We sincerely thank the reviewers for the constructive feedback.

Weakness 1 & Question 1

Some viewpoints in the paper lack factual support. For instance, in line 214, the statement "prompting the model to provide reflections on the evaluation" lacks evidence or justification.

Why is the "prompting the model to provide reflections" approach a more effective method? This actually corresponds precisely to weaknesses 1.

We acknowledge that it might not be the optimal or most effective. The question seems to arise from a misunderstanding. In the manuscript, we believe that “..is a more convenient method”, since it directly follows the thoughts of fact-checking and does not require extra steps to elicit the knowledge. Therefore, it is technically more simple.

Moreover, in the experiments, we compare baselines with various elicitation methods, including self-ask (CoVE), self-consistency (SelfCheckGPT and ChatProtect), and chain-of-thought (CoT and FaR). Compared with these methods, SelfElicit utilized reflections conditioned on the evaluation, which shows better factuality and diversity (Section 4.3 Elicitation Quality).

Weakness 2

Although mentioned in the Limitations section, it is worth emphasizing that the experiments are conducted solely on real-world medical QA datasets. This limitation narrows the scope of general conclusions and reduces the credibility of SelfElicit’s generalizability across diverse fields.

We have conducted experiments on the WikiBio $1$ dataset, which includes the biographies of 238 celebrities. In summary, the result shows that SelfElicit has strong performance across different domains. Detailed metrics and discussions can be found in our response to Weaknesses 8 & 11, Reviewer LceZ (link: https://openreview.net/forum?id=r9mYbs8RTH&noteId=ttwwA4UQJX).

Weakness 3 & Question 5

In the Main Results section, the baseline is locally implemented rather than based on data from previous research, which diminishes the persuasive strength of the experimental findings.

In the Main Results section, the baseline was implemented locally rather than using data from other studies, which may weaken the persuasive power of the experimental results. Could you provide comparative data from similar scenarios in other literature to strengthen credibility? (For example, using results from SelfCheckGPT under comparable experimental conditions).

We sincerely thank you for the suggestions. However, due to various realistic limitations, we have to implement these methods locally, which include: (1) using private datasets (CoVE), and (2) focusing on short-form datasets (IO, ChatProtect, and FaR).

More importantly, the main focus of the experiment is to compare the effectiveness of various elicitation methods. Therefore, we use the identical self-eval prompt after their original procedures to obtain the hallucination scores for a fair comparison. In other words, the only difference between all methods is the approach to elicit the intrinsic knowledge. Unfortunately, we have to implement these baselines locally rather than using metrics reported in their papers since they have used different methods to obtain the hallucination score.

We sincerely thank the reviewers for the constructive feedback.

Weakness 1 & Question 1

Weaknesses 1: The paper's focus on medical QA datasets raises concerns about the framework's ability to generalize to other domains. The lack of experimentation in diverse domains means that the robustness and applicability of SelfElicit across different types of long-form content remain unproven.

Question 1: How well does the SelfElicit framework generalize to different domains beyond medical QA? Have there been any preliminary tests or simulations to assess its performance on long-form content from other fields, such as legal or historical texts?

We have conducted experiments on the WikiBio $1$ dataset, which includes the biographies of 238 celebrities. In summary, the result shows that SelfElicit has strong performance across different domains. Detailed metrics and discussions can be found in our response to Weaknesses 8 & 11, Response to Reviewer LceZ (part 3), link: https://openreview.net/forum?id=r9mYbs8RTH&noteId=ttwwA4UQJX).

To the best of our knowledge, there are currently no publicly available datasets from other domains. And we will leave the work of synthesizing multi-domain data to future works due to time limits.

Weakness 2 & Question 2

The theoretical underpinnings of the self-elicitation process and its interaction with the knowledge hypergraph are not thoroughly explored. The paper does not provide a deep dive into the mathematical or conceptual models that would support the claims about the framework's effectiveness.

What is the theoretical foundation that underpins the self-elicitation mechanism and its interaction with the knowledge hypergraph? Could the authors provide more details on any formal proofs or mathematical models that validate the effectiveness of this approach?

self-elicitation

We acknowledge that the motivation of self-elication is largely experimental and we have not yet fully understand the mathematical model. However, we have some conceptual understandings about the effectiveness.

