Pith. sign in

REVIEW 4 major objections 5 minor 34 references

Truth or Mirage? Towards End-to-End Factuality Evaluation with LLM-Oasis

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2411.19655 v3 pith:66CGZS5V submitted 2024-11-29 cs.CL

Alessandro Scirè , Andrei Stefan Bejgu , Simone Tedeschi , Karim Ghonim , Federico Martelli , Roberto Navigli This is my paper · ORCID
classification cs.CL
keywords factualityevaluationhallucinationdetectionend-to-endclaimextractionnaturallanguageinferencebenchmarklargemodelsWikipedia
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Sec. 4.2, Appendix E.4] 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.
  2. [Sec. 4.2, Table 3] 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.
  3. [Sec. 5.1-5.3, Sec. 4.2] 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.
  4. [Table 1, Sec. 3.3] 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.
minor comments (5)
  1. [Table 1] 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.
  2. [Abstract, Table 5] 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.
  3. [Sec. 7.2, Table 2] 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.
  4. [Appendix E.3] The annotation guidelines contain a duplicated bullet defining the '<v>, <x>' combination; the duplicate should be removed.
  5. [Appendix E.2] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark labels are compositionally validated and external LLM scores are independent of the resource.

full rationale

The derivation chain is not circular. LLM-OASIS constructs unfactual texts by extracting atomic claims from Wikipedia, falsifying one claim, and generating a paraphrase that incorporates the falsified claim (Sec. 3, Table 1). The gold benchmark (Sec. 4.2) retains only instances that passed human component checks (claim extraction 96.78%, falsification 98.55%, factual paraphrase 90.36%, unfactual paraphrasing 89.20%; Sec. 4.1, Table 3). The sentence "We then labeled all the resulting factual and unfactual texts with True, and False, respectively" defines labels by pipeline conditions, but the conditions include human-verified falsification semantics and text-claim correspondence (Appx E.2, E.4), so the False label is a compositional inference, not a tautology. The main difficulty claim (GPT-4o 60% end-to-end, Llama-3.3-70B 61.7%) is measured on external LLMs not trained on the resource, so it is independent empirical evidence. The fine-tuned pipeline's higher score (69.24%) is explicitly acknowledged as partly due to training on LLM-OASIS; that is distributional familiarity, not circular derivation. The absence of a direct human factuality read on the final text (Appx E.4 asks whether the text corresponds to the claims) is a measurement-validity caveat, but it does not make an equation or prediction identical to its input. The only same-author citation (Scirè et al., 2024) is used for the easinessF1 selection metric and as an NLI inspiration, not as a load-bearing authority. No circular step is exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central benchmark claims rest on three domain assumptions: Wikipedia as ground truth, small-sample human validation as representative, and single-claim falsification as a faithful model of hallucination. No numerical free parameters are fitted to derive a theory; the listed parameters are dataset construction choices. No invented entities are introduced.

free parameters (4)
  • Passage length K = 5 sentences
    Sliding-window size chosen for resource construction; sets text length and claim density.
  • Sliding window stride s = 1 sentence
    Stride used to segment Wikipedia pages; affects total passage count.
  • Retrieval top-k = 30
    Selected from validation Recall@k analysis to balance recall and efficiency.
  • Falsified claims per passage = 1
    Design choice to make the end-to-end task harder by embedding one subtle error.
assumptions (4)
  • domain assumption Popular English Wikipedia pages are an accurate, sufficient ground-truth source for factuality labels.
    Sec 3 selects the 80k most-visited English Wikipedia pages as the sole knowledge source; unfactual texts are defined as contradicting these passages, not checked against independent sources.
  • domain assumption Human validation of 1,750 sampled instances per pipeline step is representative of the full 81k resource.
    Sec 4.1 extrapolates per-step accuracies (89.20-98.55%) to the unannotated majority; only a small, filtered subset is used for the gold benchmark.
  • domain assumption A text is unfactual if and only if it contains the single falsified claim, and this can be detected by entailment against retrieved evidence.
    Sec 3.3 and Sec 4.2 define labels from one modified claim; the modular system in Sec 5 assumes claim-level NLI aggregation captures passage-level factuality.
  • ad hoc to paper GPT-4 paraphrases generated from the same claim sets preserve semantics except for the intended falsification.
    Prompt Step 4 (Table 1) requires the unfactual text to be as similar as possible to the factual text, an assumption that controls confounds but may create generation artifacts.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Truth or Mirage? Towards End-to-End Factuality Evaluation with LLM-Oasis." pith.science (2026). https://pith.science/paper/66CGZS5V

