arXiv：cs.CL· Aaron Fainman, Gabriela Kadlecov\'a, Maciej Gryka, Bartosz Kruszczy\'nski, Usman Zafar, C\'edric Archambeau, Aaron Klein, David Salinas, Selim Nowicki, Jacek Golebiowski·· 6 小时前AI 评分37
Turnslide：用有限状态机实现可扩展多轮对话数据合成
Turnslide: Scalable Multi-Turn Data Synthesis by Walking a Finite-State Machine
AI 导读
研究者提出 Turnslide，将每个 API 建模为有限状态机，生成状态合法的工具调用序列，再用单次 LLM 调用把序列转成完整多轮对话样本。在小语言模型微调测试中，该方法在多轮工具调用任务上达到 70.7% 完整准确率，超过对比方法的 63.4% 和 53.7%，且消耗 token 少 3.6-6.6 倍。该工作被 NeurIPS 2026 SLM-Agents Workshop 接收。
正文
Abstract:Small language models are inexpensive to serve and can run on private infrastructure, but base models are often not good enough at multi-turn tool calling, and fine-tuning them needs per-API data that rarely exists. Existing synthesis methods are too expensive for high-scale fine-tuning, as they often require mock operational environments for different domains and multiple LLM calls per generated conversation turn. We introduce a fully automated, lightweight synthesis framework that models each API as a finite-state machine, representing the system as abstract states that determine when each tool may be called, producing state-valid sequences of tools; sequences are translated into complete examples with a single LLM call. Rather than optimize diversity, we set a target distribution over the number of turns, the tool sequence and task complexity. We measure data quality by fine-tuning SLMs on generated trajectories, showing that our FSM-based generation significantly improves downstream accuracy over an unmutated baseline and, against existing works, reaches 70.7% full accuracy over 63.4% and 53.7% with 3.6-6.6$\times$ fewer tokens.
| Comments: | Accepted at the SLM-Agents Workshop, NeurIPS 2026 (non-archival) |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.07070 [cs.CL] |
| (or arXiv:2610.07070v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.07070 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Gabriela Kadlecová [view email]
[v1]
Mon, 5 Oct 2026 08:48:26 UTC (2,370 KB)
来源：arXiv：cs.CL · arxiv.org