diff --git "a/20240318/2310.05155v2.json" "b/20240318/2310.05155v2.json" new file mode 100644--- /dev/null +++ "b/20240318/2310.05155v2.json" @@ -0,0 +1,629 @@ +{ + "title": "Toolink: Linking Toolkit Creation and Using through Chain-of-Solving on Open-Source Model", + "abstract": "Large Language Models (LLMs) have demonstrated remarkable progress in utilizing tools, but their closed-source nature and high inference costs pose limitations on their adaptability, necessitating a valid method that leverages smaller, open-sourced models. In this paper, we introduce Toolink, a comprehensive framework that performs task-solving by first creating a toolkit and then integrating the planning and calling of tools through a chain-of-solving (CoS) approach. We first validate the efficacy of Toolink in harnessing the model\u2019s creativity and CoS ability on ChatGPT. Subsequently, we curate CoS-GPT, a chain-of-solving dataset designed for tool-using, and finetune the LLaMA-7B model. It results in LLaMA-CoS, a powerful open-source model with advanced tool-planning and tool-calling capabilities. Evaluation of diverse tasks from BIG-bench demonstrates its CoS ability matches that of ChatGPT while its performance surpasses the chain-of-thought approach. Further studies highlight the generalization of LLaMA-CoS to unseen tasks and showcase its capability in using toolkits not explicitly tailored for the target task, affirming its robustness in real-world scenarios. All codes and data are released111https://github.com/qiancheng0/Toolink.", + "sections": [ + { + "section_id": "1", + "parent_section_id": null, + "section_name": "Introduction", + "text": "###figure_1### Large Language Models (LLMs) such as Codex (Chen et al., 2021 ###reference_b5###), ChatGPT (OpenAI, 2022 ###reference_b22###), and GPT4 (OpenAI, 2023 ###reference_b23###) have made significant strides in code generation, in-context learning, and logical reasoning. However, they still struggle with precise calculations and accessing current information (Patel et al., 2021 ###reference_b27###; Trivedi et al., 2022 ###reference_b40###; Lu et al., 2022b ###reference_b20###). To address these issues, research has focused on equipping LLMs with tools such as calculators (Cobbe et al., 2021 ###reference_b8###; Parisi et al., 2022 ###reference_b26###; Schick et al., 2023 ###reference_b33###), search engines (Carlini et al., 2021 ###reference_b4###; Thoppilan et al., 2022 ###reference_b38###; Schick et al., 2023 ###reference_b33###), scratch pads (Nye et al., 2021 ###reference_b21###), calendars (Schick et al., 2023 ###reference_b33###), and image retrievers (Sheynin et al., 2022 ###reference_b35###) to enhance their capabilities, thus benefiting various tasks including question-answering, math calculations, and long-form generation. Recent studies have also explored how LLMs can devise plans, make decisions, and perform tool invocations (Shen et al., 2023 ###reference_b34###; Lu et al., 2023 ###reference_b18###; Liang et al., 2023 ###reference_b13###). By combining them into a pipeline, these frameworks aim to construct more advanced NLP systems for improved task performance.\nHowever, current tool-using pipelines heavily rely on closed-source models with inaccessible parameters. It poses challenges particularly as follows: i) Limited adaptability: The closed-source nature of major LLMs prevents them from customization, resulting in a lack of flexibility to adapt to tasks with specific requirements. ii) Low efficiency and high inference cost: Many existing LLMs can only be accessed online, which imposes limitations on the inference rate and leads to high expense. iii) Privacy and security concerns: Each query must be submitted to these closed-source LLMs to obtain a tool-using solution, which raises concerns regarding potential privacy breaches and compromises data security.\nTo address these challenges, we propose Toolink, a comprehensive framework to boost the tool-using ability of open-source models with the help of closed-source models. As shown in Figure 1 ###reference_###, Toolink first decomposes the target task by creating a toolkit for problem-solving, and then leverages the open-source model to use tools to answer queries in a chain-of-solving (CoS) approach. Specifically, CoS disentangles the model\u2019s reasoning through two stages: CoS-Planning, which selects useful tools from the created toolkit and plans their usages based on the specific query; and CoS-Calling, which focuses on deriving the answer by performing tool invocations in code format according to the plan devised. To effectively train the open-source model in these abilities, we employ ChatGPT to curate CoS-GPT, a training dataset that aims to inspire the tool-using ability of open-source models through CoS. Specifically, we finetune LLaMA-7B (Touvron et al., 2023 ###reference_b39###) into LLaMA-CoS, which is equipped with strong tool-using capabilities by linking toolkit creation with the chain of problem-solving.\nLLaMA-CoS can solve the queries offline without uploading queries to closed-source models, ensuring data security and privacy. Experiments further illustrate that Toolink outperforms the chain-of-thought (CoT) (Wei et al., 2022 ###reference_b41###) on diverse tasks from BIG-bench (Srivastava et al., 2022 ###reference_b36###) and enables LLaMA-CoS to showcase comparable CoS ability to that of ChatGPT. In addition, LLaMA-CoS can generalize to unseen tasks by planning and calling tailored tools, and solve the target task with a toolkit not specifically tailored for it. These findings further affirm our framework\u2019s robustness in solving queries under real-world scenarios." + }, + { + "section_id": "2", + "parent_section_id": null, + "section_name": "Related Work", + "text": "" + }, + { + "section_id": "3", + "parent_section_id": null, + "section_name": "Toolink Framework", + "text": "###figure_2### As shown in Figure 2 ###reference_###, Toolink first adopts toolkit creation to break down the target task through generating potential tools for task-solving (\u00a7 3.1 ###reference_###). Then, the model links these created tools to address specific queries by selecting pertinent tools from the toolkit, planning their uses, and performing tool invocations (\u00a7 3.2 ###reference_###). This new reasoning approach, referred to as chain-of-solving (CoS), not only enables the effective and coherent application of tools but also facilitates the tool-using adaptation on the open-source model (\u00a7 3.3 ###reference_###)." + }, + { + "section_id": "3.1", + "parent_section_id": "3", + "section_name": "Toolkit Creation", + "text": "Toolkit creation decomposes a general task into modular and essential tools for problem-solving, facilitating more flexible tool utilization.\nGiven the target task , toolkit creation breaks it down into more manageable components through generating a toolkit , where represents the tool to solve the subtask . We illustrate our approach in Figure 2 ###reference_###A, where the target task is decomposed into (movement in a single direction) and (change of orientation). Each component is represented by a specific implementation encapsulated within a function tool.\nWe utilize ChatGPT for task decomposition. For each task , we provide ChatGPT with a task description and a few data samples , expecting them to facilitate the model\u2019s understanding of task \u2019s objective and identify commonalities among queries. The prompting details are presented in Appendix A ###reference_### and Figure 6 ###reference_###. Note that our design requires only a few data points as demonstrations fed into the closed-source ChatGPT, leaving the entire testing set for local processing to maintain privacy.\nEach tool within the toolkit is comprised of a concise introduction and its corresponding code implementation. The introduction provides a brief overview of \u2019s utility, inputs, and outputs, facilitating effective planning and calling in subsequent steps." + }, + { + "section_id": "3.2", + "parent_section_id": "3", + "section_name": "Chain-of-Solving", + "text": "Chain-of-solving (CoS) involves deliberate planning and decision-making for tool invocation, which bridges the gap between toolkit creation and downstream tool use for task query resolution. CoS is disentangled into CoS-Planning and CoS-Calling. This separation allows for a more transparent and interpretable reasoning path, thereby enhancing the applicability of CoS to open-source models.