAG attributions","Citation correctness doesn't guarantee faithfulness in RAG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000331,"raw_usage":{"total_tokens":1836,"prompt_tokens":929,"completion_tokens":907,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":823}},"tokens_in":545,"tokens_out":907,"duration_ms":8433,"temperature":1.0,"reasoning_tokens":823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:07:33.533464+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the planted answer snippet in the adversarial document with a factually contradicting sentence (e.g., 'The capital of Germany is Bonn') while keeping the rest of the document unchanged. If the model continues to answer 'Berlin' and continues to cite that document, the citation is not causally dependent on the cited content, confirming post-rationalization. If the answer or citation changes, the causal-use assumption of the test fails.","supporting_citations":[{"cited_title":"4198– 4205 (2020)","cited_arxiv_id":null,"evidence_quote":"Provides the definition of a faithful explanation as one that reflects the model's decision process, which the paper adapts to citations."},{"cited_title":"Computational Linguistics pp","cited_arxiv_id":null,"evidence_quote":"Establishes that faithfulness is a causal relationship, the basis for the paper's causal-impact condition."},{"cited_title":"Advances in Neural Information Processing Syst ems 36 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the evaluation strategy for unfaithful chain-of-thought explanations that the paper adapts to the adversarial-document test."},{"cited_title":"Tr ansactions of the Association for Computational Linguistics 7, 453–466 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the NaturalQuestions benchmark of 1,444 questions used to generate attributed answers in the experiment."},{"cited_title":"In: Proceedings of the 47th Inter national ACM SIGIR Conference on Research and Development in Informatio n Retrieval, pp","cited_arxiv_id":null,"evidence_quote":"Supplies the passage-chunking strategy (100-token chunks with titles prepended) used to build the retrieval index."},{"cited_title":", Zaharia, M.: Col- BERTv2: Effective and efficient retrieval via lightweight lat e interaction","cited_arxiv_id":null,"evidence_quote":"Supplies the ColBERT v2 reranker used to select the top 5 documents passed to the model."}],"review_version":1}