@misc{pith2026241119655,
  author       = {Pith},
  title        = {Pith review of: Truth or Mirage? Towards End-to-End Factuality Evaluation with LLM-Oasis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66CGZS5V}},
  note         = {Machine review of arXiv:2411.19655}
}
read the original abstract

After the introduction of Large Language Models (LLMs), there have been substantial improvements in the performance of Natural Language Generation (NLG) tasks, including Text Summarization and Machine Translation. However, LLMs still produce outputs containing hallucinations, that is, content not grounded in factual information. Therefore, developing methods to assess the factuality of LLMs has become urgent. Indeed, resources for factuality evaluation have recently emerged. Although challenging, these resources face one or more of the following limitations: (i) they are tailored to a specific task or domain; (ii) they are limited in size, thereby preventing the training of new factuality evaluators; (iii) they are designed for simpler verification tasks, such as claim verification. To address these issues, we introduce LLM-Oasis, to the best of our knowledge the largest resource for training end-to-end factuality evaluators. LLM-Oasis is constructed by extracting claims from Wikipedia, falsifying a subset of these claims, and generating pairs of factual and unfactual texts. We then rely on human annotators to both validate the quality of our dataset and to create a gold standard test set for benchmarking factuality evaluation systems. Our experiments demonstrate that LLM-Oasis presents a significant challenge for state-of-the-art LLMs, with GPT-4o achieving up to 60% accuracy in our proposed end-to-end factuality evaluation task, highlighting its potential to drive future research in the field.

Figures

Figures reproduced from arXiv: 2411.19655 by the authors.

Figure 1
Figure 1. Pipeline for the creation of LLM-OASIS. Given a passage from a Wikipedia page (original text on top), we task an LLM to: extract a list of atomic claims (1), falsify one of the extracted claims (2), and then, given the two sets of claims, produce a paraphrase of the original text (3), and an alternative version featuring the unfactual information (4). evaluation approaches. In this context, we intro￾duce LLM-OASIS, … view at source ↗
Figure 2
Figure 2. Recall@k performance of the E5base model at different values of k. All the prompts used in the various settings can be found in Appendix B. 7 Results 7.1 Task 1: End-to-End Factuality Evaluation In this section we present the results obtained in the end-to-end factuality evaluation task (cf. Section 4.2). First of all, we examine the performance of the evidence retrieval module, as this component supplies the extern… view at source ↗
Figure 3
Figure 3. Balanced Accuracy (%) of different models in the RAG setting with increasing context length. Each [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

  1. [1]

    Albert Einstein was a German-born theoretical physicist

    work page

  2. [2]

    Noam Shazeer and Mitchell Stern

    Fenice: Factuality evaluation of summariza- tion based on natural language inference and claim extraction. Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. CoRR, abs/1804.04235. Derek Tam, Anisha Mascarenhas, Shiyue Zhang, Sarah Kwan, Mohit Bansal, and Colin Raffel. 2022. Evalu- ating the factual consis...

    arXiv 2018

  3. [3]

    The theory of relativity is one of the two pil- lars of modern physics

    work page

  4. [4]

    Einstein’s work influenced the philosophy of science

    work page

  5. [5]

    He developed the theory of relativity

    work page

  6. [6]

    the world’s most famous equation

    The formula E = mc2 is dubbed “the world’s most famous equation”. Factual Text: Albert Einstein, originally from Germany, was a theoretical physicist who formulated the theory of relativity, a cornerstone of modern physics. His contributions significantly impacted the philosophy of science. The mass–energy equivalence equation E = mc2, which he is most fa...

    work page

  7. [7]

    Factual" or

    Just reply with "Factual" or "Not Factual", do not generate any additional text to the answer. Table 9: Prompt for factuality evaluation in RAG setting. The results of all models tested during the abla- tion study are summarized in Table 10, confirming the robustness and efficiency of the E5base model for claim retrieval, balancing performance with comput...

    work page 2000

  8. [8]

    He is best known for his mass–energy equiva- lence formula E = mc2

    work page

Show all 34 references
  1. [9]

    Factual" or

    In the latter setting, we prompted all the LLMs with the same pieces of evidence retrieved and used by our NLI module (cf. Sec. 7.1). Concerning the Explain-then-Answer paradigm, we expand the set of instructions with the following recommendation: Motivate your response with a...

