acceptodds
Withdrawn from ICLR 2027

Conformal Factuality with Learned Structure

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As large language models (LLMs) are increasingly used to generate content, evaluating the correctness of their outputs has become essential. We introduce a novel framework for filtering factually incorrect claims from a given piece of AI-generated text while controlling an expected loss, such as the false negative rate (FNR). Our work builds on the recent conformal factuality literature, which leverages conformal risk control (CRC) to select a scalar filtering threshold with distribution-free validity guarantees under exchangeability. However, unlike methods that assess the correctness of each claim in isolation, our approach constructs factuality scores by leveraging the natural language inference (NLI) relationships between all pairs of claims within the text. Specifically, given a piece of text and a raw factuality score function, our algorithm constructs a graphical model over claim correctness, with pairwise factors defined by the estimated NLI scores. Our algorithm then uses belief propagation to approximate the marginal probability that each claim is incorrect, and applies CRC to the resulting updated scores to select a threshold for filtering claims. We demonstrate the effectiveness of our approach on three benchmarks: FABLES, FELM, and a newly annotated Wikipedia-based dataset. The results show that leveraging learned structure between claims substantially improves claim retention over baseline methods while controlling the expected FNR at the nominal level. Remarkably, even when instantiated with open-source models, our method achieves factuality classification accuracy comparable to that of the closed-source baseline.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal Inference, Variational Inference, Multiple Hypothesis Testing, LLM Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CCpSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method for detecting incorrect claims in LLM-generated text while controlling the false negative rate (FNR). The main idea is to improve existing factuality scores by using NLI relationships between claims, such as entailment, contradiction, and neutrality. The authors first define a posterior score based on an independent Bernoulli working model and then use star approximation and belief propagation to reduce computational cost. They apply conformal risk control (CRC) to provide finite-sample marginal FNR guarantees. Experiments on four datasets show that the proposed method can retain more correct claims while improving detection accuracy compared to existing methods.

Critical strengths

S1. The paper addresses an important and practical problem.

Detecting incorrect claims while retaining as much correct information as possible is an important problem. Since FNR can be trivially controlled by removing all claims, I think it is natural to consider retention alongside FNR control.

S2. The idea of NLI-based score refinement is intuitive.

The idea of updating factuality scores using logical relationships between claims, rather than evaluating each claim independently, is interesting. Formulating this idea through a probabilistic model and considering multiple approximation methods are also meaningful aspects of the work.

Critical weaknesses

W1. The theoretical explanation connecting NLI refinement to retention improvement is insufficient.

  • While the intuition behind updating factuality scores using NLI relationships is understandable, it is unclear whether these updates actually produce better scores and lead to improved retention.
  • The FNR guarantee in Theorem 1 holds regardless of the quality of the scoring function under appropriate conditions. Therefore, I believe the main contribution of this paper lies in improving score quality and retention through NLI information. However, Theorem 2 only provides an upper bound for the star posterior and does not directly explain these improvements.
  • Of course, it would be difficult to expect a guarantee that NLI always improves performance. Nevertheless, analyzing the conditions under which improvements can be expected would help clarify the contribution. For example, Noh et al. (2026), although working in a different setting, analyze the relationship between score estimation error and the retention gap. I wonder whether a similar connection could be established between the accuracy of NLI information, the effect of refinement, and retention in this paper.

W2. The experimental contribution of the NLI component is not sufficiently demonstrated.

  • Although several scoring methods are compared, it is difficult to understand which NLI relationships actually contribute to performance improvements. For example, the retention of the Full posterior increases from 0.3500 to 0.4792 on the authors' constructed dataset (Table 6), whereas the improvement on WiCE is much smaller, from 0.2023 to 0.2093. There is insufficient analysis explaining why these differences occur.
  • It would be helpful to report the proportions of entailment, contradiction, and neutrality relationships, as well as NLI prediction accuracy, for each dataset. I also think ablation studies comparing performance when NLI relationships are removed or made inaccurate are needed.

W3. The experimental setting and sensitivity analysis are insufficient.

