pith. machine review for the scientific record. sign in

arxiv: 2604.05387 · v1 · submitted 2026-04-07 · 💻 cs.IR · cs.CL

Recognition: no theorem link

Data-Driven Function Calling Improvements in Large Language Model for Online Financial QA

Dugang Liu, Fuyuan Lyu, Hao Chen, Lingjie Li, Shiwei Li, Weihong Luo, Weijie Shi, Xiku Du, Xing Tang, Xiuqiang He

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:21 UTC · model grok-4.3

classification 💻 cs.IR cs.CL
keywords function callinglarge language modelsfinancial question answeringdata augmentationAPI integrationonline deploymentdataset construction
0
0 comments X

The pith

A periodically updated dataset with parameter augmentation and two-step training lets large language models call financial APIs more reliably.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out a practical method to adapt large language models so they can use specialized financial tools when answering user questions in an online system. Generic models often pick the wrong tool or fail on queries that contain unexpected parameter values. The approach builds a dataset from earlier work, refreshes it regularly with fresh user questions, adds an augmentation step that generates varied parameter examples, and trains the model in two stages to decide when and how to call the right functions. Tests on standard collections and live online traffic show the method outperforms earlier versions. The resulting system now runs in production for financial question answering.

Core claim

The authors claim that a data-driven pipeline of periodic dataset construction, incorporation of user queries, augmentation that explores possible parameter values, and two-step training enables an LLM to exploit a set of financial functions, with the resulting gains visible both in offline dataset tests and in actual online deployment.

What carries the argument

The pipeline of periodic dataset updates that incorporate the AugFC augmentation method, followed by two-step training on function-calling examples.

If this is right

  • The updated dataset lets the model draw on the financial toolset through data rather than manual rules.
  • AugFC increases the range of parameter values the model sees, reducing errors on unusual inputs.
  • The two-step training process produces a model that can decide both which function to call and how to format its inputs.
  • The same pipeline has been placed into live operation for an online financial question-answering service.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same maintenance loop of periodic updates plus augmentation could be tried in other domains that rely on many private APIs and variable user phrasing.
  • Because the method depends on ongoing collection of real queries, it implies that production systems will need continuous data pipelines rather than one-time training.
  • The two-step training may serve as a lighter alternative to full retraining when new functions are added to the toolset.

Load-bearing premise

The assumption that regularly adding real user queries and generating extra parameter examples will prepare the model to handle every kind of financial question that appears online.

What would settle it

A clear rise in wrong function selections or outright failures when the deployed system meets user questions whose parameter values fall outside the examples seen during training.

Figures

Figures reproduced from arXiv: 2604.05387 by Dugang Liu, Fuyuan Lyu, Hao Chen, Lingjie Li, Shiwei Li, Weihong Luo, Weijie Shi, Xiku Du, Xing Tang, Xiuqiang He.

Figure 1
Figure 1. Figure 1: The overview of the key points and corresponding challenges in the online financial QA system powered by LLM. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The data-driven pipeline consisting of data collection, data construction, data augmentation, and model training. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The prompt template for prompt isolation. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The number of blind spots to repair with different [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The heatmap illustrating how 𝜏𝑔 and 𝜏𝑏 affects the performance resepectively. Referring to Definition 1, two key hyperparameters 𝜏𝑔 and 𝜏𝑏 determine the blind spot together. Therefore, we need to investigate how these two hyperparameters affect our pipeline. We do a grid search on two hyperparameters, where 𝜏𝑔 ∈ {1.0, 1.5, 2.0, 2.5, 3.0} and 𝜏𝑏 ∈ {0.05, 0.1, 0.15, 0.2, 0.25} for each size of base model. Th… view at source ↗
Figure 6
Figure 6. Figure 6: The overview illustration of the online financial QA system. Our pipeline is marked as red in the whole system. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The online results measured by four metrics. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

