{"id":"32fed295-1b84-4771-be0e-51aaa596ea2e","arxiv_id":"2507.22915","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that restates known hallucination definitions and standard PAC-Bayes bounds, then proposes a generic detection-mitigation workflow, without new results or experiments.","lead":"This paper surveys existing ideas about why large language models make things up and how to catch or reduce those errors. It adds formal-looking definitions and a standard learning-theory bound, but the math is a restatement of prior work, not a new derivation.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The PAC-Bayes bound in Eq. 1 is formally standard, but it is applied to a 0-1 loss that is not well-defined: hallucination is not equality to a unique oracle f*(x), so the empirical hallucination rate is convention-dependent and the bound does not actually constrain hallucination risk.","rationale":"The reader's weakest-assumption analysis is on target. The paper's formal contribution is a conditional PAC-Bayes inequality, not a demonstrated bound on any observable hallucination rate. The survey sections are reasonable summaries of existing detection and mitigation work, and the proposed workflow is a diagram rather than an evaluated system; the reader correctly treats the absence of empirical validation as disqualifying for an 'accept.' I considered whether the more serious flaw is that the bound is just a standard theorem with no new derivation. That is a novelty objection, not a correctness objection. The load-bearing correctness issue is the definition of H. The paper explicitly assumes f* and then applies the 0-1-loss PAC-Bayes theorem, but the assumption does not survive contact with the object it is meant to model. A set-valued oracle could rescue the framework, but the paper does not provide one. If the authors supply such an oracle and a re-derived bound, the rejection could be revisited; as written, the central claim is not about a well-defined quantity.","tokens_in":15730,"tokens_out":9924,"duration_ms":106970,"concrete_test":"Analytical counterexample: take a two-sentence source article and write two distinct summaries, y1 and y2, both fully faithful and factually equivalent. Instantiating Definition II.B with f*(x)=y1 makes H(M,x)=1 for y2, while f*(x)=y2 makes H(M,x)=1 for y1. Thus R_hall(M) depends on an arbitrary choice of oracle, so Eq. 1 does not define a unique quantity. To settle the concern, the authors would need to replace f* with a set-valued oracle or an explicit equivalence relation and re-derive Eq. 1 from a loss function that is well-defined under that specification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Eq. 1 in Section III.B, a PAC-Bayes bound on E_{M~Q}[R_hall(M)]. The bound is a standard theorem for bounded 0-1 losses, so the mathematical derivation is not the weak point. The weak point is the definition of the loss. Section II.B assumes a ground-truth oracle f*(x) that outputs one 'fully truthful and contextually appropriate' text, and defines H(M,x)=1 iff M(x) is not completely faithful to that single output. For language generation, however, there are many equally faithful outputs for the same input: multiple correct paraphrases, summaries, or answers. Choosing different f* changes H without changing factual quality. Section II.A even concedes that whether content counts as a hallucination 'can depend on the task context and the expected scope of the output,' but the formalization does not encode this; it silently replaces a semantic judgment with a unique oracle. Consequently, the empirical estimate \\hat R_hall(M) in Section III.B is not a well-defined sample statistic: there is no objective label to average, and any human or automatic labeling is a convention, not a measurement of the paper's quantity. The PAC-Bayes inequality is therefore not false, but it is vacuous for the object it claims to bound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes formal definitions of intrinsic and extrinsic hallucination, introduces a hallucination risk R_hall(M) defined through a ground-truth oracle f*(x), and states PAC-Bayes and Rademacher-complexity bounds on that risk (Sections II and III). It then surveys detection methods (uncertainty estimation, calibration, attention alignment), mitigation strategies (RAG, fine-tuning, logit calibration, fact-verification), proposes a combined detection/mitigation workflow (Section VI), and recommends evaluation datasets and metrics (Section VII). The manuscript reports no experiments and no new mathematical proof; its main formal contribution is Eq. (1), presented as a direct application of McAllester's PAC-Bayes inequality.","tokens_in":16037,"tokens_out":7201,"duration_ms":86333,"significance":"The paper covers a broad and important topic and its survey parts cite relevant literature accurately (e.g., semantic entropy [6], FactCC [24], TruthfulQA [28], chain-of-verification [34]). The proposed workflow is coherent and could serve as a practical checklist for system builders. However, the theoretical contribution is not novel: Eq. (1) is a standard PAC-Bayes inequality, and the central definition of hallucination risk via a unique oracle is not valid for the open-ended generation tasks that the paper explicitly targets. Because the formal object that the bound is supposed to constrain is not well-defined, the claimed theoretical foundation does not currently support the paper's conclusions.","major_comments":[{"comment":"The definition of H(M,x) relies on a single ground-truth oracle f*(x) that produces one 'fully truthful and contextually appropriate' output for each input, but Section II.A states that whether content counts as a hallucination 'can depend on the task context and the expected scope of the output.' For summarization, knowledge-grounded dialogue, and open-ended QA, there are typically many fully faithful outputs (paraphrases, equally valid summaries, different correct answers), so equality to a unique f*(x) cannot define a binary hallucination indicator. The empirical rate \\v{R}_hall(M) in Section III.B is therefore convention-dependent: different choices of f* change the value on the right-hand side of Eq. (1) without changing the factual quality of M. The PAC-Bayes inequality may be valid for a fixed labeling oracle, but it does not bound an objective hallucination risk for the tasks the paper targets. This is load-bearing because Eq. (1) is the paper's central theoretical result.","section":"Section II.B and Section III.B, Eq. (1)"},{"comment":"The paper says it derives a PAC-Bayes bound, but Eq. (1) is quoted from McAllester [38] as a 'direct application' with no derivation and no precise statement of the underlying supervised protocol. In particular, no labeled training sample of pairs (x_i, y_i) is defined: \\v{R}_hall(M) is an average of H(M,x_i) over inputs x_i with labels supplied by an external oracle, whereas the PAC-Bayes theorem concerns empirical risk on a drawn sample with a fixed loss l(h,z) for z=(x,y). As written, the bound does not formally follow from the cited theorem, and the paper should either prove the instantiation or state the exact learning protocol (including what data is drawn, what the labels are, and over what randomness the probability 1-δ is taken).","section":"Section III.B, Eq. (1)"},{"comment":"The impossibility claim that hallucinations are fundamentally unavoidable is presented only as an informal paraphrase of Xu et al. [11] ('a recent theoretical result suggests...'). No theorem statement, assumptions, or proof is given, so the paper's 'theoretical takeaway' is not self-contained. If this claim is part of the paper's contribution, it should be stated as a precise theorem (e.g., in terms of computability, the distribution over inputs, and the hypothesis class) and proved or explicitly attributed as an external result outside the scope of the paper's own derivation.","section":"Section III.C"}],"minor_comments":[{"comment":"The phrase 'similar to how Fact was a BERT-based classifier' appears to refer to FactCC [24]; the typo should be corrected and the reference made explicit.","section":"Section IV.B"},{"comment":"Figure 1 is rendered as ASCII art in the manuscript; a proper figure with clearly labeled components (input, LLM, detection module, mitigation module, final answer) should be provided, since the text references it as 'Figure 1.'","section":"Section VI and Figure 1"},{"comment":"The bound R_hall(M) ≤ \\v{R}_hall(M) + O(√(C/n)) is stated with an unspecified complexity measure and unspecified constants; for Rademacher complexity the standard form is R_n + √(ln(1/δ)/(2n)) (or similar), so the exact theorem and assumptions should be cited to make the statement checkable.","section":"Section III.A"},{"comment":"Reference [30] is cited alongside [17] and [18] for retrieval-augmented generation, but [30] is a knowledge-based dialogue metrics paper; the citation appears misplaced and should be moved or replaced.","section":"Introduction, RAG description"},{"comment":"The experimental recommendations (e.g., 'use sufficiently large sample sizes and statistical tests') are sensible, but the paper itself presents no empirical evaluation; the text should explicitly state that these are recommendations for future work, not results of this paper.","section":"Section VII.C"}],"recommendation":"reject","confidential_remarks":"The paper is essentially a survey of hallucination detection and mitigation methods wrapped in a formal learning-theory framework. The central formal claim, Eq. (1), is a standard PAC-Bayes inequality restated from McAllester [38], and its application to hallucination hinges on an oracle f*(x) that is not well-defined for most of the tasks the paper discusses. This is a load-bearing conceptual flaw rather than