  • Section 3.1 defines a training/calibration split, but Appendix A.6 only reports a Train/Test split. Since pretrained VERA and DeBERTa-v3 models are used, it is unclear how the training data are specifically used and how calibration is performed.
  • Another limitation is that all main experiments are conducted at α=0.25\alpha=0.25. Comparing the FNR–retention trade-off across different values of α\alpha would help clarify the method's practical utility.
  • In addition, since the method uses NLI relationships between all pairs of claims and the computational complexity of the full posterior is O(K22K)O(K^2 2^K), I think an analysis of performance and computational cost as a function of the number of claims KK is necessary.
Questions for authors

Q1. Can the authors theoretically characterize the conditions under which NLI-based posterior updates improve score quality or retention? In particular, how does the accuracy of NLI relationships affect the final performance?

Q2. What are the distributions and prediction accuracies of NLI relationships across the datasets? Is it possible to analyze performance according to relationship type or NLI prediction error?

Q3. Could the authors compare the FNR–retention trade-off across different values of α\alpha (e.g., α∈0.01,0.05,…\alpha\in{0.01,0.05,\ldots})? In particular, I wonder whether the improvements persist at smaller values of α\alpha.

Q4. How do posterior approximation accuracy, retention, and computation time change with KK? What values of KK and padding/truncation settings were used in practice?

Q5. How exactly are the training, calibration, and test splits constructed? If pretrained scorers are used, what is the role of the training stage in Algorithm 1?

Q6. Appendix A.6 mentions modifications to BTProp. How do these changes affect performance compared with the original method?

Suggestions:

  • It would be helpful to organize the experimental comparisons according to their objectives. Existing methods can be categorized into conformal factuality (Mohri & Hashimoto, 2024; Cherian et al., 2024), NLI-based refinement (Hou et al., 2025; Marinescu et al., 2025; Rubin-Toles et al., 2025), and standalone factuality scoring. Evaluating scoring quality using AUROC/AUPRC and evaluating FNR–retention after applying CRC would help clarify the contribution. However, comparisons should account for differences in the guarantees provided and the information used by each method.
  • NLI ablation and scalability analyses are needed. Comparisons with neutral-only, corrupted NLI, and oracle NLI settings, along with runtime analysis as a function of KK, would be useful.
  • The notation and practical details could be clarified. It would be helpful to provide more explanation of the transition from xix_i to ziz_i in Section 3.1, the effects of claim ordering and truncation, and how flagged claims are used for filtering the actual text.

References

[1] Mohri, C. and Hashimoto, T. (2024). Language Models with Conformal Factuality Guarantees. ICML.

[2] Cherian, J. J., Gibbs, I., and Candès, E. J. (2024). Large Language Model Validity via Enhanced Conformal Prediction Methods. NeurIPS.

[3] Noh, K., Lee, S., Kim, I., and Song, K. (2026). Multi-LLM Adaptive Conformal Inference for Reliable LLM Response. ICLR.

[4] Hou, B., Zhang, Y., Andreas, J., and Chang, S. (2025). A Probabilistic Framework for LLM Hallucination Detection via Belief Tree Propagation. NAACL.

[5] Marinescu, R. et al. (2025). FactReasoner: A Probabilistic Approach to Long-Form Factuality Assessment for Large Language Models. Findings of EMNLP.

[6] Rubin-Toles, M. et al. (2025). Conformal Language Model Reasoning with Coherent Factuality. ICLR 2025.

Reviewer_3UQXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • ConFacTR is a structure-aware framework for detecting incorrect factual claims in generated text while controlling the false-negative proportion through conformal risk control.
  • The method exploits pairwise natural language inference relationships to refine initial claim-level incorrectness scores.
  • The authors formulate these relationships using a probabilistic model and investigate posterior inference via approximation to estimate the factuality scores.
  • The authors then calibrated the scores using some labelled examples to select a threshold with an expected FNR guarantee
  • Ran on a couple of evaluations (FELM, FABLES, WiCE) and a newly constructed dataset to report improvements in claim retention and factuality-screening perf relative to selected baselines.
Critical strengths
  1. Paper is well written and easy to follow. It is a meaningful problem although I'm not so sure of the relevance of the problem today. The problem of managing errors under risk control is important, and I appreciate well formulated derivations for the screening problem. The key is to retain as much useful information/content as possible while keeping the FNR low. This is good for providing a useful operating objective for some form of factuality assurance.

