{"id":"accbcf38-a51b-44fc-ab9d-4431e25fba89","arxiv_id":"2505.04844","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Fine-tuning Qwen2.5-7B on GPT-4o-generated perturbed multi-hop QA data yields recall 0.938 vs GPT-4o's 0.710 on RAGTruth hallucination detection, with lower precision (0.366 vs 0.446).","lead":"A Stanford team fine-tuned a 7B open-source model on a new multi-hop QA dataset with GPT-4o-generated hallucinations, reporting higher recall than GPT-4o on the RAGTruth benchmark. The claims are weakened by a confounded evaluation pipeline and missing precision and accuracy details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RAGTruth recall advantage may be an artifact of asymmetric GPT-4o mini post-processing; the paper never states whether the GPT-4o baseline receives the same repair step.","rationale":"The reader's weakest assumption is exactly the load-bearing issue I find: the two evaluation arms are not demonstrably symmetric because only Osiris outputs are described as receiving GPT-4o mini JSON repair. Table 3 is the sole empirical support for the paper's central claim, and a 7B model beating GPT-4o on hallucination-detection recall is surprising enough that evaluation symmetry must be established before the result is meaningful. This is a reporting and protocol gap, however, not a fundamental flaw in the approach: supervised fine-tuning on perturbed multi-hop data is a plausible route, the code is released, and the method is reproducible. Precision and F1 are in fact below GPT-4o (0.366 vs 0.446 and 0.527 vs 0.548), no other baselines or confidence intervals are reported, and the abstract's 'accuracy' claim is never backed by a number, which further supports a conditional rather than unconditional acceptance. The numeric inconsistency between '22.8%' and '23.8%' reinforces the need for an exact, transparent evaluation, but it does not by itself overturn the recall claim. Conditional acceptance with a required symmetric evaluation is the right outcome, so the reader's verdict does not change.","tokens_in":9801,"tokens_out":6284,"duration_ms":60995,"concrete_test":"Run a controlled RAGTruth evaluation with identical framing and parsing for both systems under four conditions: (1) both raw outputs parsed by the same rule-based extractor, (2) both outputs repaired by the GPT-4o mini JSON prompt, (3) only Osiris-7B repaired as in the paper, and (4) only GPT-4o repaired. The central claim survives only if the recall advantage persists in conditions (1) and (2). Additionally, recompute the claimed 22.8%/23.8% improvement from the underlying confusion matrix and report raw counts, precision, and an accuracy number as claimed in the abstract.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim (Osiris-7B recall 0.938 vs GPT-4o 0.710, Table 3) rests on an evaluation whose two arms are not shown to be symmetric. Section 2.2 says the fine-tuned models are used 'in conjunction with GPT-4o mini, employing a JSON prompt detailed in Appendix B.3 only when the output answers required correction,' but it never states whether GPT-4o's RAGTruth predictions were run through the same JSON repair or parsing step. The Limitations section concedes that small fine-tuned models often fail to produce correct JSON, so the repair step is not a no-op: the B.3 prompt explicitly converts free text, bullet lists, and malformed arrays into a uniform 'hallucination_list', which can directly increase measured recall. If the GPT-4o baseline was parsed with a stricter or different routine, the reported 22.8-point recall gain could be an artifact of post-processing rather than of the model. The numeric claim itself is also unstable: the abstract says '22.8%', Section 3 says '23.8%', and neither distinguishes absolute percentage-point difference (0.938 - 0.710 = 0.228) from relative improvement (32%). This does not disprove the method, but it makes the central comparison unverified without the exact harness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Osiris-7B, a Qwen2.5-7B model fine-tuned on a newly constructed perturbed MuSiQue dataset in which GPT-4o generates plausible but unsupported answers with reasoning labels. The authors evaluate the model on the human-annotated RAGTruth benchmark and report that the fine-tuned 7B model achieves recall 0.938 versus GPT-4o's 0.710, while also claiming competitive precision and accuracy and faster inference. The paper's contributions are an open-source fine-tuned detector, a data perturbation pipeline, and an evaluation against a closed-source model.","tokens_in":10027,"tokens_out":4880,"duration_ms":45093,"significance":"If the reported result holds, the paper would make a practically useful contribution: a small open-source model with higher recall than GPT-4o on a human-annotated hallucination detection benchmark, at lower inference cost, is of clear value to RAG deployment workflows. A