Firefly back-chains from real MCP APIs to generate 5,144 verified tool-call RL tasks

机器之心 · wechat · 2026-11-07

A detailed walkthrough of the NeurIPS 2026 paper Firefly, which tackles the scarcity of verifiable training data for tool-calling agents. Existing approaches each miss one requirement: human annotation scales poorly ($200–2,000 per RL task per Epoch AI), LLM-simulated environments diverge from real API behavior, and forward generation (write task first, solver's answer becomes ground truth) offers no correctness guarantee. Firefly inverts the order — Back-Chaining: a strong LLM first explores real MCP servers, then writes tasks around observed outputs, making ground-truth correctness a constructive property of the pipeline. The pipeline filters 3,000 servers to 240 (993 tools) by statelessness/no-auth/schema clarity/non-triviality, builds an 83K-edge ToolGraph as a weak semantic prior, grows tool-call DAGs on live servers (fan-out/fan-in), and validates tasks with a five-dimension LLM judge, yielding 5,144 verified tasks. A retrieval-augmented simulator (exact hash match → fuzzy retrieval with LLM-conditioned generation → conservative error) makes RL fully reproducible without live APIs.

Original post →

More from coding & agent

coding & agent channel →