Conceptually, SelfElicit has schematic similarities with multi-step reasoning. Since LLMs face difficulties in using their stored knowledge to perform multi-step inference tasks $4$ , our SelfElicit alleviates these issues in two aspects. On one hand, the model only focuses on one specific factoid at a time, which makes it easier to utilize their intrinsic knowledge. More importantly, since parts of the intrinsic knowledge have been expressed verbally in advance, the reasoning burden can be reduced with the help of the in-context information. Similar gains can also be found in a boarder context $4$

$5$

$6$

$7$ . In this work, the capacity of multi-step reasoning is integrated with the calibration capacity of LLMs for effective long-form hallucination detection.

knowledge hypergraph

The interaction with the knowledge hypergraph is for the technical convenience of knowledge organization. Compared with HistoryIO (prefixing historical query-responses as contextual information), SelfElicit further uses graph structures to enable a more careful organization of the elicited information via sampling, conflict detection, and mitigation, leading to better performance. Moreover, graphs provide more intuitive knowledge organization, better efficiency, and higher underlying expansibility than text-based storage.

Weakness 3

The paper heavily relies on quantitative metrics to evaluate the performance of the SelfElicit framework. However, there is an absence of qualitative analysis, such as case studies or detailed error analysis, which could provide a more nuanced understanding of the framework's strengths and weaknesses in practical scenarios.

We have added two case studies in Section 4.5 in the revised version to demonstrate the workflow of our method. One is an evaluation with/without contextual information, showing the effectiveness of self-elicited thoughts, and the other is a model generating inaccurate reflection and how it is mitigated by conflict detection.

Weakness 4

The visual representation in the paper, particularly in Figure 2, is not sufficiently clear or expressive, which can lead to confusion about the SelfElicit framework's methodology. The diagrams fail to effectively communicate the process, detracting from the reader's understanding and the overall quality of the presentation. A redesign of these figures is recommended for better clarity.

We have redesigned Figure 2 to make it clearer, including a reorganized diagram, clearer arrows, and example texts. We have also included a case study with a detailed data case in Section 4.5 in the revised version.

Thank you for your reply and addressed some concerns, so I will improve my scores.

Question 2

What are the computational costs associated with maintaining and updating the knowledge hypergraph, especially for large datasets? Could the authors provide specific metrics or comparisons, such as how these costs scale with the size of the input or the number of iterations?

Please refer to Question 3, reviewer GtBE (link: https://openreview.net/forum?id=r9mYbs8RTH&noteId=fFjI0DmHZz). The conclusion is that the complexity is $\mathcal{O}(F^2)$ where $F$ refers to the number of factoids in a sample (the number of iterations). However, since the conflict detection/mitigation procedures are not always activated and the scale of current datasets is not large enough, it is not the major issue for this task in real-world scenarios.

Question 4

Can the approach be adapted or extended to work with external knowledge sources to further enhance hallucination detection?

Yes. Conceptually, by incorporating external information into the knowledge hypergraph, our method can be enhanced without modifying the framework, which is impossible for the baselines.

Question 5

How does the framework perform across different domains and languages, and what adaptations might be necessary for domain-specific applications?

domain

We have conducted experiments on the WikiBio $3$ dataset, which includes the biographies of 238 celebrities. The result shows that SelfElicit has strong performance across different domains. Detailed discussions can be found in our response to Weaknesses 8 & 11, Reviewer LceZ (link: https://openreview.net/forum?id=r9mYbs8RTH&noteId=ttwwA4UQJX)

language

Table 4: qwen1.5-7B-chat results across languages. S: sentence-level metrics. R: response-level metrics. Bold: better than the other language.

It can be observed that, although these datasets have identical knowledge points and hallucinations, the performance of MedHallu-en is relatively inferior compared with MedHallu-zh in 24 over 36 cases. Considering the overall performance is dependent on the capacity of the models, we owe that the model is unfamiliar with the English version of some medical terminologies (e.g. Chinese medicine), which has been proven by our manual tests.

adaption

The only adaption required is the prompt of extracting entities/statements (Section 3.1), which is an Open Information Extraction (OpenIE) procedure. We observe that incorporating domain-specific expertise improves the quality of the extraction. All other prompts used in our framework are domain-independent and no adaptation is required.

References

[1]StructBERT. https://modelscope.cn/models/iic/nlp_structbert_nli_chinese-large.

[2]SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models. 2023.

[3]WikiBio. https://huggingface.co/datasets/potsawee/wiki_bio_gpt3_hallucination

Weakness 1

Complexity: The approach involves multiple steps and components, which may increase the complexity of implementation and computational overhead. To better understand the scalability of the framework, could the authors provide a detailed analysis of the computational complexity? Additionally, it would be helpful to see comparisons with existing methods in terms of implementation complexity and runtime. Not so sure about scalability here.

Since the evaluation of each sample is independent, the discussion is under the condition when fact-checking a single sample.

Assume there are $N$ statements (or $F$ factoids) to be fact-checked in a given sample. We categorize the LLM usages into two categories: token where only the first token or its logit matters, and open-ended where LLMs generate full reasoning given an instruction. Note that an open-ended generation will be magnitudes more costly than a token generation because the former usually generates hundreds of tokens while the latter generates only several tokens.