Large language models (LLMs) have been incorporated into numerous industrial applications. Meanwhile, a vast array of API assets is scattered across various functions in the financial domain. An online financial question-answering system can leverage both LLMs and private APIs to provide timely financial analysis and information. The key is equipping the LLM model with function calling capability tailored to a financial scenario. However, a generic LLM requires customized financial APIs to call and struggles to adapt to the financial domain. Additionally, online user queries are diverse and contain out-of-distribution parameters compared with the required function input parameters, which makes it more difficult for a generic LLM to serve online users. In this paper, we propose a data-driven pipeline to enhance function calling in LLM for our online, deployed financial QA, comprising dataset construction, data augmentation, and model training. Specifically, we construct a dataset based on a previous study and update it periodically, incorporating queries and an augmentation method named AugFC. The addition of user query-related samples will \textit{exploit} our financial toolset in a data-driven manner, and AugFC explores the possible parameter values to enhance the diversity of our updated dataset. Then, we train an LLM with a two-step method, which enables the use of our financial functions. Extensive experiments on existing offline datasets, as well as the deployment of an online scenario, illustrate the superiority of our pipeline. The related pipeline has been adopted in the financial QA of YuanBao\footnote{https://yuanbao.tencent.com/chat/}, one of the largest chat platforms in China.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a data-driven pipeline to enhance function calling in LLMs for an online financial QA system. The pipeline consists of periodic dataset construction based on a prior study, augmentation with the AugFC method to explore parameter values and increase diversity (including OOD cases), and two-step model training to enable use of financial functions. The authors claim this yields superior performance on existing offline datasets and in a live online deployment, with the pipeline now adopted in the YuanBao financial QA platform.

Significance. If the empirical claims are substantiated, the work offers a practical, deployable approach for adapting LLMs to private financial APIs while handling diverse and out-of-distribution user queries through data augmentation and periodic updates. The reported production adoption in YuanBao constitutes a concrete strength, demonstrating feasibility in a high-stakes industrial setting.

major comments (2)
  1. [Abstract] Abstract: the superiority claim is asserted on the basis of 'extensive experiments on existing offline datasets, as well as the deployment of an online scenario,' yet no metrics, baselines, ablation results, or quantitative handling of out-of-distribution parameters are supplied. This prevents evaluation of the data-to-claim link.
  2. [Pipeline description] Dataset construction and augmentation section (presumably §3): the central premise that the periodically updated dataset plus AugFC augmentation sufficiently covers the diversity and out-of-distribution parameters of real online user queries is stated at a high level but is not supported by distribution statistics, coverage metrics, or failure-mode analysis comparing the augmented data against logged production queries. This coverage assumption is load-bearing for both the two-step training and the offline/online superiority claims.
minor comments (2)
  1. [Abstract / Dataset construction] The reference to 'a previous study' for the base dataset is not accompanied by a citation; add the appropriate reference.
  2. [Augmentation method] Clarify the exact definition and implementation of AugFC (e.g., via pseudocode or parameter sampling procedure) to improve reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address the two major comments point by point below. Where the comments correctly identify gaps in quantitative support, we agree to revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the superiority claim is asserted on the basis of 'extensive experiments on existing offline datasets, as well as the deployment of an online scenario,' yet no metrics, baselines, ablation results, or quantitative handling of out-of-distribution parameters are supplied. This prevents evaluation of the data-to-claim link.

    Authors: We agree that the abstract would benefit from explicit quantitative anchors. In the revised manuscript we will insert the key offline metrics (e.g., function-calling accuracy and parameter-match F1 versus the strongest baselines), a one-sentence ablation summary, and a brief statement on how AugFC increases coverage of OOD parameter values. These numbers are already present in §4 and §5; moving a concise subset into the abstract will directly address the data-to-claim concern without altering the claims themselves. revision: yes

  2. Referee: [Pipeline description] Dataset construction and augmentation section (presumably §3): the central premise that the periodically updated dataset plus AugFC augmentation sufficiently covers the diversity and out-of-distribution parameters of real online user queries is stated at a high level but is not supported by distribution statistics, coverage metrics, or failure-mode analysis comparing the augmented data against logged production queries. This coverage assumption is load-bearing for both the two-step training and the offline/online superiority claims.

    Authors: The referee correctly notes that §3 currently describes the construction and AugFC procedure at a high level without supporting statistics. We will add a new paragraph (or short subsection) that reports (i) parameter-value coverage before and after augmentation, (ii) diversity metrics such as unique parameter combinations and entropy, and (iii) a high-level comparison against a sample of anonymized production query logs. Because raw production logs are subject to privacy constraints, the comparison will use aggregated statistics rather than raw failure cases; we believe this still substantiates the coverage claim while remaining compliant. revision: yes

Circularity Check

0 steps flagged

No circularity in empirical pipeline construction

full rationale

The paper describes a practical data-driven pipeline for enhancing LLM function calling in financial QA: periodic dataset updates from a prior study, AugFC augmentation for parameter diversity, and two-step training. No mathematical derivations, equations, or fitted parameters are presented that reduce to the inputs by construction. Claims of superiority rest on external experiments using offline datasets and online deployment results, which are independent of the construction process. Any reference to a previous study for initial dataset construction does not create circularity, as the method adds updates and augmentation explicitly and evaluates outcomes against separate benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only view prevents enumeration of specific free parameters or axioms; the approach implicitly assumes that real user queries plus synthetic parameter augmentation will generalize to production traffic.

