{"id":"d62c8053-a079-4c5a-bc6b-c9add0022372","arxiv_id":"2411.19655","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LLM-OASIS provides 81,275 pairs of factual and subtly falsified passages from Wikipedia, plus a human-validated gold benchmark on which GPT-4o reaches only about 60% balanced accuracy.","lead":"A new dataset called LLM-OASIS contains over 81,000 pairs of factual and subtly falsified Wikipedia-based passages, designed to train systems that judge whether a text is true. Early tests show large language models struggle with the task, with GPT-4o scoring only about 60% balanced accuracy, suggesting the resource can drive better factuality checks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gold-standard labels are inferred from component checks, never from a direct human read of the final text; no human baseline calibrates Task 1.","rationale":"The paper's headline contribution is a training resource plus a human-validated gold benchmark. The training pairs inherit labels from the generation pipeline, and the gold benchmark is a filtered subset of those pairs. The filtering uses human annotations from four component tasks (Claim Extraction, Claim Falsification, Factual Text Gen, Unfactual Text Gen), but none of these tasks asks annotators to read the final text and decide whether it is factually accurate. Task 4 (Unfactual Text Generation) checks only that the text 'corresponds to the claims'; it does not check for invented extra content or for overall truthfulness. Therefore, the statement in Sec. 4.2 that 'we rely on human annotators... to create a gold standard test set' overstates what was validated: the gold labels are construction-based, inferred from component-level checks. This matters because if a portion of the unfactual texts contain unintended additional errors (or the falsified claim is not actually false in context), the binary labels are still correct but the task is no longer the 'single subtle falsification' the paper intends; models may be exploiting these confounds. Conversely, a direct human baseline is absent, so we cannot tell whether GPT-4o's ~60% is poor in an absolute sense. The proposed check—an independent, blinded annotation of final texts—would settle whether the gold labels agree with holistic human judgment and how large the human headroom is. This does not reject the paper: the resource is released, the pipeline is transparent, and the concerns are empirically testable. It does mean the 'human-validated gold standard' phrasing should be tempered unless such a test is run. Since the reader's verdict was already CONDITIONAL on external validation, this stress-test does not change the verdict, but it sharpens the condition.","tokens_in":30550,"tokens_out":10060,"duration_ms":85424,"concrete_test":"Sample 200 factual and 200 unfactual texts from the gold benchmark (Sec. 4.2). Recruit annotators who were not involved in dataset construction; give them only the text (no pipeline information) and ask for a binary Factual/Not Factual judgment, with optional web access or Wikipedia look-up. Compute (a) agreement (Cohen's kappa) between annotator labels and the benchmark's construction-based labels, and (b) human balanced accuracy. If human-vs-benchmark agreement falls below ~95%, or human accuracy fails to substantially exceed GPT-4o's 60.8% (Table 5), then the gold labels are not independently validated, and all downstream model comparisons need to be re-interpreted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The gold benchmark's binary labels are assigned by construction rather than by direct human reading of the final texts. In Sec. 4.2, an unfactual text is labeled False if (i) it is generated from a valid claim set, and (ii) the falsified claim is properly falsified and paraphrased; the annotators in Task 4 (Sec. 4.1, Appx E.4) are asked whether the generated text 'corresponds to the claims', not whether the text itself is factually inaccurate. Consequently, the central validity claim—that the benchmark is a 'human-validated gold standard' for end-to-end factuality—has not actually been tested end-to-end. If annotators never independently judge the final text's truthfulness, subtle errors introduced during paraphrase (e.g., an added detail not present in the claim list) or an unfactual claim that is not in fact false can slip through, and the reported GPT-4o 60% / Llama-3.3-70B 61.7% accuracies may measure performance on labels that differ from