a local presentation issue: fixing it would require either restricting the scope to tasks with unique gold answers (which would exclude summarization and dialogue, both explicitly targeted) or redefining the risk as an expectation over a distribution of annotator labels or reference outputs, which is a substantial reformulation. The lack of any new derivation or empirical validation further weakens the case for publication as a research contribution. If the venue is open to survey papers, the authors should reposition the manuscript as a survey and remove the claim to new theoretical foundations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a competent survey wearing a theory-paper costume, and the theory part doesn't hold. The definitions come from Maynez et al. and Ji et al., the PAC-Bayes bound is McAllester's inequality restated, and the Rademacher bound is standard; the only genuinely new element is the workflow diagram. To the paper's credit, it calls Eq. (1) \"a direct application\" and later admits the bound is \"mostly conceptual,\" but the abstract's \"derive bounds\" still oversells what is here.\n\nWhat the paper does well: Sections IV, V, and VII give an accurate, well-cited map of detection methods (semantic entropy, self-consistency, calibration, attention checks), mitigation (RAG, fine-tuning, constrained decoding, verification modules), and evaluation (TruthfulQA, FactCC, QAGS, human scales). Someone new to the hallucination literature would get a fair picture. The reference list is appropriate; I don't see a citation-pattern problem.\n\nThe soft spots are real, and one is load-bearing. Section II.B assumes a unique oracle f*(x) that returns \"a fully truthful and contextually appropriate\" output, and defines hallucination as deviation from it. But Section II.A, two paragraphs earlier, concedes that whether content counts as a hallucination depends on task context and expected scope. For single-answer QA the oracle is defensible; for the summarization and grounded-dialogue settings the paper explicitly targets, there are many equally faithful outputs, so which f* you pick is a convention, not a measurement. That makes the empirical rate R-hat_hall convention-dependent, and Eq. (1), correct as math, vacuous for the quantity it claims to bound. The paper never confronts this. Neither bound is derived in any self-contained sense, the workflow has no implementation or empirical check, and Section III.C leans on an unpublished arXiv preprint for its impossibility claim.\n\nI think the reader's rejection is about right. There is no new result here that needs referee time, and the formal core doesn't survive contact with the open-ended tasks the paper itself cares about. I'd engage with the author constructively rather than dismissively — the survey writing shows clear thinking and honest engagement with the literature — but I wouldn't send this to review as a research paper. It's useful as a cautionary teaching example of how not to apply PAC-Bayes to a semantic judgment; I wouldn't spend a reading group on it.","headline":"A solid survey of hallucination detection and mitigation wrapped in a theory framing that doesn't hold: the PAC-Bayes bound is standard math applied to a 0-1 loss that is only well-defined when a unique correct output exists.","tokens_in":16488,"tokens_out":6532,"would_cite":false,"duration_ms":67371,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T50","68Q32","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Hallucination can be framed as a measurable risk with formal bounds.","keywords":["LLM hallucination","hallucination risk","PAC-Bayes bound","Rademacher complexity","retrieval-augmented generation","factuality evaluation","hallucination detection","mitigation workflow"],"falsifier":"Take a set of open-ended or subjective prompts, have independent annotators label each model output as hallucinated or faithful, and measure inter-annotator agreement. If agreement is near chance, the 0-1 oracle assumption fails for that input distribution, and the PAC-Bayes bound does not apply. A more direct check: on a distribution with a well-defined oracle, compute the bound's right-hand side from a training sample and compare it with the observed hallucination rate on a held-out sample; if the held-out rate systematically exceeds the bound, the claimed inequality would need re-examination.","tokens_in":15556,"feed_emoji":"🤖","tokens_out":7415,"duration_ms":73086,"temperature":0.7,"pith_summary":"The paper tries to put LLM hallucination on a formal footing by defining intrinsic hallucinations (output contradicts the source) and extrinsic hallucinations (output adds unverifiable content), then defining a hallucination risk as the expected probability that a model's output is unfaithful to a ground-truth oracle. Its main theoretical claim is that this risk is bounded above by the empirical hallucination