{"id":"d5edbd0f-7e2b-49a7-8708-41145760d040","arxiv_id":"2505.20767","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CogniBench is a sentence-level benchmark that labels LLM inferences, explanations, and opinions as faithful or hallucinated, expanding hallucination evaluation beyond verbatim factual claims.","lead":"This paper introduces CogniBench, a dataset and framework that separates what an LLM says into factual versus cognitive statements and scores cognitive faithfulness against three legal-inspired criteria. It reports that LLMs are 4.6 times more likely to hallucinate on cognitive statements, and releases an auto-labeled dataset plus an 8B detector.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4.6x cognitive-hallucination finding rests on an annotation boundary that is not shown to be reproducible: Table 1's 96.19% is reported as a QA/IAA figure rather than a chance-corrected pairwise agreement, and Figure 10 shows similar causal inferences receiving opposite labels.","rationale":"The reader's weakest_assumption is the same one I regard as most load-bearing: the operationalization of the three-tier annotation framework. The 64.8%-vs-13.9% comparison is the paper's headline discovery and the motivation for CogniBench-L and CogniDet; if the Speculative/Unequivocal boundary is not reproducible, every derived statistic inherits the instability. The paper's own Table 1 does not supply a proper chance-corrected IAA, and Figure 10 gives concrete reason to doubt boundary stability. I considered the auto-labeling contamination issue raised by the reader (prompt examples drawn from the same human annotations later used for evaluation, Section 3.1 and Table 5) and regard it as a serious secondary problem that also warrants a held-out evaluation, but the annotation-framework question is more fundamental because it threatens the human ground truth itself. For that reason I do not adjust the reader's CONDITIONAL verdict: the paper needs a proper reproducibility study before the central quantitative claims can be accepted, but the framework and dataset remain potentially valuable if that evidence is supplied. The suggested kappa test is the direct way to settle whether the concern lands.","tokens_in":18184,"tokens_out":9670,"duration_ms":97884,"concrete_test":"Draw a stratified random sample of 200 cognitive sentences from the human-annotated CogniBench set. Have two annotators who did not participate in the original labeling independently re-annotate them using only the Section 2.2 sequential protocol, then compute Cohen's kappa on the four-way cognitive label (Misleading, Speculative, Reliable, Unequivocal) and the pairwise confusion between Speculative and Reliable/Unequivocal. If kappa is below 0.6, or if more than 10% of sentences swap between Speculative and Reliable/Unequivocal, the 64.8% cognitive hallucination rate and the 4.6x ratio are not stable enough to support the paper's headline claim, and the statistics should be reported with confidence intervals or the boundary definitions revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Section 4.4) is that cognitive statements hallucinate at 64.8% versus 13.9% for factual statements. This number is computed from human labels assigned under the Rational/Grounded/Unequivocal framework of Section 2.2, with Speculative counted as hallucination. For that 4.6x ratio to be a finding about LLMs rather than about the annotators, the boundary between Speculative and Reliable/Unequivocal must be reproducible. The paper does not establish this. Table 1 reports 96.19% 'IAA' alongside QA instances, but no independent double-annotation sample is described and no chance-corrected metric such as Cohen's kappa is given; the third-reviewer adjudication described in A.5.2 is not an agreement measure. Figure 10 illustrates the fragility: 'Their success and reputation attract top talent...' is labeled Speculative because the context does not directly prove that success attracts talent, while 'This policy helps prevent underage customers from accessing their products' is labeled Unequivocal even though the context states only an age restriction and does not state the policy's purpose. Both are causal inferences from context, and no operational rule in Section 2.2 distinguishes them. Because 62.2% of cognitive statements fall in the Speculative category, a small shift in this boundary materially changes the 64.8% rate, the 4.6x gap, and every downstream comparison that uses these labels as ground truth.