\nThe CoS-Planning stage entails selecting useful tools from a toolkit in response to a specific query of task . It employs natural language-based reasoning chains, referred to as a plan, to determine the most effective way to utilize the selected tools to solve the given query.\nIn Figure 2 ###reference_###B, the model devises strategies for employing tools to update the location and orientation, with additional initial conditions that serve as a guiding hint. Planning plays a crucial role in establishing a link between toolkit creation and decision-making, thus reducing the cognitive burden associated with tool-use reasoning.\nThe CoS-Calling stage entails the utilization of selected tools and interpretation of the plan into program language to perform tool calls. The plan generated in the previous stage serves as the guidance for program implementation. During the tool execution, all results from the tool invocations are implicitly captured and used to extract the final answer for the given query.\nFigure 2 ###reference_###C illustrates this process, where the model simulates the entire navigation process using code as the underlying medium. In this example, the model derives the final correct answer, thereby demonstrating a successful CoS-Calling process.\n###table_1###" + }, + { + "section_id": "3.3", + "parent_section_id": "3", + "section_name": "Open-Source Model Adaptation", + "text": "Considering the limited adaptability, high inference cost, and privacy concerns posed by closed-source models, we aim to enhance the CoS ability in open-source models. We propose the CoS-GPT, a specialized training dataset that emphasizes the planning and calling of tools, along with code generation. These elements are crucial for boosting the model\u2019s CoS ability. The statistics related to CoS-GPT are presented in Table 1 ###reference_###. Furthermore, for each specific target task , we employ to generate a task-specific dataset. This is achieved by augmenting each sample query with suitable tools, thereby facilitating a more effective training of task on the open-source models.\nTo enhance the open-source model\u2019s skills in applying tools for problem-solving, we construct CoS-GPT from scratch to improve the model\u2019s CoS ability from planning, calling, and coding. We include the first two aspects as they are essential for CoS within Toolink, and the last aspect as it serves as the medium for tool-using.\nFor data points about planning and calling, we enhance the AQUA-RAT (Ling et al., 2017 ###reference_b14###), GSM8K (Cobbe et al., 2021 ###reference_b8###), and TabMWP (Lu et al., 2022a ###reference_b19###) datasets, comprising graduate-level math problems, numerical reasoning tasks, and diverse table contents, with toolkits. Each query is augmented with a toolkit containing both useful and redundant tools. The model\u2019s objective for planning is to select and plan the use of useful tools, while for calling, the objective is to learn how to call the chosen tools through codes. We apply ChatGPT to simulate this process and utilize their responses for dataset construction. Please refer to \u00a7 E.1 ###reference_### for more details.\nData points for code generation encompass diverse sources, including augmentation from existing datasets, GitHub repositories, and newly generated data, detailed in \u00a7 E.2 ###reference_###. Each query adheres to an instruction-following pattern and aims to enhance the open-source model\u2019s understanding of code while expanding its versatility in making informed decisions when performing CoS.\n###table_2### For each target task , we construct 200 tool-augmented data points (100 each for plan and call) from the publicly available samples , and use them to tune the open-source model together with CoS-GPT. Similar to the construction process for tool-using data in CoS-GPT, we first augment with a toolkit . Next, we employ ChatGPT to select useful tools for each query and generate the calling decision. The decision\u2019s output is compared against the standard answer, and minor adjustments may be made to ensure the augmented data\u2019s validity.\nTogether with CoS-GPT, we apply the tool-augmented data points from all target tasks to finetune the open-source model. We expect the derived tool-augmented open-source model to excel in applying useful tools for problem-solving. By planning and calling through CoS, this model links the created toolkit with specific queries, which realizes the final goal of the Toolink framework.\n###table_3### ###table_4###" + }, + { + "section_id": "4", + "parent_section_id": null, + "section_name": "Experiments", + "text": "To evaluate the effectiveness of Toolink, we first conduct a validation test using ChatGPT. We select eight distinct tasks from the BIG-bench dataset (Srivastava et al., 2022 ###reference_b36###) to investigate whether Toolink can effectively leverage ChatGPT\u2019s creativity and tool-using capability to improve task performance.\nSubsequently, we finetune the open-source LLaMA-7B model following the adaptation process outlined in \u00a7 3.3 ###reference_###. This results in LLaMA-CoS, a model capable of linking the created toolkit with specific tool use through CoS. We evaluate the effectiveness of LLaMA-CoS in utilizing tools on the same set of eight tasks and showcase its excellence." + }, + { + "section_id": "4.1", + "parent_section_id": "4", + "section_name": "Validation Evaluation", + "text": "To evaluate the effectiveness of Toolink, we conducted a validation test using ChatGPT on eight tasks of diverse categories from BIG-bench, as detailed in Table 2 ###reference_###. For each task, we first employ ChatGPT to create a toolkit, outlined in \u00a7 3.1 ###reference_###. The total number of tools in the toolkit created for each task is shown in Table 3 ###reference_###, with the tool\u2019s implementation details provided in Appendix B ###reference_###. Equipped with these tools, ChatGPT is presented with instructions and demonstration examples to perform CoS for problem-solving, detailed in Appendix C ###reference_###.\nWe compare CoS against two baselines: i) Vanilla baseline, where ChatGPT directly produces the final answer. ii) CoT baseline (Wei et al., 2022 ###reference_b41###), where ChatGPT performs chain-of-thought reasoning before providing an answer.\nWe evaluate the ability of ChatGPT to leverage plans and calls to perform CoS. The accuracy is measured by matching the model\u2019s final output to the correct answer. In addition, we also evaluated the individual contributions of CoS-Planning and CoS-Calling separately.\nDuring CoS-Planning, the model is asked to select useful tools and plan their uses given the created toolkit. The planning accuracy is thus measured by the following metric:\nwhere denotes the number of correct (useful) tools in the toolkit selected in the model\u2019s generated plan, while denotes the number of incorrect (redundant) tools selected.\nDuring CoS-Calling, the model is asked to implement the plan using code as the medium, after the useful tools are selected. The accuracy is thus measured by matching the output from the final execution with the correct answer. Please refer to Appendix D ###reference_### for more details regarding the separated test of CoS-Planning and CoS-Calling.\nThe results are presented in Table 3 ###reference_###. ChatGPT which uses tools through the CoS approach achieves significantly improved performance compared to other baselines, with notable margins of superiority. Further, the accuracy for CoS-Calling and CoS-Planning individually is even higher, indicating successful reasoning in each step of CoS which links toolkit creation with specific uses. These findings affirm the validity of Toolink, establishing a strong basis for its potential transferability to smaller, open-sourced models." + }, + { + "section_id": "4.2", + "parent_section_id": "4", + "section_name": "Experiments on LLaMA-CoS", + "text": "Our primary objective is to apply Toolink to smaller, open-sourced models. To this end, the models from the LLaMA family (Touvron et al., 2023 ###reference_b39###) stand out due to their capability to perform reasoning and generate codes. Considering the affordability of computational resources, we select LLaMA-7B as the representative base model to evaluate the performance of Toolink on open-source models.