  2. Statistical formulation is good. The central idea of exploiting entailment and contradiction relationships between claims is intuitive and well motivated. Probabilistic inference with calibration while preserving a formal risk guarantee forms a good baseline to build upon. The substantive math formulation on the scoring procedure, exact posterior construction and approximate inference alternatives are good usage of well-ironed theories to the factuality problem (although I don't think this is entirely novel).

Critical weaknesses
  1. My main gripe here as you would have already have expected is that there is really limited methodological novelty beyond the established components that have been formulated here. No doubt that the proposed integration is coherent, but beyond that, I find it difficult to appreciate the breadth of the solution beyond just tackling the factuality issue in fact-based textual content. Conformal risk control and factuality guarantees have been well explored by Mohri-Hashimoto (ICML 2024), and NLI-based probabilistic reasoning has been around for forever. (Hou et al. (NAACL 2025), Rubin-Toles et al. (2025), and Marinescu et al. (EMNLP 2025)) as rightly pointed out in some of the references. How would the contribution of the paper generalise beyond the problems in factuality scoring under conformal risk settings to other real-world use-cases? How does this hold up against some of the more challenging fake-news/misinformation datasets that we can find today?

  2. There are the strong assumptions on calibration data and limited applicability of factuality guarantees. Will such representative calibration dataset with ground-truth correctness labels for individual claims be always readily available? In realistic factuality assessment settings, there are heterogeneous, specialized, or evolving document corpora. To obtain such annotations is expensive, requires expertise, and is difficult to scale. Additionally, conformal FNR guarantee depends on the exchangeability between calibration and deployment samples, how would this hold across different domains of operations? How will this change with different LLMs, different factual error distributions etc. and how much can we rely on synthetic generators to reduce annotation costs? There needs to be more thought put in to the practical generalizability of the proposed framework.

  3. Evaluations - The experimental datasets are relatively small, and the newly constructed dataset deliberately introduces logical dependencies that favor the proposed structure-aware formulation. Can the authors better characterize the calibration sample size, annotation errors, or distribution shift ablations that affect downstream perf, retention rates, risk thresholds etc.?

Questions for authors

Actually, I kinda conflate my questions into the critical weaknesses above. My apologies for that. But yes, you may refer to the above and further develop ways to strengthen the paper. I think for a start, to work on expanding the evaluation test set, run more ablations to probe the framework i.e. understanding whether the logical structure genuinely contributes information about factuality? Running the experiment under miscalibrated scores? Change the calibration size, run with various document corpora length/density etc.

Will be interesting to understand these insights. :)

Reviewer_bomHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Authors propose an extension to another published paper [1], by incorporating the degree in which the claims in a document can contradict each other. The method proposed in [1], views the claims inside documents individually and ignores their relationships, but in this paper authors argue that this relationship can be used to improve the detection of false claims. The mentioned relationships are obtained by an NLI method, which are then stored in a matrix. The relationship matrix and the initial correctness vector of the claims--obtained via an llm--are used in approximation and full algorithms to update the correctness vectors. Finally the claims are inspected by learning a threshold in a held out set.

The method is evaluated in four datasets--three existing ones and a new synthetic one. The results show that the new method when incorporated into an existing baseline outperforms GPT-5-nano.

[1] Christopher Mohri and Tatsunori Hashimoto. Language models with conformal factuality guarantees. 2024.

Critical strengths
  • In my opinion authors address a very interesting problem in which the algorithm can rely only on parametric knowledge--regardless of how realistic the problem is
  • The improvements are consistent
  • The code is released
Critical weaknesses
  • The datasets are small. As a person who has worked in this area, I can also add that the proposed new dataset is not realistic
  • A large section of the method discussed in the paper already existed, the new extension adds only a small part. Even worse, there is no theoretical justification for it, it just works
Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.