MetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use

Q: The experiments are done zero-shot. I wonder if few-shot prompting (with chain of thought reasoning) can improve the performance (on both tasks) by a lot.

A: We supplemented the few-shot prompts experiments about tool usage awareness and the first three subtasks of tool selection, as outlined in Appendix C.4 of our experimental setup. Due to the lack of annotated CoT data, we had the LLMs generate a segment of reasoning to explain their answers, achieving a similar purpose of CoT.

Tool Usage Awareness:

Tool selection: Sub-task 1 & Sub-task 2:

Q: Relevant to the above two: would it be more prudent to measure “whether LLMs know when to use tools / whether they can use tools” by actually performing tool use experiments – the metric would be the accuracy on the downstream task (e.g., whether the weather actually matches internet search, or whether some calculation actually equals the correct answer, or whether a 2-hour reminder is created in some simulator)?

A: Thank you very much for your suggestion. However, we would like to emphasize that the primary focus of this paper is to assess whether LLMs exhibit tool usage awareness (knowing when to use tools) and whether they can choose the correct tools. Our emphasis is on the inherent capabilities of LLMs. In contrast, the metric you mentioned, "the accuracy on the downstream task," largely depends on the settings and capabilities of the tools themselves, which is not closely related to the central theme of this paper. Figure 1 in the paper illustrates our key points, highlighting 1 and 2 (Thought and Action), rather than 3 and 4. Furthermore, we need to reiterate the contributions of this paper:

1. We introduce ToolE, a comprehensive dataset that encompasses a wide range of 21,127 user queries, with both single-tool and multi-tool queries. Different from the previous single-method generation, these queries are generated using various prompting methods, including emotional generation, keyword generation, direct diverse generation, and detailed generation. Moreover, to address the challenge of overlapping tool functionality, we undertake tool merging and decomposition.
2. Evaluation on awareness of tool usage and tool selection. We construct a test set to evaluate the awareness of tool usage based on ToolE and existing instruction datasets. Moreover, we formulate four distinct tasks to evaluate the tool selection ability of LLMs. These tasks are thoughtfully designed to assess semantic comprehension, adaptability, reliability, and inferential capability, namely tool selection with similar choices, tool selection in specific scenario, tool selection with possible reliability issues, and multi-tool selection.
3. We rigorously evaluate the performance of eight well-known LLMs. We have observed that most LLMs struggle to recognize their capability boundaries and lack a good awareness of tool usage. Regarding tool selection, we find that while LLMs possess basic tool selection capabilities, the tool selection of most LLMs remains unreliable, with noticeable variations in performance across different daily scenarios. Moreover, the error analysis indicates there is still room for improvement in tool selection. Finally, by analysis the tool description, we gained two insights for tool developer.

Q: I wonder how the authors define a reference answer (on whether LLMs need to use a tool) if with or without the tool, LLM can both solve the question.

A: We are sorry for your confusion. To better understand the Tool Usage Awareness dataset, we will describe how we selected positive and negative samples for the dataset. Firstly, we categorized user queries into three types: The first type is "Queries must be solved by tools" (positive), such as multimodal input and real-time information retrieval. The second type is "Queries can be solved well by all LLMs" (negative), such as telling jokes, basic conversational functions, sentiment classification, and other basic NLP tasks. The third type represents the middle ground between the first and second types of user queries, i.e., queries we hope LLMs can solve but currently cannot, such as complex calculations, long text summarization, and information extraction.

With the aforementioned types of user queries, we classified user queries into positive or negative samples using human evaluation and model checking.

(1) Human evaluation: For the first and second types, we determined the samples through unanimous agreement from two human experts and referenced the four reasons in Appendix A.4 for selection.

(2) Model Checking: Regarding the third type of user query, our objective is to find those within this category that can be solved well by all LLMs (classified as negative) and those that none of the LLMs can solve (classified as positive). We discarded user queries that only a portion of the LLMs can solve them. We conduct validation in the following two steps:

We initially input the queries into GPT-4. Since GPT-4 currently has the best performance in terms of utility, if GPT-4 declines to answer (i.e., unable to solve the problem or refuse to answer), we classify it as a positive query.

