Firefly reverses task synthesis: explore real APIs first, then write verifiable tool-call tasks
In a NeurIPS 2026 paper reported by 机器之心, the Firefly team describes a pipeline that explores real MCP servers first and then back-chains questions and ground truth from the observed outputs. The team says the pipeline produced 5,144 verifiable tasks spanning 240 servers and 993 tools, and it comes with a Retrieval-Augmented Tool-Call Simulator that caches observed tool calls so reinforcement-learning training and evaluation can run fully offline and reproducibly. The report notes that stateful and authentication-requiring servers were excluded during filtering, and that the team's hypothesis that larger models such as 30B or above may benefit more from verifiable supervision remains unvalidated.
The report frames verifiable, reproducible tool-call training data as an open problem, with manual annotation hard to scale, purely simulated environments drifting from real API behavior, and forward synthesis producing ground truth with no guarantee of correctness. Firefly's exploration-first design and its simulator address those gaps, letting RL rollouts run offline without live API cost or availability concerns and with identical tool behavior in training and evaluation.
机器之心 2026-11-07 12:04 辽宁
研究社区探索可验证Agent学习的一个起点
在训练 Tool-Calling Agent 的过程中,我们反复遇到一个看似简单、却非常关键的问题:训练数据从哪里来?
Tool-Calling Agent 需要根据用户意图调用外部 API、整合中间结果,并最终给出答案。要通过 Reinforcement Learning 训练这样的 Agent,我们需要大量任务,每个任务都应当带有正确的中间 tool calls 和可以被验证的最终答案。即使是数据效率较高的 RL 方案,也往往需要数千个这样的任务;而目前主流的 agent benchmark,每个通常只提供几十到几百个任务。
#Firefly#Retrieval-Augmented Tool-Call Simulator#MCP Server#Tool Graph