FacTool: Factuality Detection in Generative AI -- A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios

We appreciate the reviewer for the thoughtful review. The following is the response to major concerns:

Limited LLM evaluation scope

We thank the reviewer for the thoughtful suggestion. While the initial evaluation focused on three prominent models (GPT-4, ChatGPT, and Flan-T5-XXL), this represents a strategic choice to validate the framework on diverse architectures - covering both OpenAI's GPT family and Google's model family. The framework is designed to be model-agnostic, and the selected models span different scales and training paradigms, providing a reasonable foundation for demonstrating generalizability. Future work can easily extend evaluation to additional models using the established framework. We will update the paper with results of newer model (including DeepSeek-R1 and Qwen) in our next version. Our current experiments show that, even without powered by state-of-the-art models, FacTool is still a very effective hallucination detection approach.

Potential bias from using ChatGPT in FacTool components

This concern about positive bias toward ChatGPT is valid and important. Currently, our empirical results may provide some evidences to mitigate this concern. Specifically, if bias were the primary factor, we would expect consistent advantages for GPT4/ChatGPT across all tasks (KBQA, math, code, scientific). Instead, we see varied performance patterns that suggest FacTool is detecting genuine differences rather than evaluation artifacts. Notably, Bard actually outperforms GPT-4 on mathematical reasoning tasks, which demonstrates that the evaluation framework does not inherently favor OpenAI models. This cross-model performance variation indicates the framework is capturing authentic differences in factual accuracy. Additionally, GPT-4's superior performance on most tasks aligns with common knowledge about its enhanced reasoning capabilities compared to earlier models. To further address this concern, future iterations could apply model-agnostic implementation using open-source alternatives or ensemble approaches, though the current results suggest bias may not be a primary confounding factor.

Pipeline error propagation concerns

Error propagation is indeed a consideration in any multi-stage pipeline. Our framework includes several mitigation strategies: (1) multiple evidence sources are collected to reduce single-point failures, (2) the agreement verification step cross-validates claims against multiple pieces of evidence, and (3) each component is designed with robustness in mind. While some error propagation is inevitable, our empirical results suggest the overall system maintains reliability across diverse tasks and domains. We discuss potential failure cases in the appendix of the paper, including potential errors in llm reasoning and tool errors.

Missing qualitative discussion in the main part of the paper

We provide detail analysis in the appendix of the paper due to page limitations. We thank the reviewer's suggestion, and will do our best to move some of the qualitative discussion from the appendix to the main part of the paper.

We respectfully appreciate the reviewer's feedback. The following is the response to major concerns:

Limited novelty in individual components

While individual components like claim extraction and evidence retrieval exist separately, FacTool's key novelty lies in creating a unified, multi-task framework that works across diverse domains (KB-QA, code, math, scientific literature). The survey paper [3] shows that prior work focuses on single tasks - for example, most of the 69 surveyed papers address only one specific fact-checking task, with fact verification and fake news detection being the most prominent (53 out of 69 papers), but typically confined to single domains. FacTool's contribution is demonstrating how tool augmentation can be systematically applied across multiple domains with domain-specific adaptations.

Bias in tool outputs not discussed

We acknowledge that search engines may have ranking biases. We discuss the potential errors of "conflicting evidence" in our failure analysis in Appendix B.2.

Computational and resource constraints

While tool augmentation does increase computational overhead, this trade-off is often worthwhile for high-stakes applications where factuality is crucial. The paper demonstrates that the framework can be implemented with different LLM backends (ChatGPT, GPT-4, Llama) to balance cost and performance. The modular design also allows selective tool usage based on resource constraints. With the development of future LLMs, we anticipate the effectiveness and efficiency of FacTool will be significantly improved.

Insufficient experiments with outdated models

The experiments use ChatGPT and GPT-4, which were state-of-the-art at the time of writing. While newer reasoning models like DeepSeek-R1 have emerged, the framework's modular design means it can readily incorporate these newer models. The core contribution - the systematic tool augmentation approach - remains valid regardless of the underlying LLM. We will update the paper with results of newer model (including DeepSeek-R1 and Qwen) in our next version.

Prompt dependency

Indeed, this is a legitimate concern for any prompt-based approach. Our paper provides detailed prompts in the appendix and shows consistent performance across different tasks, suggesting the prompting strategy is relatively robust. The framework could be enhanced by incorporating more sophisticated prompting techniques or fine-tuning approaches.