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a legal-inspired, three-tier faithfulness framework (Rational, Grounded, Unequivocal) for sentence-level assessment of 'cognitive statements' (inferences, evaluations, explanations) in knowledge-grounded dialogues. It introduces CogniBench, a human-annotated multi-turn dialogue benchmark built from GPT-4/RefGPT-generated dialogues over Wikipedia contexts, and CogniBench-L, a roughly 24k-dialogue extension produced by an LLM-based auto-labeling pipeline that uses contrastive prompts and five-way majority voting. The authors report three headline findings: cognitive statements become more frequent as dialogues lengthen; LLMs hallucinate 64.8% of cognitive statements versus 13.9% of factual statements; and an 8B detector (CogniDet) fine-tuned on CogniBench-L reaches 70.3% F1, outperforming existing hallucination detectors. The paper also releases the datasets, code/model, and detailed annotation prompts.","tokens_in":18403,"tokens_out":4692,"duration_ms":49829,"significance":"If the annotation boundary is reproducible, the benchmark fills a real gap: existing faithfulness datasets focus on verbatim factual consistency, and a tool for evaluating inferential claims would be useful both for benchmarking and for training low-cost detectors. The public release of the datasets and model, together with the detailed prompts in Appendix A.7, supports reproducibility and follow-up work. The reported 4.6x gap between cognitive and factual hallucination rates is a striking, falsifiable empirical claim. However, the central claims rest on two premises that are not yet established: the reliability of the Speculative/Unequivocal boundary, and the validity of the auto-labeling evaluation. As a resource, the paper is still valuable, but the headline statistics need to be made robust to these premises.","major_comments":[{"comment":"The auto-labeling prompts are refined using human annotations from CogniBench itself: Section 3.1 describes an 'Initial Prompting and Diagnosis' step in which LLM annotation errors are identified on CogniBench data and used to construct contrastive examples, and Table 3 then reports an 82.2% F1 for auto-labeling on the same CogniBench set. This is an in-sample estimate, and the claim that auto-labeling 'closely matches human annotation' is not supported by this evaluation. Please construct the contrastive examples on a development split and evaluate on a held-out split, or clearly label the 82.2% figure as a development-set estimate and validate it on independently annotated data. The concern propagates to Section 4.7, where auto-labeling is used as a proxy for human annotation to compare GPT-4, Gemini-Pro, and other models.","section":"Section 3.1, Table 3, Section 4.6"},{"comment":"The reproducibility of the three-tier labels is not established. Table 1 reports a 96.19% 'IAA' for the sequential decision framework, but no definition of the metric, no confidence interval, no chance-corrected statistic such as Cohen's kappa, and no description of the sample beyond the small QA-instance counts. Figure 10 shows two causal inferences that receive different labels: 'This policy helps prevent underage customers from accessing their products' is labeled Unequivocal, while 'Their success and reputation attract top talent...' is labeled Speculative, with annotation reasons that do not provide an operational rule separating the two cases. Since Section 4.4 counts Speculative statements as hallucinations, the 64.8% cognitive hallucination rate and the resulting 4.6x ratio are directly sensitive to this boundary. Please report per-category pairwise agreement, chance-corrected agreement, and an analysis of how the headline rates change under alternative boundary decisions.","section":"Section 2.2, Table 1, Figure 10, Section 4.4"},{"comment":"The headline 'hallucination rate' conflates categories with different epistemic status. Speculative statements are defined as plausible but ungrounded inferences (satisfying only Criterion 1), and labeling all such statements 'hallucinations' is a design choice rather than a measured fact. The abstract and conclusion present the 64.8% figure as an LLM deficiency, but it depends on counting rationally plausible inferences as hallucinations. Please report the four categories (Misleading, Speculative, Reliable, Unequivocal) separately, explicitly state that the 64.8% statistic uses the 'Rational but not Grounded' definition, and include a sensitivity analysis (e.g., counting only Misleading statements as hallucinations) so readers can separate the empirical finding from the definitional choice.","section":"Section 4.4, Figure 6"}],"minor_comments":[{"comment":"The heading 'Matrices' should be 'Metrics'; also, the precision/recall formulas count words while the evaluation is described as sentence-level, so the unit of analysis should be clarified in the text.","section":"Section 4.1"},{"comment":"Figure 10 contains a spelling error ('Anbreakthroughcame') and the figure is dense; please clean the text and consider enlarging the examples or splitting the figure for readability.","section":"Figure 10, Section A.5.4"},{"comment":"The threshold values 2 and 3 used for