Table 1. Complexity analysis of all methods. $N$: the number of sentences (average 6.81). $F$: the number of factoids (average 7.37). $K$: the number of generated documents. Brackets denote the components that require LLM usage.

Since an open-ended generation is magnitudes more costly than a token generation, the major overhead is the open-ended generation. In summary, SelfElicit (1+F) has a comparable complexity with CoT (N) and is faster than CoVE (N+FN), FaR (N+N), and ChatProtect (N+FN), which is also demonstrated by the experimental results in Table 2 in Section 5.1.

Weakness 2.1

Dependency on LLMs: The performance of the framework is inherently tied to the capabilities of the underlying LLMs, which may limit its effectiveness in certain scenarios. Could the authors provide a more detailed analysis of how the framework's performance varies with different LLMs?

The performance is inherently tied to the capabilities of the underlying LLMs. In Table 2, we compare two families of modern LLMs: the qwen family (qwen1.5 against qwen2.5) and the Llama family (llama2 against llama3.1) with close model scales.

(1) Llama3.1 and Qwen2.5 outperform their previous-generation counterparts in almost all cases, respectively, which is owe to their stronger capability.

(2) In some cases, with appropriate algorithms, the performance of previous-generation models surpasses latest-generation models (e.g. SelfElicit+Llama2 > IO+Llama3). This phenomenon shows that appropriate algorithms can benefit fact-checking and boost performance without upgrading the LLMs.

(3) SelfElicit beats all other baselines across different LLM backbones in most cases, which demonstrates the effectiveness of our method.

Weakness 7

The evaluated LLMs are relatively limited and small-scale. It is suggested that the authors also evaluate on SOTA models such as GPT-4 and models with a larger scale such as llama-70B. Otherwise, the effectiveness of the proposed method is relatively limited.

Model Scale

We have shown the scalability results using up to 110B model (Qwen1.5-110B) in Figure 5, Appendix A6 in the initial submission (or Figure 7, Appendix D1 in the revised version). The scalability results have demonstrated that (i) larger models tend to have better performance, (ii) the inference costs also increase nearly linearly with the model size, and (iii) our SelfElict consistently outperforms the baselines.

Comparison with other SOTAs

We have also conducted further experiments on more LLMs, including SOTA open-sourced models Qwen2.5, Llama3.1, and commercial model GPT4o-mini. The results are shown in Table 1.

The results show that our SelfElict outperforms all baselines in most cases across the latest open-source LLMs and commercial LLMs, which demonstrates the versatility of our proposed approach. Moreover, we can also observe that the latest-generation models (Qwen2.5 and Llama3.1) outperform their previous-generation counterparts (Qwen1.5 and Llama2). The full results are shown in Table 1 in our revised version.

We also notice that GPT4o-mini, one of the most efficient and powerful commercial models, has relatively inferior performance. This reason is that we are not able to obtain the model’s confidence score nor adaptively adjust the thresholds to reduce confidence biases. Specifically, when it comes to commercial models accessed by APIs, we are not able to obtain detailed token logits but only binary results through verbal outputs. Consequently, it is unable to adaptively modify the thresholds according to the results of the validation set. The confidence biases of all methods might eventually lead to sub-optimal performance $2$ . Moreover, since the training corpora of GPT are not public, we can not further assess whether it has specific knowledge in the medical domain.

Weaknesses 8 & 11

The authors only evaluate the performance of the proposed methods on their own dataset. Could the authors also test the effectiveness on public datasets?

The paper's evaluation is limited to the medical domain. The lack of evaluation across other domains makes it difficult to assess the general applicability of the approach.

We acknowledge that it is important to comprehensively evaluate the methods with public datasets in various domains.

We have added an experiment on a publicly available long-form hallucination dataset, WikiBio $1$ , which includes the biographies of 238 celebrities. Since most samples in this dataset are positive (hallucination), we only report the AUC metric here.

Table 2: Results on WikiBio. S: sentence-level metrics. R: response-level metrics. Bold: the best. Italic: the second best.

Although SelfElicit does not achieve the best performance in the starred (*) metric, the difference is less than 2%. Compared against baselines, our SelfElicit achieves leading performance (best or second best in 3/4 cases) and on average improvements of 11.1%, 12.1%, 5.6%, and 18.0% against all baselines. These results show that SelfElicit has strong performance across different domains.

Moreover, we also observe that the relative performance is not as obvious as on the MedHallu datasets. The reasoning might be that the biographies have weaker contextual relationships between sentences but features stating individual factoids, which can be proven by the inferior performance of context-argument methods (ContextIO and HistoryIO).