Overclaim about connecting "tool use" with "factuality detection"

We thank the reviewer for providing the reference. After reviewing the survey paper, we want to respectfully clarify the contribution of FacTool based on the context of the survey. The survey shows that external knowledge augmentation is significantly underexplored in fact-checking, with only five out of 69 surveyed papers employing external sources. Most prior work uses prompting (without tools) or fine-tuning approaches, with knowledge-augmented strategies being identified as a key future direction.

While tools have been used for various NLP tasks, FacTool appears (to the best of our knowledge) to be among the first to systematically design a tool-augmented framework specifically for multi-domain factuality detection. The survey categorizes approaches into prompting, fine-tuning, and augmentation with external knowledge, noting that techniques incorporating external knowledge remain underexplored. FacTool's systematic integration of domain-specific tools (Google Search for KB-QA, code interpreters for programming, Google Scholar for scientific claims) represents a novel contribution to this underexplored area. The framework's significance lies not just in using tools, but in demonstrating how different tools can be systematically selected and applied across diverse factuality detection tasks, providing a template for future tool-augmented fact-checking systems.

Typo

We thank the reviewer for pointing out the typo, we will fix the typo in our next version paper.

[3] Vykopal et al., Generative Large Language Models in Automated Fact-Checking: A Survey.

We thank the reviewer's thoughtful review. We appreciate your recognition of our paper's organization, clarity, and comprehensive experimental details. We would like to respectfully address your concerns:

The novelty of individual components in FacTool seems to be limited

We respectfully emphasize that FacTool represents one of the earliest works that pioneered the integration of external tools to systematically address hallucination in LLMs. While individual components may seem familiar now, at the time of this work, the unified framework combining:

• Multi-domain claim extraction via LLM prompting
• Tool-augmented evidence collection
• Systematic verification across diverse tasks (KB-QA, code, math, scientific)

was genuinely novel. This approach has since become the de facto standard for addressing hallucination, validating our early vision. The fact that tool-augmented approaches are now considered the most effective method for combating hallucination demonstrates the impact and foresight of our work.

Missing baselines for claim extraction and RAG comparisons

We acknowledge these suggestions would strengthen the paper. We also note that:

• For claim extraction, at the time of our work, there were limited comparable automated claim extraction methods that worked across our diverse domains. Existing methods were typically task-specific (e.g., only for summarization or QA), making fair comparison challenging.
• For RAG baselines, our focus was on demonstrating the effectiveness of dynamic tool use rather than static retrieval, as tools provide real-time verification capabilities that static datastores cannot match. Static RAG systems would struggle with dynamic content like code execution results or real-time mathematical calculations.
• For code execution baseline, we actually do compare against execution-based evaluation (HumanEval), but our approach goes beyond simple pass/fail by generating diverse test cases and using majority voting to identify subtle errors that fixed test suites might miss.

Limited transferability experiments and small sample sizes

We would like to respectfully emphasize that our experiments conducted across four distinct domains (KB-QA, code, math, scientific) already demonstrate significant generalization across domains and scenarios. We emphasize that our dataset comprises over 400 responses across tasks, containing nearly 1,000 claims. This significantly surpasses related works like RARR [1] and SelfCheckGPT [2], which are limited to 150 and 238 samples respectively, and focus on only a single KB-QA task.

Furthermore, we clarify that our work primarily focuses on providing a tool for evaluation, not on benchmarking the factuality of modern LLM chatbots - the chatbot evaluation in Section 6.3 is more of an additional demonstration of FacTool's utility rather than the core contribution.

[1] RARR: Researching and Revising What Language Models Say, Using Language Models, Gao et al.

[2] SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models, Manakul et al.

We respectfully appreciate the reviewer's feedback. While we acknowledge that computational cost is a valid consideration, we would like to emphasize several key points:

• High-stakes applications: In domains like healthcare, finance, and scientific research, the cost of factual errors far exceeds the computational cost of verification.
• Selective deployment: FacTool can be strategically deployed for critical outputs rather than every generation.
• Improving efficiency: As one of the first works in this space, future optimizations and more efficient tool integration methods will naturally follow.
• Opensource alternatives: Our experiments with FacTool powered by Llama show competitive results, demonstrating that opensource models can effectively power the framework. With the increased capabilities of opensource models, the cost of deploying FacTool will continue to significantly reduces.