majority voting are only defined in Appendix A.9; please define them when Table 5 is first cited in Section 4.6.","section":"Table 5, Section A.9"},{"comment":"The text says CogniBench-L is '100 times larger' than CogniBench, but 24084/264 is approximately 91; please state the exact factor or write 'about two orders of magnitude.'","section":"Abstract, Table 2"},{"comment":"Given the in-sample nature of the auto-labeling evaluation noted in Major Comment 1, the model-comparison numbers in Figure 8 should be accompanied by a caveat or by a validation of the auto-labeling proxy.","section":"Figure 8, Section 4.7"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the in-sample evaluation of the auto-labeling pipeline; this should be fixed before acceptance, either with a held-out split or with new independent annotations. The annotation-boundary issue is also central, but it is addressable with the existing double annotations by reporting chance-corrected agreement and robustness analyses. If the authors can provide those, the paper would be a useful benchmark contribution despite the limitations of the current headline claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about CogniBench. First, it identifies a genuine gap: faithfulness evaluation has focused on factual rephrasing while ignoring inferences, explanations, and opinions, and this paper makes a credible case that those 'cognitive statements' deserve separate scrutiny. Second, its headline numbers are not as solid as they look. The auto-labeling pipeline is tested on the same human annotations used to refine it, and the annotation boundary between Speculative and Unequivocal is not shown to be reproducible.\n\nThe good parts are real. The cognitive/factual split is a useful addition to faithfulness evaluation, and the three-tier framework (Rational, Grounded, Unequivocal) is a sensible way to match evaluation stringency to application risk. CogniBench itself is a human-annotated, sentence-level dataset that is new; CogniBench-L is a large auto-labeled corpus; and CogniDet is a trained detector that people can actually use. The observations that cognitive statements grow as dialogues lengthen and that LLMs hallucinate more on cognitive than factual statements are plausible and likely to influence how the field designs evaluation.\n\nThe soft spots are substantial but fixable. Section 3.1 describes how the authors collected common annotation errors and used them to craft contrastive examples for the auto-labeling prompt; Table 5 then evaluates auto-labeling on the same human-labeled CogniBench. That is tuning on the test set, so the reported 82.2% F1 does not establish proxy reliability. The paper also does not provide a genuine chance-corrected inter-annotator agreement. Table 1's 96.19% is based on 13 QA instances, not a full double-annotation sample. Figure 10 shows two causal inferences getting opposite labels, and the criteria give no operational rule for distinguishing 'Grounded' from 'Speculative'. Because 62.2% of cognitive statements fall into Speculative, small shifts in that boundary move the headline 64.8% hallucination rate and the 4.6x gap. And the definition of cognitive hallucination as misleading plus speculative is strict; the paper should be explicit that this is a standard-setting choice, not an empirical discovery. The limitations section is candid about domain specificity and legal analogy, but it does not mention this circularity.\n\nI would not desk-reject this. The dataset and framework have real value, and the failure mode they target is important for RAG systems in medicine, finance, and law. But the paper needs major revision before the empirical claims can be trusted: a held-out evaluation of auto-labeling, a proper kappa calculation on a larger double-annotation sample, and a clearer operationalization of the Speculative boundary. I suspect the overall direction is robust—cognitive statements are likely riskier than factual ones—but the exact ratio is not supported. With the revisions, this could become a solid resource. Worth a serious referee and a reading-group discussion.","headline":"A useful new dataset and framing of cognitive hallucination, but the headline percentages rest on circular auto-labeling evaluation and an unvalidated annotation boundary.","tokens_in":19073,"tokens_out":4804,"would_cite":true,"duration_ms":48782,"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":"Language models produce unsupported claims 4.6 times more often when they infer, evaluate, or explain than when they restate facts; this paper builds the measurement standard, dataset, and detector that reveal the gap.","keywords":["cognitive hallucination","faithfulness hallucination","knowledge-grounded dialogue","legal-inspired annotation framework","hallucination detection","LLM evaluation","sentence-level