rate plus a complexity penalty, via PAC-Bayes and Rademacher arguments, so that a low observed hallucination rate on a finite sample gives a high-probability guarantee about future inputs. The paper also argues, citing an impossibility result, that completely eliminating hallucination from a general-purpose model is not feasible, which reframes the goal as reducing and managing risk. The rest of the work surveys detection techniques, mitigation techniques, and evaluation protocols, and proposes a generate-detect-mitigate workflow. A sympathetic reader would care because this turns an elusive failure mode into a quantity one can measure, bound, and optimize.","feed_headline":"LLM hallucination gets a formal risk bound","feed_subtitle":"Low hallucination rates can generalize—but only when a ground-truth oracle exists.","key_machinery":"The load-bearing object is the hallucination indicator $H(M,x)$, which returns 1 when the output $y=M(x)$ is not completely faithful to the oracle $f^*(x)$, and the hallucination risk $R_{\\mathrm{hall}}(M)=\\mathbb{E}_x[H(M,x)]$ built from it. The argument is carried by applying the PAC-Bayes generalization bound for 0-1 loss to this indicator: the expected risk of a model distribution $Q$ is bounded by its empirical hallucination rate plus a term depending on the KL divergence between $Q$ and a prior $P$ and on the sample size. The same indicator permits Rademacher and VC-dimension complexity bounds. A separate impossibility argument for computable oracles supplies the negative result that some hallucinations are unavoidable, which is what motivates the paper's detection and mitigation workflow rather than a search for a perfectly truthful model.","core_discovery":"The central claim is that, under the assumption of a ground-truth oracle $f^*(x)$ that labels each input with a fully truthful output, hallucination becomes a binary 0-1 classification error, and the model's expected hallucination risk $R_{\\mathrm{hall}}(M)=\\mathbb{E}_x[H(M,x)]$ can be controlled by sample-based bounds. The paper derives the PAC-Bayes bound $\\mathbb{E}_{M\\sim Q}[R_{\\mathrm{hall}}(M)] \\le \\mathbb{E}_{M\\sim Q}[\\hat{R}_{\\mathrm{hall}}(M)] + \\sqrt{\\frac{\\mathrm{KL}(Q\\parallel P)+\\ln(1/\\delta)}{2n}}$, holding with probability at least $1-\\delta$, and it invokes Rademacher/VC uniform-convergence bounds of the form $R_{\\mathrm{hall}}(M) \\le \\hat{R}_{\\mathrm{hall}}(M)+O(\\sqrt{C/n})$. It further claims that for sufficiently powerful computable models no perfect reproduction of an arbitrary oracle is possible, so hallucinations are unavoidable in full generality. The paper presents these results as the theoretical foundation for the practical program: measure hallucination rates, bound them, detect likely errors, and mitigate them with retrieval, fine-tuning, calibrated decoding, and verification.","pith_inferences":["The oracle assumption suggests the formal bounds apply cleanly to closed tasks with verifiable answers, not to creative or subjective generation; extending them would require replacing the 0-1 oracle with graded human judgments, which would change the bound's meaning.","A natural next step the paper leaves implicit is to measure the bound's looseness on real LLMs by reporting the empirical rate, the KL term, and the held-out rate together; if the gap is large, data-dependent priors or tighter PAC-Bayes bounds could make the guarantee practically useful.","The generate-detect-mitigate loop can be tested as an engineering tradeoff: end-to-end hallucination rate versus latency and retrieval cost, with an optimal number of verification loops.","If the impossibility claim is right, retrieval-augmented generation does not remove hallucination but shifts it to cases where retrieval is incomplete or misread, so benchmarks should include unanswerable and out-of-knowledge queries."],"forward_implications":["A model with a low hallucination rate on a representative sample, and with a small KL penalty relative to its prior, comes with a formal guarantee that its true hallucination risk is also low, provided the inputs have a well-defined oracle.","Because the complexity term shrinks as $1/\\sqrt{n}$, collecting more labeled samples tightens the guarantee on hallucination risk at a known rate.","Fine-tuning on high-quality factual responses and keeping the model close to its pre-fine-tuning prior directly reduces the KL term, improving the bound.","The claimed impossibility of perfect truthfulness implies that retrieval, verification, and refusal are not optional extras but necessary components of any reliable deployment.","The proposed workflow formalizes the practical consequence: generate a draft, check it with uncertainty, factuality, and source-alignment signals, and then retrieve, correct, or abstain before final delivery."],"supporting_citations":[{"why":"supplies the PAC-Bayes 0-1-loss generalization bound