Following the above operation, if the query is not classified as positive, we conducted inference on eight LLMs and evaluated the answers through two human experts. If all output from LLMs solves the problem well, we classify them as negative queries.

We have added detailed information in Appendix B (blue text).

Thank you very much for your feedback. We will explain step by step to clarify your doubts and thus, as much as possible, eliminate any misunderstanding about your contribution to the article.

Q: I’m a bit confused because many of those questions are quite difficult, and intuitively they would definitely benefit from retrieval. The reason LLMs can solve them might be because they appeared in the pretraining/fine-tuning datasets already, so LLMs remember the examples. So why are those instruction-tuning datasets in the negative subset (if it wouldn’t harm if they search for info on the internet)?

A: Firstly, it is crucial to understand the significance of incorporating negative samples.

The purpose of introducing negative samples is to ensure the fairness of the tool usage awareness evaluation. Without them, models might excessively rely on tools for all types of queries, which is not the ideal performance we expect from LLMs. For example, if LLMs use tools in all situations, including answering simple mathematical questions (like "5 * 8 = ?"), they would be no different from standard search engines, losing their characteristics as advanced language processing systems. Therefore, introducing negative samples helps make task assessments more comprehensive and balanced.

Secondly, we need to redefine the fundamental principles of LLMs using tools. It is reasonable to use tools when queries (instructions) exceed the processing capabilities of LLMs. For instance, even though basic operations (such as addition, subtraction, multiplication, and division) can be completed by tools with possibly very accurate results, if LLMs can handle these problems independently, then using tools becomes redundant. Otherwise, LLMs' advanced natural language processing and generative capabilities would not be fully demonstrated, and further research and enhancement of their computational and reasoning abilities would be meaningless. Therefore, as we proposed in Appendix A.4, LLMs should consider using tools only in the following four situations: when real-time data or external databases are needed (A), for processing special inputs/outputs, such as multimodal inputs (B), for handling domain-specific tasks, such as complex computational tasks (C), and in scenarios involving user interaction, like playing games (D).

We will introduce how we select positive and negative samples in the next question.

Q: Did you use the SFT-tuned & RLHF-tuned llama (which should be llama2-chat) or the plain pretrained llama (without SFT/RLHF fine-tuning)?

A: Thanks a lot for your careful consideration. We used the chat version and have added details in the experiment setting.

Q: I wonder what the tool descriptions are. I don’t see detailed tool descriptions in the paper. I also don’t see any supplementary materials in the submission.

A: According to your suggestion, we have uploaded the description information of the tool to the supplementary materials. The developers themselves write the descriptions of these tools and you can also find them in the OpenAI plugin store.

Q: Ethics Concerns. One of the specialized scenarios in the test set (the second task in Section 2.3.2) is called “housewife” covering topics like “discount, restaurant booking tool, product search, etc.” citing videos like “AI for women today, 10 genius ways housewives can use ChatGPT to save money.” I recognize many housewives but there seems to be a hint of prejudice and stereotyping there…? It would be nice to rephrase this category of examples. Perhaps I’m saying this because of my current physical location or social environment?

A: We sincerely apologize for using the term "housewife" in our paper, which may carry a biased connotation, and we deeply appreciate your feedback. Based on your suggestion, we realize that biased terminology might inadvertently reinforce inappropriate stereotypes. Therefore, in the relevant sections of our paper, we plan to replace "housewife" with "home manager" to more fairly and broadly represent those who manage daily household affairs and budgets. This change will be reflected in our revised manuscript. We hope this modification will more accurately convey the intent of our research while avoiding any potential bias.

Finally, heartfelt thanks for the comments you provided on this paper. We hope that our responses can help resolve any confusion you may have.

Q: What would be the human performance on these tasks?

A: We evaluated human abilities through questionnaires to investigate human performance in tool selection. Specifically, we mixed questions from four sub-tasks, asking participants to select 0 to 2 tools for each question. Each questionnaire comprised 10 or 15 questions, with participants making choices based on provided queries and candidate tools as options. We collected a total of 340 valid responses due to the time limitation.