Weakness 6

"While insightful, we contend that these methods either require complex manual prompts or involve intricate reasoning processes, which limit their elicitation capacity and increase the risk of accumulated inaccuracies and hallucinations." (Line 67) needs more explanations

We analyze the limitations of existing methods as follows:

• ChatProtect requires multiple reasoning steps: generate triples, elicit intrinsic knowledge via cloze, and detect hallucinations by consistency. We observe that the process of generating triples introduces most inaccuracies because some sentences include conditioned or constrained knowledge (e.g. sentence “Gliclazide is used when diabetes cannot be controlled by proper dietary management…” ). Simple triples discussed in ChatProtect, i.e. (sub, pred, obj), are unable to handle such cases, or careful prompt designs are required to handle such complex cases. In summary, the multi-step fact-checking reasoning process increases extra manual labor for prompt design, the cost of LLM usage, and the risk of accumulated inaccuracies.
• SelfCheck also requires multiple reasoning steps: target extraction, information collection, step generation, and result comparison. Similarly, it has a higher cost of LLM usage and the risk of inaccuracy accumulation.
• EVER, CoVE, and QuestGen generate fact-validation questions for each factoid, elicit intrinsic knowledge via answering these questions, and compare the results. We observe that most uncertainties and inaccuracies occur when generating these questions. For example, given a factoid “Angina pectoris can be relieved by sublingual nitroglycerin tablets in emergency situations.” The model sometimes generates opaque, inarticulate articulate questions “In what situations, can Angina pectoris be relieved by sublingual nitroglycerin tablets?” or open-ended answers “What medications are commonly used for relieving angina pectoris?” that might have several answers.
• SelfCheckGPT (-prompt) and FaR ask the model to generate documents related to the factoid and check whether the factoid is supported by these documents. We observe that the generated documents tend to be lengthy and have much irrelevant information, increasing the cost and introducing noises to the evaluation reasoning.

Compared with the above methods, our SelfElicit uses a simple elicitation prompt and directly expresses relevant knowledge conditioned on the evaluation result, which reduces the overhead of LLM usage and the risk of inaccuracy accumulation. Experimentally, these baselines use 76%-1244% more token generations (Table 2) while having inferior performance (Table 1). We will elaborate on their limitations more clearly in the revised manuscript.

Table 1: Qwen2.5, Llama3.1, and GPT4o-mini results on MedHallu-EN. S: sentence-level metrics. R: response-level metrics. Bold: the best. Italic: the second best.

We sincerely thank the reviewers for the constructive feedback.

Weakness 1

There is no related work section. It is strongly suggested that the authors conduct a comprehensive literature review, especially on hallucination detection. Otherwise, this paper is actually not complete.

We apologize for the confusion. The related work section is included in Appendix A in our initial submission.

Weakness 2

"For the convenience in writing, we slightly abuse the term “non-factual” to represent both factuality and faithfulness hallucinations (Huang et al., 2023). Yet our work is not limited to either one" is confusing. What type of hallucination does this paper focus on? It is suggested that the authors explicitly state the type of hallucinations rather than mix them.

This work mainly focuses on factuality (external) hallucinations and investigates how to use intrinsic knowledge to detect non-factualities. Moreover, we also observe that in several cases, our method is capable of detecting faithfulness (internal) hallucinations. We owe it to the inconsistency detection component between the former and the latter statements, which can detect internal hallucinations to some extent.

We have differentiated the terminologies in line 53 in our manuscript to avoid confusion.

Weakness 3

"they are unable to learn the inherent semantic continuity in long-form content." (Line 19) is unclear. It is suggested that the authors define "semantic continuity" before they use it, even in the abstract.

We have updated the abstract by replacing “semantic continuity” with “in-context semantics”, which is more understandable.

Weakness 4

"However, there remains a concern regarding their tendency to generate hallucinations (Bang et al., 2023), producing sentences with plausible looking yet factually1 unsupported content (Huang et al., 2023) and hurting their faithfulness in real-world scenarios expecting factually-accurate response (Wei et al., 2024)." How do the authors define "faithfulness"? It is suggested that the authors differentiate "faithfulness" and "factuality". Otherwise, it is confusing.

We apologize for the confusion. The meaning here is that such non-factuality will weaken the reliability of LLMs in scenarios such as medicine. We replace the word to reliability for clearness. As stated in the above Weakness 2, this work mainly focuses on factuality (external) hallucinations and we have differentiated these terminologies in the revised version.

Weakness 5

The example “Gliclazide can be taken at any time of the day, regardless of whether it is on an empty stomach or after meals” needs more explanations why it is misleading.

“Swallow your gliclazide tablets whole with a drink of water. Do not chew them. If you are taking 2 doses a day, take 1 dose with your breakfast and 1 dose with your evening meal. If you are taking slow-release gliclazide, take your dose once a day with breakfast.” $5$ In summary, gliclazide should be taken with the meal, rather than “at any time”. Therefore, this statement is factually incorrect. We have included a brief explanation in line 39.