\nWe follow the adaptation process outlined in \u00a7 3.3 ###reference_### and finetune LLaMA-7B with CoS-GPT and eight sets of task-specific tool-augmented data. The eight tasks are the same ones as those we test in \u00a7 4.1 ###reference_###. Applying the training setting detailed in Appendix F ###reference_###, we derive a powerful variant, LLaMA-CoS, that excels in using tools through CoS.\nWe use LLaMA-CoS as the representative finetuned open-source model for testing. Building upon the validation test conducted on ChatGPT, we evaluate LLaMA-CoS\u2019s performance on the same set of eight tasks from BIG-bench. We keep all the metrics the same as in \u00a7 4.1 ###reference_###\nAs a comparison to CoS, we employ the chain-of-thought (CoT) reasoning as the baseline. We evaluate both methods under two scenarios: i) Prompting with demonstrations on Alpaca, LLaMA-7B, and ChatGPT, and ii) Finetuning specifically on the LLaMA-7B model. We referred to the LLaMA-7B tuned with CoT data as LLaMA-CoT, while our model, LLaMA-CoS, is specially tuned for tool use through CoS.\nWe present the results in Table 4 ###reference_###. Notably, LLaMA-CoS achieves an impressive average accuracy of 94.74%, outperforming all the CoT baselines, whether tuned or not, by a substantial margin. Compared to ChatGPT, which exhibits strong reasoning and tool-using capabilities under the CoS setting, our tuned model can still achieve comparable performance. These results highlight the effectiveness of CoS in outperforming traditional CoT methods and demonstrate the successful transfer of tool-using abilities from closed-source LLMs to smaller, open-source models.\n###table_5###" + }, + { + "section_id": "4.3", + "parent_section_id": "4", + "section_name": "Results Analysis", + "text": "To comprehensively assess the CoS method, we similarly report the individual contribution of CoS-Planning and CoS-Calling in Table 5 ###reference_###. Our results demonstrate that CoS-Planning and CoS-Calling separately surpass the performance achieved by CoT-based models on all tasks. This validates the model\u2019s proficiency in both stages during CoS and underscores the rationale behind designing CoS-Planning and CoS-Calling to promote effective tool use under the Toolink framework.\n###figure_3### To evaluate the impact of code generation data in CoS-GPT, we compare the LLaMA-7B tuned with or without them. The results in Figure 3 ###reference_### indicate that LLaMA-CoS trained with code generation data achieves higher accuracy, with an average improvement of 1.4%. This validates the necessity of training on code generation together with CoS ability. By incorporating these data points, the model learns to leverage codes as the medium for tool-using more effectively, which ultimately enhances task performance.\n###figure_4### We discover from the results that the raw LLaMA-7B and Alpaca\u2019s performance lags far behind. To provide insights into why they fail to do CoS-Planning and CoS-Calling even with demonstrations, we conduct a detailed error analysis in fig. 4 ###reference_###.\nUpon analyzing the errors made by both models, we identified two primary tendencies: i) the models tend to learn the pattern but often overlook the utility of the tool and the purpose of the task; ii) they frequently exhibit disarray in reasoning and a misalignment between the tool plan and the tool call. These issues contribute significantly to the subpar performance of both models.\nThroughout the experiments, LLaMA-CoS exhibits diverse CoS-Calling patterns. It is capable of sequentially calling different tools to achieve a specific purpose, using tools in a non-linear logic (in a loop or with conditions), or performing nested tool calls, where the output from one tool directly serves as the other tool\u2019s input. These abilities illustrate the robustness and adaptability of LLaMA-CoS across diverse scenarios. We provide case studies and more details in Appendix G ###reference_### and Figure 5 ###reference_###.\n###figure_5### ###figure_6### ###figure_7###" + }, + { + "section_id": "5", + "parent_section_id": null, + "section_name": "Further Studies", + "text": "###table_6### ###table_7### In this section, we show the generalization of LLaMA-CoS to novel tasks and how it can use toolkits that are not specially tailored for solving the target task. These studies aim to illustrate the robustness of LLaMA-CoS in utilizing tools through planning and calling." + }, + { + "section_id": "5.1", + "parent_section_id": "5", + "section_name": "Generalization to Novel Tasks", + "text": "The eight evaluation tasks (Srivastava et al., 2022 ###reference_b36###) we previously used all have data points presented during training, despite only leveraging a few publicly available samples. To showcase the generalization ability of LLaMA-CoS, we further test it on two new tasks: FinQA (Chen et al., 2022b ###reference_b7###) and GSM8K (Cobbe et al., 2021 ###reference_b8###). FinQA involves questions based on financial report data, while GSM8K involves grade school math problems.\nTogether with AQUA-RAT, MATH, and TabMWP, whose data are presented in CoS-GPT (as detailed in \u00a7 3.3 ###reference_###), we randomly select a maximum of 400 test data points from each of the five tasks, ensuring they do not appear in CoS-GPT. We augment each data point with a toolkit, following the method outlined in \u00a7 3.3 ###reference_### regarding how CoS-GPT is constructed. In experiments, we follow the CoS-Planning and CoS-Calling tests outlined in \u00a7 4.1 ###reference_###.\nWe show in Table 6 ###reference_### that LLaMA-CoS achieves high accuracy in both planning and calling stages and could even beat ChatGPT in performance. This affirms the effectiveness and robustness of its CoS ability even applied to unseen tasks. As our tests encompass math, finance, table reasoning, etc, this finding also emphasizes the robustness of LLaMA-CoS across diverse types of tasks." + }, + { + "section_id": "5.2", + "parent_section_id": "5", + "section_name": "CoS on Generic Toolkit", + "text": "We further explore the ability of LLaMA-CoS to use a generic toolkit instead of the one specifically tailored for the target task. In real-world scenarios, toolkits are usually designed to address diverse tasks rather than tailored for a single task. We expect LLaMA-CoS and ChatGPT can apply toolkits borrowed from other tasks to solve the target query in a CoS approach.\nTo validate this, we source two additional tasks from BIG-bench: Dynamic Counting and Unit Interpretation. For each task, we provide a toolkit created explicitly for the target task or borrowed from another task. Specifically, we pair Dynamic Counting and Unit Interpretation respectively with Dyck Language and Arithmetic.\nUnder these settings, we evaluate the performance of both LLaMA-CoS and ChatGPT in Table 7 ###reference_### and show that both LLaMA-CoS and ChatGPT can utilize a generic toolkit borrowed from another task to solve target queries through CoS. Though the performance still lags, these findings nevertheless confirm our assumption that CoS can increase the robustness of tool-using, and make our Toolink more applicable to real-world scenarios. We present more details in Appendix H ###reference_###." + }, + { + "section_id": "6", + "parent_section_id": null, + "section_name": "Conclusions", + "text": "We present Toolink, a tool-training framework that effectively applies toolkits to solve problems leveraging small, open-source language models. Toolink offers increased flexibility in adapting to diverse downstream tasks while addressing concerns related to high inference costs and privacy. Our main contributions include i) empirically implementing a framework that can effectively leverage open-source models\u2019 tool-using ability, ii) devising the chain-of-solving (CoS) method that links toolkit creation and tool use through robust planning and calling, and iii) releasing the CoS-GPT dataset that aims to enhance the model\u2019s CoS capabilities.