Model$_{Max}$ is the best performance of LLMs.

Model$_{Avg}$ is the average performance of eight LLMs.

We observe a notable discrepancy between the CSR of humans and LLMs in sub-task 1, sub-task 2, and sub-task 3. Human CSR surpasses both the average and maximum CSR of LLMs. In sub-task 3, human performance reaches an impressive 96%, starkly contrasting to the model's meager 9%. This discrepancy highlights LLMs' challenges, particularly in addressing issues like hallucination, significantly impacting their reliability.

Moreover, in sub-task 4, while surpassing the average level of LLMs, human performance falls short of reaching their maximum CSR. This implies that LLMs still maintain a distinct advantage when confronted with intricate language tasks, such as multiple-choice questions.

We have added detailed information in Appendix C.5 (blue text).

(In continuation of the previous answer)

Sub-task 3:

We observed that, in the tool usage awareness task, the accuracy of LLMs showed both improvements and declines, but the F1 Score increased. This indicates a significant enhancement of few-shot prompts on the tool usage awareness task.

Additionally, in the tool selection task, for subtask 1, we found that, apart from Vicuna-7b (possibly due to the model's poor robustness and high sensitivity), few-shot prompts improved the performance of other LLMs. However, in subtask 2, this improvement was limited for most models. ChatGPT achieved a remarkable 28.14% improvement, and ChatGLM2 also saw a 9.05% improvement. The Vicuna series and LLama2 series models showed less than 5% improvements. In subtask 3, few-shot prompts enhanced the performance of most LLMs, especially for the LLama2 series models, which demonstrated improvements in CSR across various scenario scenarios.

Dear Reviewer,

Thank you for your invaluable assistance and support. Given the constraints of time, we wish to ensure that our responses have effectively addressed any concerns you may have had. If there are still lingering issues, please feel free to inform us. We eagerly anticipate your additional feedback and hope that, if all your primary concerns have been resolved, you may reconsider raising your score.

Once again, we appreciate your time and effort in reviewing our paper.

I really appreciate the detailed response. I'm still going over the results and other reviewers' comments, and I do plan on increasing the score in the next day or two.

Dear reviewer xJ5F,

We have addressed all concerns raised in the initial review comprehensively, including the incorporation of additional experiments and detailed explanations. Your feedback is crucial to us, and we kindly request your prompt attention to our rebuttal. If there are any further questions or points of clarification needed, please do not hesitate to let us know. Your timely response would be greatly appreciated.

Q: Most statements in the experiments are not well explained.

A: To address your concern, we have carefully reviewed the analysis of our experimental results and found that we indeed did not effectively analyze the experimental outcomes. We have implemented the following improvements:

1. Additional Experiments: We added few-shot prompt experiments for three subtasks of the tool usage awareness task and the tool selection task; we introduced a new prompt template in multi-tool selection to specify the number of tools returned by LLMs.
2. More In-Depth Analysis of Experimental Results: For both the original and the additional experiments mentioned above, we conducted a comprehensive and profound analysis of our experimental results. We have reached the following conclusions:

(1) Even under few-shot prompts, the majority of LLMs still perform poorly in tool usage awareness. In Table 2, we observe that under the zero-shot prompt, only ChatGPT has both accuracy and F1 score exceeding 70%, while the performance of other models is relatively poor, with the F1 score of llama2-13b being only 11.53%. Under the five-shot prompt, some models show significant improvement in F1 scores, for example, llama2-13b increased by 42.79%, and vicuna-7b by 42.28%. This indicates that though few-shot learning generally improves the performance of LLMs in tool usage awareness, they still lack sufficient tool usage awareness.