that Eq. 1 applies to hallucination.","marker":"[38]"},{"why":"supplies the Rademacher complexity bounds used for the sample-dependent generalization estimate.","marker":"[36]"},{"why":"supplies the VC-dimension and statistical learning theory basis for the uniform-convergence bound.","marker":"[37]"},{"why":"supplies the result that hallucinations are unavoidable for sufficiently general computable models.","marker":"[11]"},{"why":"supplies the intrinsic/extrinsic hallucination taxonomy and the summarization faithfulness annotations.","marker":"[5]"},{"why":"supplies the survey-level taxonomy of hallucination in natural language generation referenced throughout.","marker":"[4]"},{"why":"supplies retrieval-augmented generation as the main grounding-based mitigation strategy.","marker":"[17]"},{"why":"supplies TruthfulQA, the open-domain truthfulness benchmark recommended for evaluation.","marker":"[28]"},{"why":"supplies FactCC, an automatic factual-consistency verifier used as a detection and evaluation metric.","marker":"[24]"},{"why":"supplies the chain-of-verification approach that underlies the paper's verify-and-edit mitigation step.","marker":"[34]"}],"fun_headline_variants":["Formal risk bound for LLM hallucination","Hallucination risk bounded, but only with an oracle","PAC-Bayes bound reveals hallucination limits","No perfect model avoids hallucination, theory proves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire bound depends on there being a fixed ground-truth oracle that can label every output as hallucinated or faithful; for open-ended, subjective, or creative tasks no such oracle exists, so the empirical hallucination rate the bound starts from is undefined.","fun_headline_variants_meta":{"raw":{"variants":["Formal risk bound for LLM hallucination","Hallucination risk bounded, but only with an oracle","PAC-Bayes bound reveals hallucination limits","No perfect model avoids hallucination, theory proves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1693,"prompt_tokens":991,"completion_tokens":702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":641}},"tokens_in":607,"tokens_out":702,"duration_ms":7035,"temperature":1.0,"reasoning_tokens":641,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:42:16.304131+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of open-ended or subjective prompts, have independent annotators label each model output as hallucinated or faithful, and measure inter-annotator agreement. If agreement is near chance, the 0-1 oracle assumption fails for that input distribution, and the PAC-Bayes bound does not apply. A more direct check: on a distribution with a well-defined oracle, compute the bound's right-hand side from a training sample and compare it with the observed hallucination rate on a held-out sample; if the held-out rate systematically exceeds the bound, the claimed inequality would need re-examination.","supporting_citations":[{"cited_title":"PAC-Bayesian model averaging,","cited_arxiv_id":null,"evidence_quote":"supplies the PAC-Bayes 0-1-loss generalization bound that Eq. 1 applies to hallucination."},{"cited_title":"Rademacher and gaussian complexi- ties: Risk bounds and structural results,","cited_arxiv_id":null,"evidence_quote":"supplies the Rademacher complexity bounds used for the sample-dependent generalization estimate."},{"cited_title":"Vapnik, Statistical Learning Theory","cited_arxiv_id":null,"evidence_quote":"supplies the VC-dimension and statistical learning theory basis for the uniform-convergence bound."},{"cited_title":"Layer-Adapted Implicit Distribution Alignment Networks for Cross-Corpus Speech Emotion Recognition","cited_arxiv_id":"2310.03992","evidence_quote":"supplies the result that hallucinations are unavoidable for sufficiently general computable models."},{"cited_title":"On faithfulness and factuality in abstractive summarization,","cited_arxiv_id":null,"evidence_quote":"supplies the intrinsic/extrinsic hallucination taxonomy and the summarization faithfulness annotations."},{"cited_title":"Survey of hallucination in natural language generation,","cited_arxiv_id":null,"evidence_quote":"supplies the survey-level taxonomy of hallucination in natural language generation referenced throughout."},{"cited_title":"Retrieval-augmented generation for knowledge-intensive NLP,","cited_arxiv_id":null,"evidence_quote":"supplies retrieval-augmented generation as the main grounding-based mitigation strategy."},{"cited_title":"TruthfulQA: Measuring how models mimic human falsehoods,","cited_arxiv_id":null,"evidence_quote":"supplies TruthfulQA, the open-domain truthfulness benchmark recommended for evaluation."},{"cited_title":"Evaluating the factual consistency of abstractive text summarization,","cited_arxiv_id":null,"evidence_quote":"supplies FactCC, an automatic factual-consistency verifier used as a detection and evaluation metric."}],"review_version":1}