what an independent human would assign. The lack of any human baseline on Task 1 (Sec. 4.2) also prevents us from determining whether the benchmark is 'challenging' because LLMs are poor at factuality or because the labels are noisy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LLM-OASIS, a large synthetic resource for factuality evaluation built from popular English Wikipedia pages. The authors use GPT-4 to extract atomic claims from 5-sentence passages, falsify one claim per passage, and generate a factual paraphrase and an unfactual paraphrase, yielding 81,275 paired texts (plus 681,201 claims). They create a human-filtered gold benchmark of about 2,000 instances and define two tasks: end-to-end binary factuality evaluation of a whole text (Task 1) and evidence-based claim verification (Task 2). They benchmark several open-weight LLMs and GPT-4o under zero-shot, few-shot, explain-then-answer, and RAG settings, and build a modular pipeline (T5 claim extractor, E5 retriever, DeBERTa NLI verifier) trained on LLM-OASIS. The headline results are that GPT-4o reaches only 60.8% balanced accuracy on Task 1 zero-shot (68.0% in RAG), and the fine-tuned modular pipeline reaches 69.24% on Task 1 and 93.30% on Task 2.","tokens_in":30823,"tokens_out":7324,"duration_ms":61648,"significance":"If the benchmark is valid, this is a substantial contribution: LLM-OASIS is an order of magnitude larger than prior factuality resources such as FACTOR and FELM, it addresses the limitation of claim-level evaluation by targeting whole-text factuality, and it ships with two evaluation tasks, human component-level validation, extensive LLM baselines, and released code and data. The paper is also transparent about several limitations, including the silver nature of the training data and the reliance on Wikipedia. The main uncertainty is whether Task 1 measures factuality or the ability to detect GPT-4 generation artifacts, and whether the gold labels are truly human-validated in the end-to-end sense claimed by the paper.","major_comments":[{"comment":"The gold-standard labels for Task 1 are assigned by construction, not by a direct human read of the final text. Section 4.2 labels a retained unfactual text as False if it is generated from a valid claim set and if the falsified claim is properly falsified and paraphrased, while Appendix E.4 instructs annotators to check whether the unfactual text 'corresponds to the claims', not whether the text itself is factually accurate. This leaves open the possibility that a paraphrase introduces an unintended factual error, or that a 'falsified' claim is still true, so the binary label of the final text is not directly validated. The phrase 'human-validated gold standard' therefore overstates what was measured, and the reported GPT-4o 60.8% and Llama-3.3-70B 61.7% accuracies are relative to labels that have not been verified by humans end-to-end.","section":"Sec. 4.2, Appendix E.4"},{"comment":"No human baseline or human upper bound is reported for Task 1. Section 4.2 and Table 3 report accuracies and inter-annotator agreement for the four pipeline components, but annotators are never asked to perform the end-to-end binary factuality judgment that LLMs are evaluated on. Without such a baseline, the 'significant challenge' claim is underdetermined: the low LLM accuracies could reflect label noise, generation artifacts, or genuine task difficulty. Please add a human-annotator study on the final gold benchmark, ideally with per-class accuracy, and report it alongside the model results.","section":"Sec. 4.2, Table 3"},{"comment":"The fine-tuned modular system is trained on LLM-OASIS silver data and evaluated on a gold benchmark that is a human-filtered subset of the same GPT-4 pipeline. The claim extractor, retriever, and verifier (Secs. 5.1-5.3) are trained on data generated by the prompt in Table 1, and the gold benchmark in Sec. 4.2 is built by filtering instances from that same generation process. The reported Task 1 score of 69.24% and Task 2 score of 93.30% may therefore partly reflect distributional familiarity with GPT-4 paraphrase and falsification patterns rather than factuality. This is not a logical circularity because the gold labels are human-filtered, but it is a distributional overlap that limits external validity. Please evaluate the pipeline on out-of-distribution data, such as human-annotated outputs from a different LLM or human-written factual and unfactual texts, or