notable strength is that the evaluation target, RAGTruth, is human-annotated and independent of the training labels, so the main recall claim is not circular by construction; the released code and dataset construction also help reproducibility. However, the central comparison currently rests on an underspecified and potentially asymmetric evaluation pipeline, and several headline claims in the abstract are not backed by the reported numbers.","major_comments":[{"comment":"The RAGTruth evaluation is described asymmetrically: the paper states that the fine-tuned models are used 'in conjunction with GPT-4o mini, employing a JSON prompt detailed in Appendix B.3 only when the output answers required correction,' but it never states whether the GPT-4o baseline receives the same JSON repair or parsing step. The Appendix B.3 prompt is not a no-op: it explicitly converts free text, bullet lists, numbered lists, and line-separated items into a uniform hallucination_list, and the Limitations section concedes that smaller fine-tuned models often fail to produce correct JSON. If the GPT-4o baseline was parsed with a different routine, the reported 0.938 vs 0.710 recall advantage could be an artifact of post-processing rather than of the model. I ask the authors to specify exactly which post-processing each arm received and to rerun both arms under identical post-processing.","section":"Section 2.2, Appendix B.3, Table 3"},{"comment":"The paper never defines the prediction format or the matching rule used to compute recall, precision, and F1 on RAGTruth, which is described as a word-level hallucination corpus. It is not stated whether predictions are span-level exact matches, token-level overlaps, or sentence-level binary decisions, nor how the hallucination_list output is aligned with the human-annotated spans. Without this definition, the headline numbers in Table 3 are not reproducible or interpretable. Please provide the evaluation script or a precise statement of the matching criterion and any thresholds.","section":"Section 2.2, Table 3"},{"comment":"The abstract claims 'competitive performance on precision and accuracy,' but Table 3 reports precision 0.366 for Osiris-7B versus 0.446 for GPT-4o (i.e., lower), and accuracy is never reported anywhere in the paper. The claim should either be removed or supported by a corresponding table entry and a definition of how accuracy is computed.","section":"Abstract, Table 3"},{"comment":"The headline recall improvement over GPT-4o is stated inconsistently: the abstract says 22.8%, while Section 3 and the Conclusion say 23.8%. Neither number is the relative improvement from 0.710 to 0.938 (about 32.1%), and the text does not distinguish absolute percentage points (22.8 points) from relative improvement. The headline metric should be stated consistently and unambiguously.","section":"Abstract, Section 3, Conclusion"}],"minor_comments":[{"comment":"The verified example text is garbled: 'confirming the the that the country Tepuka is located in (Tuvalu) was Álvaro de Mendaña' should be rewritten, and there is a duplicated 'the' earlier in the same paragraph.","section":"Section 2.1, Table 1"},{"comment":"The inference speed comparison is not apples-to-apples: Osiris-7B is measured with 4-bit quantization on an A100 80GB via vLLM, while GPT-4o's number is cited from a third-party source without the same hardware and serving conditions. Please state the conditions for both or clearly label the comparison as indicative only.","section":"Section 3, Inference Speed"},{"comment":"The contribution bullet 'average increases of 32.98% in recall, 4.55% in precision, and 17.98% in F1' appears to report average percentage-point changes across four model sizes, yet the 7B model's precision actually decreases (0.402 to 0.366). Please state the aggregation method explicitly and note the per-size direction of change.","section":"Abstract, Table 3"},{"comment":"Figure 1 has no axis labels or legend; adding them would make the token-length distribution comparison interpretable.","section":"Section 2.1, Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the asymmetric post-processing in the RAGTruth evaluation. If the authors can demonstrate that GPT-4o received the same JSON repair step, the main recall claim may survive; if not, the headline result is unverified. I would not accept without that clarification and a rerun of both arms under identical conditions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a legitimate extension of the Lynx line of work: take a perturbation pipeline for inducing hallucinations, apply it to MuSiQue multi-hop QA, fine-tune a small Qwen model, and test on RAGTruth. The dataset construction is described in enough detail to reproduce, the code is released, and the fine-tuning recipe is concrete. That is real value. The improvement over the base Qwen model is also credible and shows the method does something.