pith-pipeline@v0.9.0 · 5610 in / 1070 out tokens · 46734 ms · 2026-05-10T19:21:11.738342+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 13 canonical work pages · 6 internal anchors

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Lada A Adamic and Bernardo A Huberman. 2000. Power-law distribution of the world wide web.science287, 5461 (2000), 2115–2115

    2000

  3. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners.Advances in neural information processing systems33 (2020), 1877–1901

    2020

  5. [5]

    Jian Chen, Peilin Zhou, Yining Hua, Loh Xin, Kehui Chen, Ziyuan Li, Bing Zhu, and Junwei Liang. 2024. FinTextQA: A Dataset for Long-form Financial Question Answering. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 6025–6047

    2024

  6. [6]

    Guanting Dong, Yifei Chen, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Yutao Zhu, Hangyu Mao, Guorui Zhou, Zhicheng Dou, and Ji-Rong Wen. 2025. Tool-Star: Empowering LLM-Brained Multi-Tool Reasoner via Reinforcement Learning.arXiv preprint arXiv:2505.16410(2025)

    work page arXiv 2025

  7. [7]

    Sorouralsadat Fatemi and Yuheng Hu. 2024. Enhancing Financial Question Answering with a Multi-Agent Reflection Framework. InProceedings of the 5th ACM International Conference on AI in Finance(Brooklyn, NY, USA)(ICAIF ’24). Association for Computing Machinery, 530–537

    2024

  8. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al . 2025. DeepSeek-R1 in- centivizes reasoning in LLMs through reinforcement learning.Nature645, 8081 (2025), 633–638

    2025

  9. [9]

    Bingguang Hao, Maolin Wang, Zengzhuang Xu, Yicheng Chen, Cunyin Peng, Jinjie Gu, and Chenyi Zhuang. 2025. Exploring Superior Function Calls via Reinforcement Learning.CoRRabs/2508.05118 (2025). https://doi.org/10.48550/ arXiv.2508.05118

    work page arXiv 2025

  10. [10]

    Bingguang Hao, Maolin Wang, Zengzhuang Xu, Cunyin Peng, Yicheng Chen, Xiangyu Zhao, Jinjie Gu, and Chenyi Zhuang. 2025. FunReason: Enhancing Large Language Models’ Function Calling via Self-Refinement Multiscale Loss and Automated Data Refinement.arXiv preprint arXiv:2505.20192(2025)

    work page arXiv 2025

  11. [11]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.ICLR1, 2 (2022), 3

    2022

  12. [12]

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. 2023. API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Singapore, 3102–3116

    2023

  13. [13]

    Qiqiang Lin, Muning Wen, Qiuying Peng, Guanyu Nie, Junwei Liao, Jun Wang, Xiaoyun Mo, Jiamu Zhou, Cheng Cheng, Yin Zhao, Jun Wang, and Weinan Zhang

  14. [14]

    InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025

    Robust Function-Calling for On-Device Language Model via Function Masking. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net

    2025

  15. [15]

    Weiwen Liu, Xu Huang, Xingshan Zeng, Shuai Yu, Dexun Li, Shuai Wang, Weinan Gan, Zhengying Liu, Yuanqing Yu, Zezhong WANG, et al . [n. d.]. ToolACE: Winning the Points of LLM Function Calling. InThe Thirteenth International Conference on Learning Representations

  16. [16]

    Zuxin Liu, Thai Hoang, Jianguo Zhang, Ming Zhu, Tian Lan, Shirley Kokane, Juntao Tan, Weiran Yao, Zhiwei Liu, Yihao Feng, Rithesh Murthy, Liangwei Yang, Silvio Savarese, Juan Carlos Niebles, Huan Wang, Shelby Heinecke, and Caiming Xiong. 2024. APIGen: Automated PIpeline for Generating Verifiable and Diverse Function-Calling Datasets. InAdvances in Neural ...