alternatively compare with a system trained on data produced by a different generator.","section":"Sec. 5.1-5.3, Sec. 4.2"},{"comment":"The data-generation design may introduce surface artifacts that models can exploit instead of reasoning about factuality. Step 4 in Table 1 instructs GPT-4 to make the unfactual text 'as similar as possible' to the factual text except for the unfactual part, and the examples in Sec. 3.3 show unfactual texts that differ from their factual counterparts at exactly one segment. This single-falsified-claim, minimal-edit construction means that unfactual texts may contain local lexical or syntactic anomalies at the altered segment. The paper should test for this possibility, for example by evaluating models on unfactual texts with multiple planted errors, or by checking whether accuracy degrades when the minimal-edit constraint is relaxed.","section":"Table 1, Sec. 3.3"}],"minor_comments":[{"comment":"In Step 4 of the prompt, the instruction says the output text 'must be as similar as possible to the output of Step 2', but Step 2 returns a pair of claims, not a text; this should almost certainly refer to Step 3, and the current wording is confusing.","section":"Table 1"},{"comment":"The abstract states that GPT-4o achieves 'up to 60% accuracy', but Table 5 reports 68.0% in the RAG setting; the abstract should be reconciled with the reported results.","section":"Abstract, Table 5"},{"comment":"Section 7.2 says that end-to-end verification involves 'reading and reasoning on a huge context (4k tokens on average)', which appears inconsistent with Table 2's average of 86.5 tokens for unfactual texts; if 4k refers to the RAG setting with 30 retrieved passages, this should be stated explicitly.","section":"Sec. 7.2, Table 2"},{"comment":"The annotation guidelines contain a duplicated bullet defining the '<v>, <x>' combination; the duplicate should be removed.","section":"Appendix E.3"},{"comment":"The claim-falsification annotation example is referred to as 'Table 13' in the text, but the table caption in the manuscript is numbered 'Table 12' and labeled as task 3; the cross-reference and task label should be corrected.","section":"Appendix E.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a CL venue, and the resource is likely to be useful to the community. The main risk is that the headline difficulty numbers are interpreted as measuring factuality when they may partly reflect the limited distribution of GPT-4-generated texts and the absence of a direct human end-to-end label. I would encourage the editor to require the additional human end-to-end validation and an out-of-distribution evaluation before acceptance, rather than rejecting the paper. I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a useful resource paper. LLM-Oasis gives the field something it didn't have: roughly 81k paired factual/unfactual passages large enough to train end-to-end factuality evaluators, plus a human-filtered benchmark. The construction pipeline is transparent, the release is real, and the limitations section is honest. I'd send it to review.\n\nWhat's genuinely new: the scale (81k pairs vs 4k for FACTOR and 847 for FELM), and the decision to target whole-text evaluation rather than claim-only verification. The two-task setup (end-to-end and evidence-based claim verification) is sensible, and the deliberate single-claim falsification makes the task harder than FEVER-style isolated claims. The human evaluation of each pipeline step is reasonably careful: five linguists, agreement in the 0.72-0.84 range, and the gold benchmark only keeps instances that no annotator flagged. That's above the usual bar for resources of this kind.\n\nThe soft spots are real but not fatal. The main one: the gold benchmark's final labels are assigned by construction, not by direct human reading of the finished text. Annotators check that the claims correspond to the source, that the falsified claim is unfactual, and that the generated text reflects the claims. But nobody reads the final unfactual passage and independently says \"this is false.\" So subtle paraphrase-injected errors, or a falsified claim that annotators missed, could make some labels wrong, and we can't estimate how often. This also means the reported \"challenging\" numbers (GPT-4o ~60%, fine-tuned pipeline ~69%) are not calibrated against a human upper bound. A small human-reading study on a sample of the gold set would fix this and is worth asking for.