Reproduce “Turnslide: Scalable Multi-Turn Data Synthesis by Walking a Finite-State Machine” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aaron Fainman, Gabriela Kadlecová, Maciej Gryka, Bartosz Kruszczyński, Usman Zafar, Cédric Archambeau, Aaron Klein, David Salinas and 2 more
Turnslide generates synthetic multi-step tool-use conversations for small language models by inferring a finite-state machine of API conditions, planning valid tool-call routes, and generating each conversation in one LLM call. On the BFCL v4 TradingBot task, Qwen3.5-2B trained on its data achieved 70.7% full-conversation accuracy, exceeding two compared methods under a fixed 20-million-token budget; this comparison is limited to TradingBot.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.