    2024

  17. [17]

    Zhuang Liu, Degen Huang, Kaiyu Huang, Zhuang Li, and Jun Zhao. 2020. Fin- BERT: A Pre-trained Financial Language Representation Model for Financial Text Mining. InProceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020. ijcai.org, 4513–4519

    2020

  18. [18]

    Yuqi Nie, Yaxuan Kong, Xiaowen Dong, John M Mulvey, H Vincent Poor, Qing- song Wen, and Stefan Zohren. 2024. A survey of large language models for financial applications: Progress, prospects and challenges.arXiv preprint arXiv:2406.11903(2024)

    work page arXiv 2024

  19. [19]

    Karmvir Singh Phogat, Chetan Harsha, Sridhar Dasaratha, Shashishekar Ra- makrishna, and Sai Akhil Puranam. 2023. Zero-Shot Question Answering over Financial Documents using Large Language Models.CoRRabs/2311.14722 (2023). https://doi.org/10.48550/arXiv.2311.14722

    work page doi:10.48550/arxiv.2311.14722 2023

  20. [20]

    Karmvir Singh Phogat, Sai Akhil Puranam, Sridhar Dasaratha, Chetan Harsha, and Shashishekar Ramakrishna. 2024. Fine-tuning Smaller Language Models for Question Answering over Financial Documents. InFindings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16,

    2024

  21. [21]

    Association for Computational Linguistics, 10528–10548

  22. [22]

    Cheng Qian, Emre Can Acikgoz, Qi He, Hongru Wang, Xiusi Chen, Dilek Hakkani- Tür, Gokhan Tur, and Heng Ji. 2025. Toolrl: Reward is all tool learning needs. arXiv preprint arXiv:2504.13958(2025)

    work page internal anchor Pith review arXiv 2025

  23. [23]

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al . [n. d.]. ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs. InICLR 2024

    2024

  24. [24]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools.Advances in Neural Information Processing Systems36 (2023), 68539–68551

    2023

  25. [25]

    Zhengliang Shi, Shen Gao, Lingyong Yan, Yue Feng, Xiuyi Chen, Zhumin Chen, Dawei Yin, Suzan Verberne, and Zhaochun Ren. 2025. Tool learning in the wild: Empowering language models as automatic tool agents. InProceedings of the ACM on Web Conference 2025. 2222–2237

    2025

  26. [26]

    Venkat Krishna Srinivasan, Zhen Dong, Banghua Zhu, Brian Yu, Damon Mosk- Aoyama, Kurt Keutzer, Jiantao Jiao, and Jian Zhang. 2023. Nexusraven: a commercially-permissive language model for function calling. InNeurIPS 2023 Foundation Models for Decision Making Workshop

    2023

  27. [27]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, et al . 2025. Stop overthinking: A survey on efficient reasoning for large language models.arXiv preprint arXiv:2503.16419(2025)

    work page internal anchor Pith review arXiv 2025

  28. [28]

    Qiaoyu Tang, Ziliang Deng, Hongyu Lin, Xianpei Han, Qiao Liang, Boxi Cao, and Le Sun. 2023. Toolalpaca: Generalized tool learning for language models with 3000 simulated cases.arXiv preprint arXiv:2306.05301(2023)

    work page arXiv 2023

  29. [29]

    Zichen Tang, Haihong E, Ziyan Ma, Haoyang He, Jiacheng Liu, Zhongjun Yang, Zihua Rong, Rongjin Li, Kun Ji, Qing Huang, Xinyang Hu, Yang Liu, and Qianhe Zheng. 2025. FinanceReasoning: Benchmarking Financial Numerical Reasoning More Credible, Comprehensive and Challenging. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguis...

    2025

  30. [30]

    Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, et al . 2025. Kimi k2: Open agentic intelligence.arXiv preprint arXiv:2507.20534(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    Maolin Wang, Yingyi Zhang, Cunyin Peng, Yicheng Chen, Wei Zhou, Jinjie Gu, Chenyi Zhuang, Ruocheng Guo, Bowen Yu, Wanyu Wang, et al. 2025. Function Calling in Large Language Models: Industrial Practices, Challenges, and Future Directions. (2025)

    2025

  32. [32]

    Mengsong Wu, Tong Zhu, Han Han, Chuanyuan Tan, Xiang Zhang, and Wenliang Chen. 2024. Seal-Tools: Self-instruct Tool Learning Dataset for Agent Tuning and Detailed Benchmark. Springer-Verlag, Berlin, Heidelberg, 372–384

    2024

  33. [33]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. 2025. The rise and potential of large language model based agents: A survey.Science China Information Sciences 68, 2 (2025), 121101

    2025

  34. [34]

    Qianqian Xie, Weiguang Han, Zhengyu Chen, Ruoyu Xiang, Xiao Zhang, Yueru He, Mengxi Xiao, Dong Li, Yongfu Dai, Duanyu Feng, et al . 2024. Finben: A holistic financial benchmark for large language models.Advances in Neural Information Processing Systems37 (2024), 95716–95743

    2024

  35. [35]