\nSpecifically, our LLaMA-CoS model outperforms traditional CoT and achieves a comparable performance to ChatGPT concerning tool-using. We believe our study provides a solid foundation for future researchers to explore and enhance the tool-using capabilities of open-source models." + } + ], + "appendix": [ + { + "section_id": "Appendix x1", + "parent_section_id": null, + "section_name": "Appendix", + "text": "" + }, + { + "section_id": "Appendix 1", + "parent_section_id": null, + "section_name": "Appendix A Prompt Pattern for ChatGPT Toolkit", + "text": "We show the pattern of the prompt we apply for the creation of toolkits leveraging GPT-3.5-turbo in Figure 6 ###reference_###. The temperature is set to 0.3 to ensure the model clearly follows the instructions while retaining its creativity to a certain extent. The max length during generation is set to 1024. The prompt shown mainly consists of the instruction for toolkit creation, the demonstration of the format, sample public data, and the task description.\n###figure_8###" + }, + { + "section_id": "Appendix 2", + "parent_section_id": null, + "section_name": "Appendix B Toolkits for tasks from BIG-bench", + "text": "We show in Figures 8 ###reference_###, 9 ###reference_###, 11 ###reference_###, 12 ###reference_###, 10 ###reference_###, 14 ###reference_###, 13 ###reference_### and 15 ###reference_### the toolkits that GPT-3.5-turbo created leveraging the prompt mentioned in the previous section. Notice that we show the final version of the toolkit, which may contain certain modifications based on human feedback. For instance, in Figure 10 ###reference_###, we have integrated addition, subtraction, and hadamard operation into one single tool, as all of them do not change the shape of the given matrix. This will effectively reduce the redundant tools and help the model learn with ease." + }, + { + "section_id": "Appendix 3", + "parent_section_id": null, + "section_name": "Appendix C Settings for Chain-of-Solving", + "text": "To inspire the models\u2019 ability to plan and call the tools during chain-of-solving (CoS), we apply clear instructions to prompt the model. For CoS-Planning, we choose the instruction \"You are presented with a question and several tools that may be useful. Select the useful tools and plan how to solve the problem.\", while for CoS-Calling, we choose the instruction \"Use the tool given in the input to write code to solve the problem.\". This applies to all the settings, including the LLaMA-CoS because it is also tuned in an instruction-following way.\nFor all the experiments leveraging ChatGPT, despite the instructions, we also provide the model with demonstration examples to showcase the format of planning and calling, as well as to better leverage its potential. The temperature is set to 0.3 during generation, and the max output length is set to 1024.\nFor the raw LLaMA-7B and Alpaca baselines without being tuned, the demonstration examples are also applied to provide guidance, while the LLaMA-CoS tuned under our Toolink framework does not need demonstration examples as it is already tuned under the instruction-following paradigm." + }, + { + "section_id": "Appendix 4", + "parent_section_id": null, + "section_name": "Appendix D Separated Test of CoS-Planning and CoS-Calling", + "text": "In Toolink, planning and calling are combined as a whole CoS process, where the plans generated by the model are again fed back to itself to help guide the generation of the final tool calling decision. To disentangle their functions and better understand their role, we employ tests to measure their accuracy separately.\n###figure_9### For the CoS-Planning test, we provide the model with the instructions and all the available tools in the toolkit. In Figure 7 ###reference_###, we showcase the format of the CoS-Planning prompt given to the model.\nHowever, plans are generated in the form of natural language, whose accuracy is hard to measure. For simplicity, we instead only measure if the correct tools are called upon to solve the given problem.\nSuppose is the toolkit with tools for task . For a specific query, we denote the set of useful tools as and other redundant tools as . Suppose the set of tools called upon during planning is , then the correct tools called is denoted as , and the erroneous tools called . These are the exact definitions of the variables that we apply in Equation 1 ###reference_###.\nIf all the useful tools are called correctly and precisely, where , the accuracy will be 1.00. Note that this metric is relatively strict because wrong calls will result in a reduction of accuracy.\nFor the CoS-Calling test, the standard (correct) plans will be provided to the model, instead of the plans that the model previously generated. The CoS-Calling test solely aims to investigate the model\u2019s ability to follow plans and generate the correct calling decisions. Besides the plans and instructions, only the useful tools with respect to the given query are provided in the prompt, instead of all the tools from the toolkit. We showcase the format of the prompt given to the model in Figure 7 ###reference_###.\nThe accuracy of CoS-Calling is based on the matching of the model\u2019s output to the correct answer. For tasks Arithmetic and Chinese Remainder, the accuracy is evaluated in numerical format; for Matrix Shape, the accuracy is evaluated based on the matching of dimensions list; for all other tasks from BIG-bench, the accuracy is based on the matching of strings between the model\u2019s output and the correct answer." + }, + { + "section_id": "Appendix 5", + "parent_section_id": null, + "section_name": "Appendix E Dataset Construction", + "text": "In this section, we provide more details about how CoS-GPT is constructed. We introduce respectively the construction of tool-using data (including planning and calling) and code generation data. All the data points aim to enhance the open-source model\u2019s CoS ability.\nFor each query in AQUA-RAT, GSM8K, and TabMWP, we first utilize ChatGPT to create a diverse set of tools that are potentially relevant to the given query, forming the toolkit. We then provide this toolkit to ChatGPT and allow it to select the most suitable tools. Subsequently, we prompt ChatGPT to generate decision calls based on the selected tools and manually verify the correctness of the resulting outputs. If the final answer is correct, we divide ChatGPT\u2019s responses into two distinct components, representing the planning stage and the calling stage, which are then individually added to the dataset. In this manner, the validity of our data points can thus be guaranteed.\nThroughout these steps of data construction, we also incorporate demonstration examples sampled from the constructed dataset, thereby expanding the dataset in a self-iterative manner. Figure 7 ###reference_### shows detailed information about the format of the query. Besides the query, we also provide the corresponding CoS-Planning or CoS-Calling response and the implementation of the toolkit with useful and redundant tools.\nThe code generation data in CoS-GPT are sourced from 6 different venues, including Python-Simple, Python-Specific, Math, Algorithm, LeetCode, and Rectification. The objective behind these categories is to enhance the model\u2019s proficiency in problem-solving through code utilization, calling existing packages, applying reasoning, employing algorithms, completing codes of challenging competitions, and engaging in self-rectification.\nFor Python-Simple and Python-Specific, the former aims to boost the models\u2019 ability to solve simple problems using codes, while the latter aims to enhance the model\u2019s ability to leverage code packages to solve more complex problems. Both these two sets are generated using ChatGPT. We prompt the model with instructions and demonstrations and gather the code snippets the model generated to solve the given problem.\nThe queries for the Math set are sampled from the training set of MathQA (Amini et al., 2019 ###reference_b1###) and augmented with a code solution based on the given query and reasoning, leveraging ChatGPT. The generated programs are verified to ensure the output answer is the same as the correct one originally, thus ensuring the validity of the augmented data points. The Algorithm set is extracted from the open-source Python algorithm repository, with over 40 categories and more than a hundred diverse algorithms. For each algorithm, we ask ChatGPT to generate a query related to it and use a code snippet to solve the problem. The codes and corresponding queries are then gathered and formed into the instruction-following format.