\n\nSecond, the fine-tuned pipeline is trained and evaluated on data from the same GPT-4 pipeline, so its 69.24% and 93.30% scores partly reflect distributional familiarity. The paper acknowledges this in passing but could be more explicit. Third, GPT-4o results have no variance estimates (single run), which is a minor reporting gap.\n\nWho's this for? Anyone building or benchmarking factuality detectors for open-domain text. The resource is the main deliverable; the specific numbers are secondary. I'd accept for peer review with the request above: a human end-to-end label check and an external validation set from real LLM outputs. Without that, the benchmark is still usable, but the \"gold\" label needs that caveat.","headline":"Large, carefully built training resource for end-to-end factuality evaluation; the gold benchmark is the soft spot because the final labels are never directly human-read.","tokens_in":31366,"tokens_out":2140,"would_cite":true,"duration_ms":26255,"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":"LLM-OASIS introduces 81,275 paired factual and subtly falsified passages, on which state-of-the-art LLMs like GPT-4o reach only 60% accuracy in end-to-end factuality evaluation, while small fine-tuned models outperform them.","keywords":["factuality evaluation","hallucination detection","end-to-end evaluation","claim extraction","natural language inference","benchmark","large language models","Wikipedia"],"falsifier":"Train a model on LLM-OASIS and evaluate it on a held-out set of human-annotated unfactual texts from real LLM outputs (not GPT-4-generated pairs); if balanced accuracy drops to chance, the resource measures artifact detection rather than factuality.","tokens_in":30349,"feed_emoji":"🔍","tokens_out":9110,"duration_ms":63291,"temperature":0.7,"pith_summary":"LLM-OASIS is a dataset of 81,275 pairs of factual and subtly unfactual texts, built by extracting atomic claims from Wikipedia, falsifying one claim per passage, and generating two paraphrases that differ only in that claim. The paper argues this is the largest resource for training end-to-end factuality evaluators—systems that judge whether a whole text is truthful, rather than verifying isolated claims. On the resulting gold benchmark, state-of-the-art LLMs stall near 60% balanced accuracy (GPT-4o reaches 60.8% zero-shot and 68% with retrieval), while a small fine-tuned pipeline based on the resource reaches 69.24%. The takeaway is that whole-text factuality evaluation is a distinct, still-unsolved task and that dedicated training data can help smaller models beat much larger general-purpose LLMs.","feed_headline":"GPT-4o catches only 60% of subtle factual errors in texts","feed_subtitle":"The largest training set for whole-text factuality evaluation shows small fine-tuned models beat much larger LLMs.","key_machinery":"The central mechanism is a four-step generation pipeline executed in a single GPT-4 prompt: (1) atomic claim extraction from a Wikipedia passage, (2) subtle falsification of one claim, (3) factual paraphrase generation grounded in the original claims, and (4) unfactual text generation grounded in the claims with the falsified one, instructed to be 'as similar as possible' to the factual text except for the unfactual part. This isolates the factual error as the only difference between paired texts. The same pipeline supplies training data for a modular evaluator: a T5-base claim extractor, an E5-base dense retriever trained with DPR-style contrastive loss, and a DeBERTa-v3-large NLI-based claim verifier, whose per-claim outputs are aggregated to label the whole text as factual or not.","core_discovery":"The central claim is that end-to-end factuality evaluation—deciding whether a natural-language text contains any factual error—can be trained and benchmarked at scale, and that current LLMs are far from solving it. The authors construct LLM-OASIS by prompting GPT-4 to extract atomic self-contained claims from Wikipedia passages, falsify one claim with a subtle edit (forbidding changes to dates, numbers, names, or trivial negation), and then generate a factual paraphrase plus an unfactual version that is instructed to be 'as similar as possible' except for the falsified content. Human annotators validate each stage and produce a gold benchmark of roughly 2,000 error-free instances. On that benchmark, the best open-weight LLM (Llama-3.3-70B) reaches 61.7% balanced accuracy in the end-to-end task, GPT-4o