    Siqiao Xue, Fan Zhou, Yi Xu, Ming Jin, Qingsong Wen, Hongyan Hao, Qingyang Dai, Caigao Jiang, Hongyu Zhao, Shuo Xie, et al. 2023. Weaverbird: Empowering financial decision-making with large language model, knowledge base, and search engine.arXiv preprint arXiv:2308.05361(2023)

    work page arXiv 2023

  36. [36]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR)

    2023

  37. [37]

    Junjie Ye, Zhengyin Du, Xuesong Yao, Weijian Lin, Yufei Xu, Zehui Chen, Zaiyuan Wang, Sining Zhu, Zhiheng Xi, Siyu Yuan, Tao Gui, Qi Zhang, Xuanjing Huang, and Jiecao Chen. 2025. ToolHop: A Query-Driven Benchmark for Evaluating Large Language Models in Multi-Hop Tool Use. InProceedings of the 63rd Annual Meeting of the Association for Computational Lingui...

    2025

  38. [38]

    Guancheng Zeng, Wentao Ding, Beining Xu, Chi Zhang, Wenqiang Han, Gang Li, Jingjing Mo, Pengxu Qiu, Xinran Tao, Wang Tao, and Haowen Hu. 2024. Adaptable and Precise: Enterprise-Scenario LLM Function-Calling Capability Training Pipeline.CoRRabs/2412.15660 (2024). https://doi.org/10.48550/arXiv. 2412.15660

    work page internal anchor Pith review doi:10.48550/arxiv 2024

  39. [39]

    N., Zeyuan Chen, Ran Xu, Juan Carlos Niebles, Shelby Heinecke, Huan Wang, Silvio Savarese, and Caiming Xiong

    Jianguo Zhang, Tian Lan, Ming Zhu, Zuxin Liu, Thai Hoang, Shirley Kokane, Weiran Yao, Juntao Tan, Akshara Prabhakar, Haolin Chen, Zhiwei Liu, Yihao Feng, Tulika Manoj Awalgaonkar, Rithesh R. N., Zeyuan Chen, Ran Xu, Juan Carlos Niebles, Shelby Heinecke, Huan Wang, Silvio Savarese, and Caiming Xiong. 2025. xLAM: A Family of Large Action Models to Empower A...

    2025

  40. [40]

    Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. 2021. TAT-QA: A Question Answer- ing Benchmark on a Hybrid of Tabular and Textual Content in Finance. InACL. Association for Computational Linguistics. WWW ’26, April 13–17, 2026, Dubai, United Arab Emirates Xing Tang, et al. A Prompt list used in...

    2021

  41. [41]

    Carefully read the generated user query and understand theintended action

  42. [42]

    Review the tool definition to understand each parameter’smeaning and constraints

  43. [43]

    Check the parameter values in tool_call: - Do they match the intent and details in the user query? - Are they internally consistent (no contradictions between parameters)? - Do they comply with the tool definition (value types, required fields, allowed ranges)?

  44. [44]

    Consistent

    Decide if the tool_call is logically correct: - If all parameters reflect the query correctly and satisfy the tool’s definition, return"Consistent". - If there is any mismatch or logical conflict, return"Inconsistent". ## Output format requirements: - Return a (JSON list) containing exactly one object: [{ "analysis": "...your reasoning here...", "result":...

    2026

  45. [45]

    A set of FIVE few-shot examples, each containing: - A user query - A toolset with tool names, descriptions, and parameters - The correct tool calls for that query in strict JSON list format

  46. [46]

    The CURRENT user query we need to process

  47. [47]

    name": "func_name1

    The CURRENT toolset specification ## Your task: - Carefully study the five examples to understand how queries are mapped to tool calls. - For the CURRENT query, use ONLY the tools provided in the CURRENT toolset, along with their descriptions, to determine the exact functions to call and the correct values of their parameters. - Parameter values MUST be d...

    2026

  48. [48]

    **Learn from previous rounds**: Analyze what values were generated and their impact

  49. [49]

    **Focus on current gaps**: Target parameter values that are still underrepresented in local distribution

  50. [50]

    **Avoid redundancy**: Don’t generate values that were already created in previous rounds

  51. [51]

    **Strategic selection**: Choose values that will maximally increase the entropy ratio

  52. [52]

    **Maintain coherence**: Ensure semantic consistency within the cluster context

  53. [53]

    **Diversify query expressions**: Generate queries with varied linguistic forms but similar semantics, avoiding mere entity substitution

  54. [54]

    new_query

    **Preserve tool_call logical coherence**: All parameter values in the generated tool_call must maintain logical consistency. # JSON Output Format: [{ "new_query": "...", "new_value_for_{tool_param.split('.')[-1]}": "...", "new_tool_call": "...", "step_rationale": "Strategic explanation for Round {step}" }]