annotation","auto-labeling"],"falsifier":"Re-annotate a random sample of several hundred sentences from CogniBench with a fresh team of annotators using only the published protocol; if the Speculative/Unequivocal split—and with it the overall 64.8% cognitive hallucination rate—does not reproduce within a few percentage points, the headline gap is an artifact of the annotation scheme. A complementary check: run the trained CogniDet detector on out-of-domain, high-stakes dialogues (for example, medical discharge or financial advisory text) and see whether its cognitive-hallucination F1 collapses below the 73.8% reported on CogniBench.","tokens_in":17824,"feed_emoji":"🧠","tokens_out":8007,"duration_ms":67116,"temperature":0.7,"pith_summary":"Faithfulness benchmarks for large language models mostly check whether a model's words match the supplied context—rephrasing is easy to verify. This paper argues that the harder, more consequential statements are 'cognitive' ones: inferences, explanations, and evaluations that go beyond the text. Drawing on how courts weigh circumstantial evidence, it defines three escalating standards of faithfulness (Rational, Grounded, Unequivocal) and builds a human-annotated dataset, CogniBench, that applies them sentence by sentence. The headline result is that LLMs hallucinate cognitive statements at a 64.8% rate versus 13.9% for factual ones—a 4.6x gap that fact-only benchmarks cannot see. The paper then scales the protocol with an LLM-judge auto-labeling pipeline into CogniBench-L and trains an 8B-parameter detector, CogniDet, that reaches 70.3% overall F1 in a single forward pass.","feed_headline":"LLMs hallucinate 4.6x more when reasoning than restating","feed_subtitle":"A legal-inspired benchmark finds 64.8% of inferences unsupported, vs 13.9% of restated facts.","key_machinery":"The load-bearing device is the three-tier sequential decision framework adapted from legal treatment of circumstantial evidence. Annotators (or an LLM judge) classify each sentence by asking, in order: Is it rational (plausible, non-misleading speculation)? Is it grounded (logically supported by the context)? Is it unequivocal (admitting no reasonable alternative interpretation)? The sequence yields four categories—Misleading, Speculative, Reliable, and Unequivocal—and the paper shows that breaking the task into this ordered set of decisions raises inter-annotator agreement to 96.19% while cutting QA effort. A second mechanism is the auto-labeling pipeline: contrastive examples fix the boundary errors the LLM judge makes, sentence-marked formatting keeps annotation cheap, and majority voting over five independent GPT-4 judgments turns a single judge into a stable proxy for human annotation.","core_discovery":"The central claim is that LLM faithfulness cannot be assessed by factual consistency alone. The paper introduces a three-tier, legal-inspired standard for judging cognitive statements—statements that infer, explain, or evaluate beyond the provided context. A statement is Rational if it is plausible speculation, Grounded if it is logically derivable from the context, and Unequivocal if it is the only reasonable conclusion; the tier a statement satisfies determines whether it is Misleading, Speculative, Reliable, or Unequivocal. Using this protocol on knowledge-grounded multi-turn dialogues, the paper reports that 64.8% of cognitive statements are hallucinations (2.6% misleading, 62.2% speculative) compared with 13.9% for factual statements, that cognitive statements grow from about 15% to 50% of model output as conversations lengthen, and that existing hallucination detectors lose roughly 31% F1 on cognitive statements. The paper further claims that an automatic LLM-judge pipeline with contrastive prompting and five-way majority voting matches human annotation accuracy, and that a detector fine-tuned on the resulting 24k-dialogue corpus, CogniDet, achieves 70.3% overall F1 on CogniBench.","pith_inferences":["If the cognitive hallucination gap is real, retrieval-augmented applications that rely on model reasoning—medical triage, financial advice, legal analysis—are much less trustworthy than their fact-rephrasing accuracy suggests, and deployment tests should measure cognitive faithfulness rather than citing accuracy alone.","The three-tier standard suggests a testable hierarchy for application design: creative assistants could be held to the Rational tier, helpful assistants to the Grounded tier, and high-stakes decision systems to the Unequivocal tier; one could validate whether model failures concentrate in the tier an application actually requires.","A natural extension the paper does not run: apply the auto-labeling pipeline to domain-specific corpora (medical, financial, legal) to test whether the cognitive hallucination gap widens when the context is dense with specialized terminology, where