(2) When selecting similar tools, there is a significant performance disparity among existing LLMs, and the improvement brought by few-shot prompts is limited. Table 3 shows that under the zero-shot prompts, the best-performing LLM is Vicuna-7b, with nearly a 30% difference compared to the worst-performing Llama2-13b. The gap between the best-performing ChatGPT and the worst-performing Llama2-13b still exceeds 20% under 5-shot prompts. Additionally, the maximum improvement brought by 5-shot prompts does not exceed 7%. Moreover, the performance of Vicuna-7b even declined by 10% under the five-shot condition, suggesting a potential bias in its 0-shot performance, which reflects either a lack of robustness or over-sensitivity of the model.

(3) LLMs still face serious challenges in dealing with reliability issues, such as reducing hallucinations. As seen from Table 3, although few-shot prompts improve the performance of all LLMs, the CSR of most LLMs remains below 20%. We find that LLMs sometimes fabricate non-existent tools, a severe hallucination issue that has a significant impact on LLM-based agents. Additionally, the potential sycophancy of LLMs may lead them to avoid returning a "none" answer, instead choosing irrelevant tools to respond to users.

(4) LLMs perform poorly in processing long texts. From Figure 4, we can see that the CSR of almost all LLMs decreases as the length of the tool list increases, especially in the range from the top 5 to the top 10. This indicates that LLMs still need improvement in understanding long texts. LLMs exhibit imbalances and biases in tool selection across different scenarios. For example, in Figure 5, LLMs generally have a higher CSR in tool selections related to the elderly and artists & designers, while their CSR is lowest for tools related to students. This means that developers still need to enhance the generalization capabilities of LLMs. At the same time, for downstream applications, it is best to choose suitable LLMs based on different applied fields.

(5) There are significant performance differences among LLMs in multi-tool selection. As shown in Table 4, ChatGPT, the top-performing model, outperforms ChatGLM2, the worst-performing model, by nearly 70%, highlighting the variability in the capabilities of different language models for this task. Furthermore, the most common error made by the models is omitting tool selection, such as in the case of Vicuna-33b, which only selected one tool in 48.49% of cases. Moreover, several LLMs overly rely on the explicitly specified number of tools they should select in the prompts. When explicitly instructed to return two tools, Vicuna-33b's correct selection rate increased to over 90%, and Vicuna-7b also improved by over 20%. This indicates that these LLMs still possess good multi-tool selection capabilities but require prior knowledge, which makes it challenging to apply in LLM-based agents.

The detailed analysis for the above conclusions is updated in the PDF document, marked in blue font.

Q: I am a little bit confused about the second conclusion you have obtained in Section 3.3.

A: Thank you very much for your suggestion and for expressing such detailed concerns! To address your concerns, we have rewritten all tool descriptions using GPT-4 and LLama2-70b and examined the performance of LLMs on subtask 1 of tool selection. The performance is as follows (we have also displayed this result in Figure 6 in the PDF):

We observed that different rewritten LLMs yielded varying benefits for different groups. For instance, descriptions rewritten by Llama2-70b resulted in a 7.83% improvement for llama2-13b but did not significantly enhance the performance of the Vicuna series models. In contrast, descriptions rewritten by GPT-4 caused a sharp decline in the performance of ChatGLM and Llama2 series, while significantly boosting the Vicuna series, possibly due to the Vicuna series' training corpus being largely sourced from ShareGPT.

For a more accurate expression, we have modified the original statement "descriptions generated by chatgpt are better than those written by the tool developers themselves" to "we strongly recommend that tool developers choose an appropriate rewrite model for generating new descriptions based on the downstream LLM the tool will apply to.“

Thank you very much for your suggestion! We apologize for the confusion caused, and next, we will gradually explain your concerns, to give you a clearer understanding of your contributions to the work:

Q: I think the comparison with existing benchmarks still needs polishing.

A: First, we acknowledge that there are some deficiencies in the comparison with benchmarks, which may be due to differences in understanding from various perspectives of different persons. Based on the issues identified after your careful review, we have revisited the related literature and engaged in in-depth discussions. We have made modifications to the comparison as per your request.

Thank you for your efforts during the rebuttal stage. My concerns have been resolved. I will maintain my positive score and vote for acceptance. Also, I noticed that all the rebuttals are not available to the public. Please double-check.