\nFor the LeetCode set, we directly extract the official open-sourced problems and the code answers from the website and form our data. The Rectification set is gathered from the error codes generated in the five sets before. The error tracebacks and the bad code snippet are fed into ChatGPT, and we leverage it to rectify the codes and generate a correct code snippet that can solve the given query successfully. We gather the generated codes and execute them again, retaining only the ones that give a correct answer finally and form the set based on these valid data points.\n###figure_10###" + }, + { + "section_id": "Appendix 6", + "parent_section_id": null, + "section_name": "Appendix F Main Experiment Setting Details", + "text": "For our main experiment, we finetuned the LLaMA-7B model on four A100-80G GPUs, with a total batch size of 32 and a learning rate of 1e-5. For the model whose performance we demonstrate in Table 4 ###reference_### and Table 5 ###reference_###, its training dataset consists of 1.6K target task-specific data points (8 tasks, 100 for planning and 100 for calling each), 4K tool-using data and 3K code-generation data randomly sampled respectively from the CoS-GPT dataset. We trained the LLaMA-7B on these data for 3 epochs and obtained LLaMA-CoS.\nIn addition, for the ablation study about the training on codes we perform in \u00a7 4.3 ###reference_###, we apply 7K tool-using data and remove all the code-generation data points. We keep all the other settings the same in this study." + }, + { + "section_id": "Appendix 7", + "parent_section_id": null, + "section_name": "Appendix G Case Studies of Diverse CoS Patterns", + "text": "In Figure 5 ###reference_###, we present three case studies highlighting the diverse nature of LLaMA-CoS in applying planning and calling for tool-using.\nFirstly, LLaMA-CoS exhibits the ability to generate sequential plans involving different tools. In the first case, the model simulates the operation on matrices step by step in a linear way and finally gets the correct result.\nSecondly, LLaMA-CoS demonstrates proficiency in executing complex tool calls within branch-loop structures. In the second case, the model learns to use different stack operations based on the character met in the expression, and can call the useful tool in a loop structure.\nLastly, the model showcases its competence in performing nested tool invocations. In the third case, the model is able to directly pass the converted hour retrieved from the previous tool as the input parameter for the next tool, which illustrates a successful nested tool call.\nThese examples serve to show the robustness, versatility, and adaptability of LLaMA-CoS across a wide range of scenarios." + }, + { + "section_id": "Appendix 8", + "parent_section_id": null, + "section_name": "Appendix H CoS on Generic Toolkit Details", + "text": "We source two new tasks, Dynamic Counting and Unit Interpretation, from the BIG-bench. We apply all the problems in Dynamic Counting for our test of toolkit generalization. However, for Unit Interpretation, we specifically select the data from LV 1 in order for the tools from task Arithmetic to be properly applied. To ensure fairness, we expand the dataset by interactively sampling new questions with similar patterns from ChatGPT and incorporating them until the dataset reaches its original full size. Note that we only aim to showcase the toolkit\u2019s generalization ability and compare the performance of LLaMA-CoS and ChatGPT within this work, so we deem expanding the dataset as fair and reasonable under our settings.\nWe show the toolkits specially tailored for these two new tasks in Figures 16 ###reference_### and 17 ###reference_###. The LLaMA-CoS model we apply is still the model we have trained in the main experiment, detailed in Appendix F ###reference_###. All the other settings, including the ChatGPT applied under our framework, are kept the same as that in the main experiment.\n###figure_11### ###figure_12### ###figure_13### ###figure_14### ###figure_15### ###figure_16### ###figure_17### ###figure_18### ###figure_19###" + } + ], + "tables": { + "1": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
CategorySet Name\n \n\n\nSource\n \n\n \n\n\nNumber\n \n
\n\n\nTool-Using\n Tool-PlanningAugmented4.4K
Tool-CallingAugmented4.4K
\n\n\nCode\n\nGeneration\n Python-SimpleNew2.0K
Python-SpecificNew2.0K
MathAugmented2.5K
AlgorithmGithub2.3K
LeetCodeLeetCode0.8K
RectificationSources Above1.6K
\n\n\nTotal\n --20.0K
\n
Table 1: The statistics about the sources and number of data points in each category of CoS-GPT. Augmented represents augmented from an existing dataset.
\n
", + "capture": "Table 1: The statistics about the sources and number of data points in each category of CoS-GPT. Augmented represents augmented from an existing dataset." + }, + "2": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
CategoryTask Name
\n\n\nMathematics\n Arithmetic, Matrix Shape,
Chinese Remainder
\n\n\nCommon Sense\n Date Understanding, Navigate
\n\n\nLogical Reasoning\n Dyck Language, Boolean Expression
\n\n\nDecomposition\n Tracking Shuffled Objects
\n
Table 2: The categories of 8 BIG-bench tasks tested.
\n
", + "capture": "Table 2: The categories of 8 BIG-bench tasks tested." + }, + "3": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
TaskArith.\n \n\n\nDate U.\n \n\n \n\n\nMatrix S.\n \n\n \n\n\nNavigate\n \n\n \n\n\nChinese R.\n \n\n \n\n\nDyck L.\n \n\n \n\n\nBoolean E.\n \n\n \n\n\nTracking S.\n \n\n \n\n\nAverage\n \n
\n\n\nNum. of Tools\n 535224243.38
\n\n\nVanilla\n 77.7868.6740.9065.160.019.4080.7023.6747.03
\n\n\nCoT\n 79.4468.6780.4687.960.019.4275.8840.7856.58
\n\n\nCoS\n 100.0069.2893.6785.3095.1452.4697.3799.1186.54
\n\n\nCoS-Planning\n 100.0066.1695.1894.78100.0074.5895.3999.8590.74
\n\n\nCoS-Calling\n 100.0090.9697.4488.4495.6798.5593.42100.0095.56
\n
Table 3: We record the number of tools in the toolkit created for each task and demonstrate the accuracy (%) of ChatGPT on 8 BIG-bench tasks under different settings. We report the results of Vanilla, CoT baselines, and our CoS method, and report the performance of CoS-Planning and CoS-Calling separately.
\n
", + "capture": "Table 3: We record the number of tools in the toolkit created for each task and demonstrate the accuracy (%) of ChatGPT on 8 BIG-bench tasks under different settings. We report the results of Vanilla, CoT baselines, and our CoS method, and report the performance of CoS-Planning and CoS-Calling separately." + }, + "4": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
MethodModelArith.\n \n\n\nDate U.\n \n\n \n\n\nMatrix S.\n \n\n \n\n\nNavigate\n \n\n \n\n\nChinese R.\n \n\n \n\n\nDyck L.\n \n\n \n\n\nBoolean E.\n \n\n \n\n\nTracking S.\n \n
\n\n\nCoT\n\n(Prompting\n\nw/ demo)\n Alpaca19.8939.765.6247.110.00.057.460.44
LLaMA-7B39.4433.7312.5839.700.02.9050.4414.22
ChatGPT79.4468.6780.4687.960.019.4275.8840.78
\n\n\nCoT\n (Tuned)LLaMA-CoT50.4449.4070.8271.640.035.2762.7228.44
\n\n\nCoS\n\n(Prompting\n\nw/ demo)\n Alpaca17.787.833.0048.607.561.0094.746.78
LLaMA-7B55.8917.4710.6545.9023.8035.8399.120.67
ChatGPT100.0069.2893.6785.3095.1452.4697.3799.11
\n\n\nCoS\n (Tuned)LLaMA-CoS100.0074.1091.0199.5695.4498.21100.0099.56
\n
Table 4: The accuracy (%) of baselines and LLaMA-CoS on the 8 BIG-bench tasks. LLaMA-CoS employs planning and calling of tools, which beats all CoT baselines by large margins and is on par with ChatGPT\u2019s CoS ability.
\n
", + "capture": "Table 4: The accuracy (%) of baselines and LLaMA-CoS on the 8 BIG-bench tasks. LLaMA-CoS employs planning and calling of tools, which beats all CoT baselines by large margins and is on par with ChatGPT\u2019s CoS ability." + }, + "5": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
MethodModelArith.\n \n\n\nDate U.\n \n\n \n\n\nMatrix S.\n \n\n \n\n\nNavigate\n \n\n \n\n\nChinese R.\n \n\n \n\n\nDyck L.\n \n\n \n\n\nBoolean E.\n \n\n \n\n\nTracking S.\n \n
\n\n\nCoS-Whole\n LLaMA-CoS100.0074.1091.0199.5695.4498.21100.0099.56
\n\n\nCoS-Planning\n\n(Prompting\n\nw/ demo)\n Alpaca18.2227.4124.1577.16100.0076.397.5999.37
LLaMA-7B74.1127.7125.0277.16100.0093.8097.59100.00
ChatGPT100.0066.1695.1894.78100.0074.5895.3999.85
\n\n\nCoS-Planning\n LLaMA-CoS100.0085.8489.6297.14100.0099.1997.59100.00
\n\n\nCoS-Calling\n\n(Prompting\n\nw/ demo)\n Alpaca99.4424.7030.0848.6017.971.5689.916.78
LLaMA-7B74.7051.2055.4943.7724.8125.6794.301.56
ChatGPT100.0090.9697.4488.4495.6798.5593.42100.00
\n\n\nCoS-Calling\n LLaMA-CoS100.0091.5795.5698.8894.1898.5595.6188.44
\n
Table 5: The accuracy (%) of CoS-Planning and CoS-Calling separately on 8 BIG-bench tasks. Results show LLaMA-CoS has excellent ability in understanding and using tools through CoS.