\n\nThe problem is the headline comparison. The claim that Osiris-7B beats GPT-4o by 22.8% (or 23.8%? the paper says both) in recall rests on an evaluation harness that is not shown to be symmetric. Section 2.2 says the fine-tuned models are used in conjunction with GPT-4o mini, with a JSON repair prompt from Appendix B.3, but it never states whether the GPT-4o baseline receives the same post-processing. Given the Limitations section openly admits that small models often struggle with JSON formatting, the repair step is not a no-op; it can convert free text, bullet lists, and malformed output into a uniform hallucination_list, which directly inflates recall. If the baseline was parsed more strictly, the 22.8-point gap could be largely an artifact of the harness.\n\nOther soft spots: the abstract claims competitive precision and accuracy, but accuracy is never reported and precision drops from 0.446 to 0.366. The only baseline is GPT-4o; no other hallucination detector, open-source or closed, is compared. The paper also conflates absolute percentage points with relative improvement, and the 22.8% vs 23.8% inconsistency does not inspire confidence in the numbers. None of this disproves the method, but it means the central empirical claim is unverified without a clean, symmetric evaluation.\n\nOn the positive side, the RAGTruth labels are human-annotated, so the recall claim is not circular by construction. The citation pattern is fine, leaning heavily on preprints but appropriately crediting Lynx, MuSiQue, and RAGTruth. The writing is a bit rough but understandable.\n\nThis paper deserves a serious referee because the question it asks is important — can a small open model match or beat a frontier model on hallucination detection recall — and because the artifact is concrete. A referee should require a symmetric evaluation, explicit precision/accuracy numbers, at least one additional baseline, and a consistent reporting of the gain. With those revisions it could be a useful data point for the field.\n\nI would bring it to a reading group as an example of an evaluation pitfall, not as a strong result, and I would not cite it in its current form.","headline":"A reasonable extension of Lynx with a large recall claim over GPT-4o that is currently undermined by an asymmetric evaluation pipeline and inconsistent reporting.","tokens_in":10566,"tokens_out":2309,"would_cite":false,"duration_ms":23437,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a 7B model fine-tuned on a perturbed multi-hop QA dataset detects RAG hallucinations with higher recall than GPT-4o on the RAGTruth benchmark.","keywords":["hallucination detection","retrieval-augmented generation","multi-hop question answering","data perturbation","supervised fine-tuning","RAGTruth","MuSiQue","open-source LLM"],"falsifier":"Re-run the same RAGTruth evaluation twice: once with the JSON-repair post-processing applied to both Osiris-7B and GPT-4o, and once with it applied to neither. If Osiris's recall advantage disappears or reverses under symmetric scoring, the claimed edge is an artifact of asymmetric post-processing.","tokens_in":9578,"feed_emoji":"🎯","tokens_out":8281,"duration_ms":73217,"temperature":0.7,"pith_summary":"This paper tries to establish that a lightweight, openly available 7B model can beat a much larger closed model at the practical task of telling whether a RAG-generated answer is supported by the retrieved context. The authors build a training set from the multi-hop MuSiQue question set: correct answers are verified with reasoning, and GPT-4o is prompted to replace them with plausible but unsupported answers that can even be drawn from the same context. After fine-tuning Qwen 2.5 7B on that data, they report 0.938 recall on the RAGTruth benchmark, compared with GPT-4o's 0.710, with competitive precision and F1. If this holds, production RAG systems could screen for hallucinations locally, cheaply, and in real time instead of paying for or waiting on a closed reviewer model.","feed_headline":"7B open model catches more RAG hallucinations than GPT-4o","feed_subtitle":"Fine-tuned on perturbed multi-hop questions, it scores 93.8% recall on RAGTruth versus GPT-4o's 71%.","key_machinery":"The central object is the perturbed MuSiQue dataset: roughly 39,876 question-answer pairs, split about evenly between verified answers and induced hallucinations, each carrying a reasoning explanation that cites the context and says why the answer is or is not supported. The perturbations are deliberately subtle, replacing the gold answer with another entity mentioned in the same evidence, so surface overlap alone cannot decide correctness. Fine-tuning on these contrastive pairs is what transfers multi-hop verification skill into the detector; the evaluation then runs on RAGTruth, a word-level hallucination corpus built from real RAG outputs, which