We greatly appreciate your feedback! We will address your concerns one by one:

Q: Section 2 is rich in substance but the presentation is disorderly.

A: Thanks for your feedback. We acknowledge that we did not organize the structure of this section well. However, swapping the positions of dataset generation and task formulation might not be suitable. This is mainly because many definitions in task formulation rely on the premise that the reader is already clear about the structure of the dataset. Therefore, placing dataset generation after task formulation might cause more confusion for the reader.

However, we appreciate your expression of confusion, so we have added a brief introduction at the beginning of Section 2 (already updated in the PDF in blue font) to help readers better understand the structure of the section.

Q: It's unclear (and seemingly undiscussed) how fundamentally the generation of the test queries with LLMs biases the evaluation for/against particular models or reduces the true diversity of these queries.

A: Thank you for pointing out this issue! We acknowledge this is a concern and still exists in the current LLMs-generated dataset.

Specifically, we attempted to use various models to generate a mixed dataset, such as Llama2-70b. However, in the actual generation process, we found that the diversity of samples generated by Llama2-70b was lacking, and the quality of generated samples was low (e.g., not relevant to the tool's description information). However, MT-Bench [1] indicated that ChatGPT has a low bias on self-enhancement. Therefore, we believe ChatGPT does not strongly favor its own generated content, making it the least biased choice for generating data while ensuring data quality.

Q: What kind of dev vs. test guidance do you recommend for this benchmark?

A: Your suggestion is forward-thinking. Inspired by your advice, we have started constructing dev. Specifically, we are going to create a fine-tune dataset for tool usage awareness. We plan to use all positive samples (all LLMs must use the tool to answer questions) as input to GPT-4 and obtain answers (output should be a refusal to answer, e.g., “Sorry, I cannot answer this…”). We intend to use positive samples and their corresponding answers as the fine-tune dataset to improve LLMs' tool usage awareness. This ensures LLMs can accurately identify inputs beyond their ability and refuse to answer, allowing external tools to assist in answering rather than generating illusory or low-quality answers. Once the dataset is completed, we will release it together.

Q: Which Llama2 models were used, vanilla or chat?

A: We used the chat version, and we have added details in the experiment setting. Thanks for your careful consideration.

Q: You find that "the worst-performing model, Llama2-13b, has an F1 Score of only 11.53%." This is surprisingly low. Is this some kind of unintended interaction between the chat format or some other minor detail and the task?

A: We admit that we were surprised when we first saw this data, so we manually reviewed the results. Since we require LLMs to explain their answers, we consider this answer relatively reliable. We believe the main reason for this result is the lack of capability awareness, meaning the model fails to have a clear understanding of its capability boundaries. The drawback of Llama2-13b lies in its overconfidence; for questions it cannot solve, it still believes it can.

Here is an example of Llama2-13b’s answer:

Question: Can you check if there are any trending discussions related to the Sakura festival occurring in Japan on Google Trends or Twitter?

Answer: No, I do not need to use external tools to address the user's query. The information the user is seeking can be found through a simple search on Google Trends or Twitter, and I have the ability to perform such searches within my language model capabilities. I can provide the user with relevant information related to the Sakura festival in Japan, such as recent discussions or trends, without the need for external tools.

Q: Maybe rename the four tool selection tests to "sub-tasks."

A: Thank you for your suggestion; we have made the modification.

[1] Zheng, Lianmin, et al. "Judging LLM-as-a-judge with MT-Bench and Chatbot Arena." arXiv preprint arXiv:2306.05685 (2023).

Thank you very much for your valuable feedback, especially regarding "constructing a fine-tune dataset," which will guide our work to make a more significant contribution. If you have any further questions, please feel free to ask.

Thank you very much for your valuable feedback. We apologize for any confusion caused by certain details in the paper. We will address each of your concerns and provide explanations to help you better understand the contributions of this paper step by step:

Q: One of the things mentioned is that other works lack diverse user inputs whereas this work does have diverse user inputs. Experiments or evidence for this is missing?