    2022

  2. [10]

    The Amazon Rainforest is also known as Ama- zonia

  3. [11]

    It is a moist broadleaf forest in the Amazon biome

  4. [12]

    The Amazon Rainforest covers most of the Amazon basin of South America

  5. [13]

    The region includes territory belonging to nine nations

  6. [14]

    The majority of the forest is contained within Brazil

  7. [15]

    Brazil contains 60% of the rainforest. Factual Text: Amazonia, widely known as the Amazon Rainfor- est, is a damp broadleaf forest located within the Amazon biome, covering a significant portion of the Amazon basin in South America. This vast region spans across nine countries...

  8. [17]

    Evaluate the factual accuracy of the input text based on your training data and knowledge

  9. [18]

    supported by known information, respond with "Factual"

    If the input text is factually-accurate, i.e. supported by known information, respond with "Factual"

  10. [20]

    Factual" or

    Just reply with "Factual" or "Not Factual", do not generate any additional text to the answer. Table 8: Zero-Shot Prompt for factuality evaluation of a text. Determine whether the given text is factual or not using the provided evidence. If the information is not present in th...

  11. [21]

    Read the evidence if provided

  12. [22]

    Assess whether the input text is factual based on the evidence if present

  13. [23]

    If the evidence are not provided or is insufficient, use your prior knowledge to determine the factuality

  14. [24]

    Not Factual

    Respond with "Not Factual" if the input text contains even a single inaccuracy

  15. [25]

    If the evidence is not related to the text to verify, rely on your prior knowledge to provide the answer

  16. [27]

    Read the original text and claims thoroughly

  17. [28]

    For each claim, determine if it is accurately represented in the original text

  18. [29]

    v" in the third column if the claim is present in the original text, otherwise mark it with an

    Place a "v" in the third column if the claim is present in the original text, otherwise mark it with an "x" Annotation Example We report an example of annotated instance in Table 11. Additional Guidelines Annotators are required to discard an entire instance, composed of the o...

  19. [30]

    Compare the two claims provided

  20. [31]

    Determine if the unfactual claim introduces new, untrue information compared to the orig- inal claim

  21. [32]

    v" if unfactual infor- mation is introduced, otherwise mark it with

    Mark column 3 with "v" if unfactual infor- mation is introduced, otherwise mark it with "x". Annotation Example We report an example of annotated instance in Table 13. Additional Guidelines If the original claim con- tains a word that is replaced with its hyponym in the candid...

  22. [33]

    Compare each claim with its representation in the paraphrased text

  23. [34]

    v" in the third column. • Use

    Determine if its semantics is preserved. • If it is preserved (regardless of whether it is reported identically in the paraphrase), place a "v" in the third column. • Use "x" otherwise

  24. [35]

    v". • If not paraphrased (e.g. identical), mark column 4 with

    Determine if it is paraphrased. • If a claim is paraphrased, mark the fourth column with "v". • If not paraphrased (e.g. identical), mark column 4 with "x". In other words: • <"v", "v"> in the last two columns means that the semantics is preserved and the text is para- phrased...

    2018

  25. [2023]

    the world’s most famous equation

    Alignscore: Evaluating factual consistency with a unified alignment function. Biao Zhang, Barry Haddow, and Alexandra Birch. 2023a. Prompting large language model for machine translation: A case study. Shiyue Zhang and Mohit Bansal. 2021. Finding a bal- anced degree of automat...

    2021

  26. [2024]

    In Proceedings of the 18th Conference of the European Chapter of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 49–66, St

    Generating benchmarks for factuality evalua- tion of language models. In Proceedings of the 18th Conference of the European Chapter of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 49–66, St. Julian’s, Malta. Associa- tion for Computational Lin...

    2024 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.