reaches 60.8% zero-shot and 68% with retrieval-augmented generation, and a fine-tuned pipeline of small models (T5-base claim extractor, E5-base retriever, DeBERTa-v3-large NLI verifier) attains 69.24%. The paper's claim is that this resource is the largest for the task and that its difficulty demonstrates real headroom in factuality evaluation.","pith_inferences":["The single-falsified-claim design makes the task a needle-in-a-haystack detection problem; real-world LLM hallucinations often contain multiple errors, so the benchmark may be harder or easier than practice depending on the failure mode.","If the GPT-4 unfactual texts carry subtle surface-level artifacts distinct from the factual versions, models could be scoring artifact detection rather than factuality; a control using off-topic but factual claims would test this.","The resource could support richer tasks than binary labeling, such as localizing the unfactual span or explaining the error, which would be more actionable for correcting LLM outputs.","Extending the pipeline to less popular Wikipedia pages or non-Wikipedia corpora would test whether the difficulty persists for rare facts, a limitation the paper's discussion acknowledges."],"forward_implications":["Factuality evaluation can be treated as an end-to-end task on whole texts, not just claim verification, and can be trained on synthetic pairs at scale.","State-of-the-art LLMs remain far from reliable at this task (GPT-4o around 60% zero-shot), giving clear headroom for future work.","Small, specialized models fine-tuned on LLM-OASIS outperform much larger general LLMs on both end-to-end evaluation and evidence-based claim verification.","Because the generation pipeline only needs raw texts, it can be adapted to other domains and languages beyond English Wikipedia.","The gold benchmark provides a fixed, human-validated testbed for comparing factuality evaluators across settings."],"supporting_citations":[{"why":"FEVER, the prior claim-verification benchmark that LLM-OASIS contrasts with and extends.","marker":"Thorne et al., 2018"},{"why":"FACTOR, the prompting framework for generating factuality benchmarks that this work scales up.","marker":"Muhlgay et al., 2024"},{"why":"FELM, a small factuality benchmark whose limited size motivates training-scale data.","marker":"Chen et al., 2023"},{"why":"FActScore, which decomposes factuality evaluation into claim extraction, retrieval, and verification, the architecture the paper follows.","marker":"Min et al., 2023"},{"why":"DPR, the dense retrieval method used to train the evidence retriever.","marker":"Karpukhin et al., 2020"},{"why":"DeBERTa, the encoder fine-tuned as the NLI-based claim verifier.","marker":"He et al., 2021"},{"why":"T5, the sequence-to-sequence model fine-tuned as the claim extractor.","marker":"Raffel et al., 2019"},{"why":"GPT-4, the model used for data generation and whose GPT-4o variant is the strongest closed-source baseline.","marker":"OpenAI et al., 2024"}],"fun_headline_variants":["LLMs flunk fact-checking: best only 60% accuracy","Small models beat GPT-4o at whole-text fact-checking","Largest factuality dataset shows LLMs still hallucinate","End-to-end factuality: LLMs hit 60% ceiling, small models pass"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark's validity rests on the assumption that the only meaningful difference between each factual and unfactual passage is the single falsified claim, with no other surface cues for models to exploit.","fun_headline_variants_meta":{"raw":{"variants":["LLMs flunk fact-checking: best only 60% accuracy","Small models beat GPT-4o at whole-text fact-checking","Largest factuality dataset shows LLMs still hallucinate","End-to-end factuality: LLMs hit 60% ceiling, small models pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000596,"raw_usage":{"total_tokens":2857,"prompt_tokens":1082,"completion_tokens":1775,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":1709}},"tokens_in":698,"tokens_out":1775,"duration_ms":11508,"temperature":1.0,"reasoning_tokens":1709,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:58:49.924920+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model on LLM-OASIS and evaluate it on a held-out set of human-annotated unfactual texts from real LLM outputs (not GPT-4-generated pairs); if balanced accuracy drops to chance, the resource measures artifact detection rather than factuality.","supporting_citations":[],"review_version":1}