A: Our diversity is manifested in several aspects. The first aspect is the diversity of user inputs in terms of language style as shown in 2.2.1 (prompt methods). For example, we used different emotional prompts to generate user inputs with different emotions, such as: "I’m feeling really down today, can you summarize this YouTube video for me?" which conveys a depressed emotion. Additionally, the direct diverse generation method generates queries in various expressions, including order (Tell me the ...) and requests (Could you please help me...). Secondly, diversity is reflected in the level of detail, and our length distribution is presented in Figure 11 in the appendix. Furthermore, our generated data covers over a hundred different topics, as showcased in Figure 12, or you can refer to the links at the end of the abstract for data interaction.

Q: Keyword Generation and Details Generation: What’s the exact difference?

A: The main distinction between Keyword generation and Detail generation lies in their respective focuses and approaches. Keyword generation involves extracting key terms or phrases from the description of a tool, such as the platform (e.g., YouTube) it is intended for or the specific regions (e.g., USA) it is applicable to. It aims to expand upon localized and crucial aspects of the tool's description.

For instance, consider the tool "Now" with the following description: "Get Google Trends. In Japan, you can also get Twitter trends and search Twitter keywords." Using ChatGPT, we extract 5 keywords: Trend, Tracker, Japan-Specific, InsightsTwitter, and TrendTrend Explorer. As a result, we can generate a query based on "Japan-Specific: What are the trending topics on Twitter in Japan? This demonstrates how Keyword generation extracts keywords from the tool description and uses them to formulate queries that delve into specific details related to those keywords, in this case, focusing on Twitter trends in Japan.

On the other hand, details generation aims to enrich the overall query by adding diverse and intricate information, not necessarily confined to extracting keywords from the tool description. For the same tool "Now," a details generation approach results in a query: "Get me the latest Google Trends in Japan, including the top search queries, rising search topics, and trending searches across various categories such as news, entertainment, and sports." This example showcases how details generation provides a more comprehensive enhancement of the query, incorporating random yet detailed information on a broader scale.

(All the prompt templates using ChatGPT are provided in Appendix D.1.)

Q: Questions about Multi-tool selection: How do you select the popular tools? Manual?

A: We use the original tool number of merged tools as a reflection of its popularity. You can find this information in Table 9 in the appendix. For example, we can see that FinanceTool has a total of 22 tools merged, so we consider it to have the highest popularity.

Q: Questions about Multi-tool selection: What do you think of the quality of multi-tool queries? Are there any human validations on the percentage of tool combinations that are useful in the dataset?

A: We conducted manual screening, focusing primarily on two aspects: 1. Is the combination of tools reasonable? 2. Does the query of the tool correspond to the tool description itself? Similar to single-tool queries, we also manually verified multi-tool queries, removing nearly 2/3 of them. Thanks for your feedback, we have added more details in the paper.

Q: Awareness dataset: It’s unclear how this dataset can be useful.

A: The specific reasons for constructing this dataset are explained in Section 2.1. We aim to create a dataset that includes queries that LLMs can handle well on their own, as well as queries that all LLMs cannot address. Evaluating whether LLMs possess tool usage awareness is crucial because one of the primary advantages of LLM-based agents is their ability to use tools. For instance, if an LLM has a good understanding of tool usage, it can resort to external tools for assistance when it realizes it cannot effectively solve a user's problem. This helps reduce hallucination issues while enhancing utility. Additionally, the introduction of negative examples aims to make the task of tool usage awareness more fair. Once negative examples are excluded, the ultimate goal is for the model to seek tools for all questions, which is not an ideal representation of an LLM.

Q: Furthermore, it’s unclear if the negative examples truly do not require any tools.

A: We are aware of the confusion caused by our failure to clearly express the selection of negative samples in the paper. In short, for each question, we conducted inference with all nine LLMs, and through manual inspection, ensured that each LLM could effectively address the problem. This ensures that negative examples do not require additional tools. To better understand the Tool Usage Awareness dataset, we will describe how we selected positive and negative samples for the dataset. Firstly, we categorized user queries into three types: The first type is "Queries must be solved by tools" (positive), such as multimodal input and real-time information retrieval. The second type is "Queries can be solved well by all LLMs" (negative), such as telling jokes, basic conversational functions, sentiment classification, and other basic NLP tasks. The third type represents the middle ground between the first and second types of user queries, i.e., queries we hope LLMs can solve but currently cannot, such as complex calculations, long text summarization, and information extraction.