\n
", + "capture": "Table 5: The accuracy (%) of CoS-Planning and CoS-Calling separately on 8 BIG-bench tasks. Results show LLaMA-CoS has excellent ability in understanding and using tools through CoS." + }, + "6": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Task\n \n\n\nIf in\n\nCoS-GPT\n \n\n \n\n\nCoS\n\nStage\n \n\n \n\n\nLLaMA\n\n-CoS\n \n\n \n\n\nChatGPT\n \n
\n\n\nAQUA-RAT\n \\usym2713Planning58.8052.90
Calling56.1265.94
\n\n\nMATH\n \\usym2713Planning65.8352.61
Calling50.7543.25
\n\n\nTabMWP\n \\usym2713Planning90.0060.75
Calling66.0032.75
\n\n\nFinQA\n \\usym2717Planning70.5150.15
Calling22.3832.38
\n\n\nGSM8K\n \\usym2717Planning61.2953.83
Calling57.2536.50
\n
Table 6: The accuracy (%) of CoS-Planning and CoS-Calling on five diverse datasets applying LLaMA-CoS or ChatGPT. Results show that LLaMA-CoS generally beats ChatGPT and is robust to unseen tasks.
\n
", + "capture": "Table 6: The accuracy (%) of CoS-Planning and CoS-Calling on five diverse datasets applying LLaMA-CoS or ChatGPT. Results show that LLaMA-CoS generally beats ChatGPT and is robust to unseen tasks." + }, + "7": { + "table_html": "
\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Task\n \n\n\nToolkit Origin\n \n\n \n\n\nLLaMA-CoS\n \n\n \n\n\nChatGPT\n \n
\n\n\nDynamic\n\nCnt.\n Raw97.5080.83
From Dyck L.73.3079.17
\n\n\nUnit\n\nInterp.\n Raw70.8380.83
From Arith.65.8380.00
\n
Table 7: The accuracy (%) of ChatGPT and LLaMA-CoS, with toolkit newly created for the target task (Raw) or borrowed from other tasks. Our results show that both ChatGPT and LLaMA-CoS can utilize tools not specifically tailored for the target task through CoS.
\n
", + "capture": "Table 7: The accuracy (%) of ChatGPT and LLaMA-CoS, with toolkit newly created for the target task (Raw) or borrowed from other tasks. Our results show that both ChatGPT and LLaMA-CoS can utilize tools not specifically tailored for the target task through CoS." + } + }, + "image_paths": { + "1": { + "figure_path": "2310.05155v2_figure_1.png", + "caption": "Figure 1: An illustration of Toolink, which decomposes tasks via toolkit creation and resolves queries through Chain-of-Solving (CoS). Toolink can be adapted to open-source LLaMA for enhanced tool usage.", + "url": "http://arxiv.org/html/2310.05155v2/x1.png" + }, + "2": { + "figure_path": "2310.05155v2_figure_2.png", + "caption": "Figure 2: A problem solving chain of Toolink pipeline. We show an example from task Navigate. Toolink first creates a toolkit generally applicable to the task, and then approaches the specific query through CoS, which involves planning and calling of the created tools.", + "url": "http://arxiv.org/html/2310.05155v2/x2.png" + }, + "3": { + "figure_path": "2310.05155v2_figure_3.png", + "caption": "Figure 3: The improvement of performance when code generation data points are involved during training.", + "url": "http://arxiv.org/html/2310.05155v2/x3.png" + }, + "4": { + "figure_path": "2310.05155v2_figure_4.png", + "caption": "Figure 4: Detailed error analysis of Alpaca and LLaMA-7B regarding CoS-Planning and CoS-Calling.", + "url": "http://arxiv.org/html/2310.05155v2/x4.png" + }, + "5(a)": { + "figure_path": "2310.05155v2_figure_5(a).png", + "caption": "(a) Sequential Tool Calling.\nFigure 5: Case Studies on the diverse CoS-Calling patterns in the main experiment.", + "url": "http://arxiv.org/html/2310.05155v2/x5.png" + }, + "5(b)": { + "figure_path": "2310.05155v2_figure_5(b).png", + "caption": "(b) Conditional Tool Calling.\nFigure 5: Case Studies on the diverse CoS-Calling patterns in the main experiment.", + "url": "http://arxiv.org/html/2310.05155v2/x6.png" + }, + "5(c)": { + "figure_path": "2310.05155v2_figure_5(c).png", + "caption": "(c) Nested Tool Calling.\nFigure 5: Case Studies on the diverse CoS-Calling patterns in the main experiment.", + "url": "http://arxiv.org/html/2310.05155v2/x7.png" + }, + "6": { + "figure_path": "2310.05155v2_figure_6.png", + "caption": "Figure 6: The pattern of the prompt given to GPT-3.5-turbo to generate the toolkit.", + "url": "http://arxiv.org/html/2310.05155v2/x8.png" + }, + "7": { + "figure_path": "2310.05155v2_figure_7.png", + "caption": "Figure 7: The format of the data (and prompt) for CoS-Planning and CoS-Calling.", + "url": "http://arxiv.org/html/2310.05155v2/x9.png" + }, + "8": { + "figure_path": "2310.05155v2_figure_8.png", + "caption": "Figure 8: The toolkit for task Arithmetic.", + "url": "http://arxiv.org/html/2310.05155v2/x10.png" + }, + "9": { + "figure_path": "2310.05155v2_figure_9.png", + "caption": "Figure 9: The toolkit for task Date Understanding.", + "url": "http://arxiv.org/html/2310.05155v2/x11.png" + }, + "10": { + "figure_path": "2310.05155v2_figure_10.png", + "caption": "Figure 10: The toolkit for task Matrix Shape.", + "url": "http://arxiv.org/html/2310.05155v2/x12.png" + }, + "11": { + "figure_path": "2310.05155v2_figure_11.png", + "caption": "Figure 11: The toolkit for task Navigation.", + "url": "http://arxiv.org/html/2310.05155v2/x13.png" + }, + "12": { + "figure_path": "2310.05155v2_figure_12.png", + "caption": "Figure 12: The toolkit for task Chinese Remainder.", + "url": "http://arxiv.org/html/2310.05155v2/x14.png" + }, + "13": { + "figure_path": "2310.05155v2_figure_13.png", + "caption": "Figure 13: The toolkit for task Dyck Language.", + "url": "http://arxiv.org/html/2310.05155v2/x15.png" + }, + "14": { + "figure_path": "2310.05155v2_figure_14.png", + "caption": "Figure 14: The toolkit for task Boolean Expression.", + "url": "http://arxiv.org/html/2310.05155v2/x16.png" + }, + "15": { + "figure_path": "2310.05155v2_figure_15.png", + "caption": "Figure 15: The toolkit for task Tracking Shuffled Objects.", + "url": "http://arxiv.org/html/2310.05155v2/x17.png" + }, + "16": { + "figure_path": "2310.05155v2_figure_16.png", + "caption": "Figure 16: The toolkit for task Dynamic Counting.", + "url": "http://arxiv.org/html/2310.05155v2/x18.png" + }, + "17": { + "figure_path": "2310.05155v2_figure_17.png", + "caption": "Figure 17: The toolkit for task Unit Interpretation (LV 1).", + "url": "http://arxiv.org/html/2310.05155v2/x19.png" + } + }, + "validation": true, + "references": [ + { + "1": { + "title": "Mathqa: Towards interpretable math word problem solving with operation-based formalisms.", + "author": "Aida Amini, Saadia Gabriel, Peter Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019.", + "venue": null, + "url": "http://arxiv.org/abs/1905.13319" + } + }, + { + "2": { + "title": "Improving language models by retrieving from trillions of tokens.", + "author": "Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. 2022.", + "venue": "In International conference on machine learning, pages 2206\u20132240. PMLR.", + "url": null + } + }, + { + "3": { + "title": "Large language models as tool makers.", + "author": "Tianle Cai, Xuezhi Wang, Tengyu Ma, Xinyun Chen, and Denny Zhou. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2305.17126" + } + }, + { + "4": { + "title": "Extracting training data from large language models.", + "author": "Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Colin Raffel. 2021.", + "venue": null, + "url": "http://arxiv.org/abs/2012.07805" + } + }, + { + "5": { + "title": "Evaluating large language models trained on code.", + "author": "Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. 2021.", + "venue": "arXiv preprint arXiv:2107.03374.", + "url": null + } + }, + { + "6": { + "title": "Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks.", + "author": "Wenhu Chen, Xueguang Ma, Xinyi Wang, and William W Cohen. 