a model's inferences are harder to verify."],"forward_implications":["Fact-only benchmarks systematically underestimate hallucination risk: cognitive statements hallucinate at 64.8% versus 13.9% for factual ones, a gap invisible to rephrasing checks.","Multi-turn evaluation is necessary: cognitive statements grow from about 15% of model output in early turns to about 50% in later turns, and both factual and cognitive hallucinations concentrate in longer dialogues.","Existing detectors degrade by about 31% F1 on cognitive statements, so hallucination detection needs training data that includes cognitive statements; CogniDet reaches 70.3% overall F1 in a single forward pass, versus 58.0% for the strongest prompting baseline tested.","The auto-labeling pipeline (contrastive prompting plus five-way majority vote) is claimed to match human annotation accuracy (82.2% overall F1), offering a low-cost proxy to audit newly released models without manual annotation."],"supporting_citations":[{"why":"Supplies the RefGPT dialogue-generation approach used to build the knowledge-grounded multi-turn conversations in CogniBench.","marker":"(Yang et al., 2023)"},{"why":"GPT-4 is the primary model for dialogue generation and serves as the LLM judge in the auto-labeling pipeline.","marker":"(Achiam et al., 2023)"},{"why":"Black's Law Dictionary provides the legal definitions of direct and circumstantial evidence that the three-tier framework adapts.","marker":"(Garner et al., 2004)"},{"why":"Supplies the 'beyond a reasonable doubt / only reasonable inference' standard behind the Unequivocal criterion.","marker":"(v Villaroman, 2016)"},{"why":"RAGTruth is the prior factual-hallucination corpus and detector that the paper contrasts with and uses as a baseline.","marker":"(Niu et al., 2023)"},{"why":"FAVA is the prior hallucination taxonomy and detector used as a baseline, and the source of the 'invented' label for factual hallucinations.","marker":"(Mishra et al., 2024)"},{"why":"SelfCheckGPT is a zero-resource black-box hallucination detector used as a baseline.","marker":"(Manakul et al., 2023)"},{"why":"Bloom's Taxonomy grounds the factual-versus-cognitive statement distinction used throughout the paper.","marker":"(Bloom et al., 1956)"},{"why":"Semantic-entropy hallucination detection motivates the multi-response sampling majority-vote step in auto-labeling.","marker":"(Farquhar et al., 2024)"}],"fun_headline_variants":["LLM reasoning unsupported 64.8% vs 13.9% factual claims","Legal-inspired benchmark exposes cognitive hallucination gap","Long conversations triple LLM cognitive output share","Existing detectors lose 31% F1 on cognitive hallucinations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire measurement rests on the boundary between a Speculative statement (plausible but unsupported) and an Unequivocal statement (the only reasonable conclusion): if different annotators, or the LLM judge, cannot reliably draw that line, then the 64.8% versus 13.9% hallucination gap is an artifact of the labeling scheme rather than a property of the models.","fun_headline_variants_meta":{"raw":{"variants":["LLM reasoning unsupported 64.8% vs 13.9% factual claims","Legal-inspired benchmark exposes cognitive hallucination gap","Long conversations triple LLM cognitive output share","Existing detectors lose 31% F1 on cognitive hallucinations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000514,"raw_usage":{"total_tokens":2501,"prompt_tokens":957,"completion_tokens":1544,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":1476}},"tokens_in":573,"tokens_out":1544,"duration_ms":12712,"temperature":1.0,"reasoning_tokens":1476,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:46:48.118395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a random sample of several hundred sentences from CogniBench with a fresh team of annotators using only the published protocol; if the Speculative/Unequivocal split—and with it the overall 64.8% cognitive hallucination rate—does not reproduce within a few percentage points, the headline gap is an artifact of the annotation scheme. A complementary check: run the trained CogniDet detector on out-of-domain, high-stakes dialogues (for example, medical discharge or financial advisory text) and see whether its cognitive-hallucination F1 collapses below the 73.8% reported on CogniBench.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Black's Law Dictionary provides the legal definitions of direct and circumstantial evidence that the three-tier framework adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the 'beyond a reasonable doubt / only reasonable inference' standard behind the Unequivocal criterion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Bloom's Taxonomy grounds the factual-versus-cognitive statement distinction used throughout the paper."}],"review_version":1}