measures recall and precision of flagged hallucinated spans.","core_discovery":"On the paper's own terms, the discovery is that hallucination detection can be learned from a perturbed multi-hop QA dataset rather than from distilled single-hop examples. Osiris-7B, a Qwen2.5-7B model fine-tuned on the 39,876-example perturbed MuSiQue set, reaches recall 0.938 on RAGTruth while GPT-4o reaches 0.710, a gap the authors quote as 22.8% in the abstract and 23.8% in the conclusion; precision is 0.366 versus 0.446 and F1 is 0.527 versus 0.548. The paper argues that multi-hop structure matters: the model is forced to verify each reasoning step against the evidence, and the subtle perturbations, wrong answers that appear in the context but do not support the question, teach it to reject plausible but unsupported claims. The result is positioned as a practical detector, not a general fact-checker, with high recall so that human reviewers can focus on flagged spans.","pith_inferences":["The reported recall advantage could partly come from the JSON-repair post-processing applied to Osiris outputs but not, as far as the paper states, to the GPT-4o baseline; a symmetric evaluation is the direct test.","Because the same closed model that verifies correct answers also writes the hallucinated ones, the fine-tuned model may learn a house style of hallucination rather than a general capacity; checking against hallucinations produced by other generators would show generalization.","The 'open-source' label is partial, since perturbation generation and inference-time JSON repair both depend on closed models; a fully open pipeline would need a local source of perturbations and a local repair step."],"forward_implications":["A locally hosted 7B detector can screen RAG outputs at scale, removing the need to route every generation through a closed API for review.","The recall-oriented design means nearly all hallucinations are flagged, allowing human reviewers to concentrate on flagged spans; the lower precision means some false positives will also land in that queue.","Because the detector is small and fast, it can operate in real time, before an answer is shown to a user, rather than as a post hoc audit.","The data recipe, perturb multi-hop QA with in-context distractors and add reasoning labels, can be transferred to other base models and domains, provided a reliable generator for the perturbations exists."],"supporting_citations":[{"why":"Provides the RAGTruth benchmark and its manual word-level hallucination annotations, the evaluation set for every reported recall, precision, and F1 number.","marker":"(Niu et al., 2023)"},{"why":"Supplies the MuSiQue multi-hop QA dataset with unanswerable contrast questions that the perturbation pipeline starts from.","marker":"(Trivedi et al., 2021)"},{"why":"The reasoning-centric hallucination evaluation approach that the dataset construction is explicitly inspired by.","marker":"(Ravi et al., 2024)"},{"why":"Documents the Qwen 2.5 base models that are fine-tuned into Osiris.","marker":"(Yang et al., 2024)"},{"why":"Supplies the GPT-4o inference-speed figure used in the speed comparison.","marker":"(Artificial Analysis, 2024)"}],"fun_headline_variants":["7B open model surpasses GPT-4o in RAG hallucination recall","Osiris-7B beats GPT-4o on RAGTruth recall at 93.8%","Multi-hop training makes 7B detector beat GPT-4o on recall","Open-source 7B model reaches 93.8% recall, topping GPT-4o","7B open detector outperforms GPT-4o on hallucination recall"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the evaluation treats both systems alike: Osiris outputs are repaired with a GPT-4o-mini JSON prompt before scoring, and the paper never says the GPT-4o baseline gets the same repair, so the recall gap may partly reflect asymmetric post-processing.","fun_headline_variants_meta":{"raw":{"variants":["7B open model surpasses GPT-4o in RAG hallucination recall","Osiris-7B beats GPT-4o on RAGTruth recall at 93.8%","Multi-hop training makes 7B detector beat GPT-4o on recall","Open-source 7B model reaches 93.8% recall, topping GPT-4o","7B open detector outperforms GPT-4o on hallucination recall"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000747,"raw_usage":{"total_tokens":3322,"prompt_tokens":932,"completion_tokens":2390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2279}},"tokens_in":548,"tokens_out":2390,"duration_ms":16317,"temperature":1.0,"reasoning_tokens":2279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:19:39.423931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same RAGTruth evaluation twice: once with the JSON-repair post-processing applied to both Osiris-7B and GPT-4o, and once with it applied to neither. If Osiris's recall advantage disappears or reverses under symmetric scoring, the claimed edge is an artifact of asymmetric post-processing.","supporting_citations":[],"review_version":1}