With the aforementioned types of user queries, we classified user queries into positive or negative samples using human evaluation and model checking.

(1) Human evaluation: For the first and second types, we determined the samples through unanimous agreement from two human experts and referenced the four reasons in Appendix A.4 for selection.

(2) Model Checking: Regarding the third type of user query, our objective is to find those within this category that can be solved well by all LLMs (classified as negative) and those that none of the LLMs can solve (classified as positive). We discarded user queries that only a portion of the LLMs can solve them. We conduct validation in the following two steps:

We initially input the queries into GPT-4. Since GPT-4 currently has the best performance in terms of utility, if GPT-4 declines to answer (i.e., unable to solve the problem or refuse to answer), we classify it as a positive query.

Following the above operation, if the query is not classified as positive, we conducted inference on eight LLMs and evaluated the answers through two human experts. If all output from LLMs solves the problem well, we classify them as negative queries.

We have added detailed information in Appendix B (blue text).

Q: There are concerns about how this experiment is set up or the motivation behind tool selection with possible reliability issues.

A: I apologize for not making it clear in the paper. We designed this subtask because we wanted to test the reliability of LLMs in tool selection. That is, when there are no suitable tools to solve the user's problem in the tool list, it should return "none" instead of fabricating a non-existent tool (hallucination), thus spreading misinformation, which is a realistic issue. This is crucial for trustworthy LLMs and trustworthy LLM-based agents.

To ensure the impact of other overlapped tools in the tool list, we not only addressed overlapped issues at the beginning (details can be found in 2.2.1) but also removed the top $t$ tools with the highest semantic similarity to the ground-label tool. This was done to ensure that the tool list is clean (this operation is explained in the last sentence of our description of subtask 3).

Q: No Related work

A: Compared to other datasets, we believe ToolE has two advantages: (1) Our dataset exhibits greater diversity, and this diversity is tailored to real user scenarios, such as expression style, mood, and level of detail. As mentioned in Q1, we employed various prompt methods to induce LLMs to generate more diverse user inputs, ensuring that ToolE covers a wide range of inputs representative of real users. (2) By employing a pipeline process to address the overlapped issue, we can ensure the rigor of the data. As outlined in section 2.2.1 of the paper, we adopted multiple steps to resolve the overlapped issue, ensuring that there is no overlap between tools, which is crucial for ensuring the quality of the dataset and task evaluation. We provide additional information on dataset comparison in A.5 and Table 8 to better illustrate the distinctions between the datasets.

Q: The aspect of arguments for each of the tools and execution is not focused on in this work.

A: I apologize for the confusion. We did not focus on arguments in this paper because our emphasis was on the awareness of tool usage by LLMs and whether they could make correct tool selections, not the tools themselves. Arguments, on the other hand, focus on whether LLMs can correctly use tools, addressing how they handle tool inputs and outputs. Figure 1 in the paper illustrates our focal points, emphasizing 1 and 2 (Thought and Action), rather than 3 and 4.

Dear Reviewer,

Thank you for your invaluable assistance and support. Given the constraints of time, we wish to ensure that our responses have effectively addressed any concerns you may have had. If there are still lingering issues, please feel free to inform us. We eagerly anticipate your additional feedback and hope that, if all your primary concerns have been resolved, you may reconsider raising your score.

Once again, we appreciate your time and effort in reviewing our paper.

Dear reviewer MgrT,

We have addressed all concerns raised in the initial review comprehensively, including the incorporation of additional experiments and detailed explanations. Your feedback is crucial to us, and we kindly request your prompt attention to our rebuttal. If there are any further questions or points of clarification needed, please do not hesitate to let us know. Your timely response would be greatly appreciated.