2022a.", + "venue": "arXiv preprint arXiv:2211.12588.", + "url": null + } + }, + { + "7": { + "title": "Finqa: A dataset of numerical reasoning over financial data.", + "author": "Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. 2022b.", + "venue": null, + "url": "http://arxiv.org/abs/2109.00122" + } + }, + { + "8": { + "title": "Training verifiers to solve math word problems.", + "author": "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021.", + "venue": null, + "url": "http://arxiv.org/abs/2110.14168" + } + }, + { + "9": { + "title": "Pal: Program-aided language models.", + "author": "Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2022.", + "venue": "arXiv preprint arXiv:2211.10435.", + "url": null + } + }, + { + "10": { + "title": "Parameter-efficient transfer learning for nlp.", + "author": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019.", + "venue": "In International Conference on Machine Learning, pages 2790\u20132799. PMLR.", + "url": null + } + }, + { + "11": { + "title": "Lora: Low-rank adaptation of large language models.", + "author": "Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021.", + "venue": null, + "url": "http://arxiv.org/abs/2106.09685" + } + }, + { + "12": { + "title": "Generalization through memorization: Nearest neighbor language models.", + "author": "Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis.", + "venue": "In International Conference on Learning Representations.", + "url": null + } + }, + { + "13": { + "title": "Taskmatrix. ai: Completing tasks by connecting foundation models with millions of apis.", + "author": "Yaobo Liang, Chenfei Wu, Ting Song, Wenshan Wu, Yan Xia, Yu Liu, Yang Ou, Shuai Lu, Lei Ji, Shaoguang Mao, et al. 2023.", + "venue": "arXiv preprint arXiv:2303.16434.", + "url": null + } + }, + { + "14": { + "title": "Program induction by rationale generation: Learning to solve and explain algebraic word problems.", + "author": "Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. 2017.", + "venue": "In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 158\u2013167.", + "url": null + } + }, + { + "15": { + "title": "Visual instruction tuning.", + "author": "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023.", + "venue": "arXiv preprint arXiv:2304.08485.", + "url": null + } + }, + { + "16": { + "title": "Goat: Fine-tuned llama outperforms gpt-4 on arithmetic tasks.", + "author": "Tiedong Liu and Bryan Kian Hsiang Low. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2305.14201" + } + }, + { + "17": { + "title": "The flan collection: Designing data and methods for effective instruction tuning.", + "author": "Shayne Longpre, Le Hou, Tu Vu, Albert Webson, Hyung Won Chung, Yi Tay, Denny Zhou, Quoc V Le, Barret Zoph, Jason Wei, et al. 2023.", + "venue": "arXiv preprint arXiv:2301.13688.", + "url": null + } + }, + { + "18": { + "title": "Chameleon: Plug-and-play compositional reasoning with large language models.", + "author": "Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. 2023.", + "venue": "arXiv preprint arXiv:2304.09842.", + "url": null + } + }, + { + "19": { + "title": "Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning.", + "author": "Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. 2022a.", + "venue": "arXiv preprint arXiv:2209.14610.", + "url": null + } + }, + { + "20": { + "title": "A survey of deep learning for mathematical reasoning.", + "author": "Pan Lu, Liang Qiu, Wenhao Yu, Sean Welleck, and Kai-Wei Chang. 2022b.", + "venue": "arXiv preprint arXiv:2212.10535.", + "url": null + } + }, + { + "21": { + "title": "Show your work: Scratchpads for intermediate computation with language models.", + "author": "Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. 2021.", + "venue": "arXiv preprint arXiv:2112.00114.", + "url": null + } + }, + { + "22": { + "title": "Chatgpt.", + "author": "OpenAI. 2022.", + "venue": null, + "url": null + } + }, + { + "23": { + "title": "Gpt-4 technical report.", + "author": "OpenAI. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2303.08774" + } + }, + { + "24": { + "title": "Training language models to follow instructions with human feedback.", + "author": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022.", + "venue": "Advances in Neural Information Processing Systems, 35:27730\u201327744.", + "url": null + } + }, + { + "25": { + "title": "Art: Automatic multi-step reasoning and tool-use for large language models.", + "author": "Bhargavi Paranjape, Scott Lundberg, Sameer Singh, Hannaneh Hajishirzi, Luke Zettlemoyer, and Marco Tulio Ribeiro. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2303.09014" + } + }, + { + "26": { + "title": "Talm: Tool augmented language models.", + "author": "Aaron Parisi, Yao Zhao, and Noah Fiedel. 2022.", + "venue": null, + "url": "http://arxiv.org/abs/2205.12255" + } + }, + { + "27": { + "title": "Are nlp models really able to solve simple math word problems?", + "author": "Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021.", + "venue": "In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2080\u20132094.", + "url": null + } + }, + { + "28": { + "title": "Instruction tuning with gpt-4.", + "author": "Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2304.03277" + } + }, + { + "29": { + "title": "Adapterhub: A framework for adapting transformers.", + "author": "Jonas Pfeiffer, Andreas R\u00fcckl\u00e9, Clifton Poth, Aishwarya Kamath, Ivan Vuli\u0107, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020.", + "venue": "In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46\u201354.", + "url": null + } + }, + { + "30": { + "title": "Creator: Disentangling abstract and concrete reasonings of large language models through tool creation.", + "author": "Cheng Qian, Chi Han, Yi R. Fung, Yujia Qin, Zhiyuan Liu, and Heng Ji. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2305.14318" + } + }, + { + "31": { + "title": "Toolllm: Facilitating large language models to master 16000+ real-world apis.", + "author": "Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2307.16789" + } + }, + { + "32": { + "title": "Exploring the limits of transfer learning with a unified text-to-text transformer.", + "author": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020.", + "venue": null, + "url": "http://arxiv.org/abs/1910.10683" + } + }, + { + "33": { + "title": "Toolformer: Language models can teach themselves to use tools.", + "author": "Timo Schick, Jane Dwivedi-Yu, Roberto Dess\u00ec, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023.", + "venue": "arXiv preprint arXiv:2302.04761.", + "url": null + } + }, + { + "34": { + "title": "Hugginggpt: Solving ai tasks with chatgpt and its friends in huggingface.", + "author": "Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023.", + "venue": "arXiv preprint arXiv:2303.17580.", + "url": null + } + }, + { + "35": { + "title": "Knn-diffusion: Image generation via large-scale retrieval.", + "author": "Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. 2022.", + "venue": "arXiv preprint arXiv:2204.02849.", + "url": null + } + }, + { + "36": { + "title": "Beyond the imitation game: Quantifying and extrapolating the capabilities of language models.", + "author": "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adri\u00e0 Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshat Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlm\u00fcller, Andrew Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut Erdem, Ayla Karaka\u015f, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bart\u0142omiej Bojanowski, Batuhan \u00d6zyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald,\nCameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, C\u00e9sar Ferri Ram\u00edrez, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Chris Waites, Christian Voigt, Christopher D. Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Mosegu\u00ed Gonz\u00e1lez, Danielle Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David Drakard, David Jurgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho Mollo, Diyi Yang, Dong-Ho Lee, Ekaterina Shutova, Ekin Dogus Cubuk, Elad Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodola, Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A. Chi, Ethan Dyer, Ethan\nJerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar, Fernando Mart\u00ednez-Plumed, Francesca Happ\u00e9, Francois Chollet, Frieda Rong, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, Germ\u00e1n Kruszewski, Giambattista Parascandolo, Giorgio Mariani, Gloria Wang, Gonzalo Jaimovitch-L\u00f3pez, Gregor Betz, Guy Gur-Ari, Hana Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar, Henry Shevlin, Hinrich Sch\u00fctze, Hiromu Yakura, Hongming Zhang, Hugh Mee Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jaehoon Lee, Jaime Fern\u00e1ndez Fisac, James B. Simon, James Koppel, James Zheng, James Zou, Jan Koco\u0144, Jana Thompson, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason Phang, Jason Wei, Jason Yosinski, Jekaterina Novikova, Jelle Bosscher, Jennifer Marsh, Jeremy Kim, Jeroen Taal, Jesse Engel, Jesujoba Alabi, Jiacheng Xu, Jiaming Song, Jillian Tang, Joan Waweru, John Burden, John Miller, John U. Balis,\nJonathan Berant, J\u00f6rg Frohberg, Jos Rozen, Jose Hernandez-Orallo, Joseph Boudeman, Joseph Jones, Joshua B. Tenenbaum, Joshua S. Rule, Joyce Chua, Kamil Kanclerz, Karen Livescu, Karl Krauth, Karthik Gopalakrishnan, Katerina Ignatyeva, Katja Markert, Kaustubh D. Dhole, Kevin Gimpel, Kevin Omondi, Kory Mathewson, Kristen Chiafullo, Ksenia Shkaruta, Kumar Shridhar, Kyle McDonell, Kyle Richardson, Laria Reynolds, Leo Gao, Li Zhang, Liam Dugan, Lianhui Qin, Lidia Contreras-Ochando, Louis-Philippe Morency, Luca Moschella, Lucas Lam, Lucy Noble, Ludwig Schmidt, Luheng He, Luis Oliveros Col\u00f3n, Luke Metz, L\u00fctfi Kerem \u015eenel, Maarten Bosma, Maarten Sap, Maartje ter Hoeve, Maheen Farooqi, Manaal Faruqui, Mantas Mazeika, Marco Baturan, Marco Marelli, Marco Maru, Maria Jose Ram\u00edrez Quintana, Marie Tolkiehn, Mario Giulianelli, Martha Lewis, Martin Potthast, Matthew L. Leavitt, Matthias Hagen, M\u00e1ty\u00e1s Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael A. Yee, Michael Cohen, Michael Gu,\nMichael Ivanitskiy, Michael Starritt, Michael Strube, Micha\u0142 Sw\u0119drowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Mimee Xu, Mirac Suzgun, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini, Mukund Varma T, Nanyun Peng, Nathan Chi, Nayeon Lee, Neta Gur-Ari Krakover, Nicholas Cameron, Nicholas Roberts, Nick Doiron, Nikita Nangia, Niklas Deckers, Niklas Muennighoff, Nitish Shirish Keskar, Niveditha S. Iyer, Noah Constant, Noah Fiedel, Nuan Wen, Oliver Zhang, Omar Agha, Omar Elbaghdadi, Omer Levy, Owain Evans, Pablo Antonio Moreno Casares, Parth Doshi, Pascale Fung, Paul Pu Liang, Paul Vicol, Pegah Alipoormolabashi, Peiyuan Liao, Percy Liang, Peter Chang, Peter Eckersley, Phu Mon Htut, Pinyu Hwang, Piotr Mi\u0142kowski, Piyush Patil, Pouya Pezeshkpour, Priti Oli, Qiaozhu Mei, Qing Lyu, Qinlang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahel Habacker, Ram\u00f3n Risco Delgado, Rapha\u00ebl Milli\u00e8re, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe\nRaymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan LeBras, Rosanne Liu, Rowan Jacobs, Rui Zhang, Ruslan Salakhutdinov, Ryan Chi, Ryan Lee, Ryan Stovall, Ryan Teehan, Rylan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou, Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel S. Schoenholz, Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey, Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan, Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane Gu, Shubh Pachchigar, Shubham Toshniwal, Shyam Upadhyay, Shyamolima, Debnath, Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-Hwan Lee, Spencer Torene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon, Stella Biderman, Stephanie Lin, Stephen Prasad, Steven T. Piantadosi, Stuart M. Shieber, Summer Misherghi, Svetlana Kiritchenko, Swaroop Mishra, Tal\nLinzen, Tal Schuster, Tao Li, Tao Yu, Tariq Ali, Tatsu Hashimoto, Te-Lin Wu, Th\u00e9o Desbordes, Theodore Rothschild, Thomas Phan, Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Timothy Telleen-Lawton, Titus Tunduny, Tobias Gerstenberg, Trenton Chang, Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera Demberg, Victoria Nyamai, Vikas Raunak, Vinay Ramasesh, Vinay Uday Prabhu, Vishakh Padmakumar, Vivek Srikumar, William Fedus, William Saunders, William Zhang, Wout Vossen, Xiang Ren, Xiaoyu Tong, Xinran Zhao, Xinyi Wu, Xudong Shen, Yadollah Yaghoobzadeh, Yair Lakretz, Yangqiu Song, Yasaman Bahri, Yejin Choi, Yichi Yang, Yiding Hao, Yifu Chen, Yonatan Belinkov, Yu Hou, Yufang Hou, Yuntao Bai, Zachary Seid, Zhuoye Zhao, Zijian Wang, Zijie J. Wang, Zirui Wang, and Ziyi Wu. 2022.", + "venue": null, + "url": "http://arxiv.org/abs/2206.04615" + } + }, + { + "37": { + "title": "Stanford alpaca: An instruction-following llama model.", + "author": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023.", + "venue": "https://github.com/tatsu-lab/stanford_alpaca.", + "url": null + } + }, + { + "38": { + "title": "Lamda: Language models for dialog applications.", + "author": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. 2022.", + "venue": null, + "url": "http://arxiv.org/abs/2201.08239" + } + }, + { + "39": { + "title": "Llama: Open and efficient foundation language models.", + "author": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth\u00e9e Lacroix, Baptiste Rozi\u00e8re, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2302.13971" + } + }, + { + "40": { + "title": "Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.", + "author": "Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022.", + "venue": "arXiv preprint arXiv:2212.10509.", + "url": null + } + }, + { + "41": { + "title": "Chain-of-thought prompting elicits reasoning in large language models.", + "author": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed H Chi, Quoc V Le, Denny Zhou, et al. 2022.", + "venue": "In Advances in Neural Information Processing Systems.", + "url": null + } + }, + { + "42": { + "title": "Visual chatgpt: Talking, drawing and editing with visual foundation models.", + "author": "Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. 2023.", + "venue": "arXiv preprint arXiv:2303.04671.", + "url": null + } + }, + { + "43": { + "title": "Agenttuning: Enabling generalized agent abilities for llms.", + "author": "Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang. 2023.", + "venue": null, + "url": "http://arxiv.org/abs/2310.12823" + } + }, + { + "44": { + "title": "Llama-adapter: Efficient fine-tuning of language models with zero-init attention.", + "author": "Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. 2023.", + "venue": "arXiv preprint arXiv:2303.16199.", + "url": null + } + } + ], + "url": "http://arxiv.org/html/2310.05155v2" +} \ No newline at end of file