4B Model BFCL Jumps 15%: Snowflake Proves Power of Mid-Tool Training
TheTuringPost · x · 2026-08-23
Snowflake's MidTool paper demonstrates that introducing tool learning during mid-training—before SFT and RL—significantly boosts model performance. The study constructed a 20.3B token corpus combining context-grounded examples and native agent trajectories.
Key Findings:
- 4B Model: BFCL rose from 39.73% to 50.25% after SFT, and from 39.51% to 54.18% after RL.
- 8B Model: Pass@1 on τ²-Bench improved from 13.04% to 19.96%.
- Data Nuance: Executable trajectories were best for function calling, while documentation-grounded data transferred better to unfamiliar environments.
Limitation: All models scored 0% on the MCP-Universe web-search subset, indicating that learning tools/workflows is insufficient for long-horizon research tasks.
More from coding & agent
- raylib 6.2 ships with 300+ improvements and 100 new contributors, 20 new API functions — jsuarez · 2026-10-16
- "Vibecoding" is as misleading a label as "AI" and "SI," dev complains — eschadiol · 2026-10-16
- Doc2Harness: New benchmark finds best coding agent passes only 53% across 329 cases — _akhaliq · 2026-10-16
- Perplexity open-sources SPACE, the Firecracker microVM sandbox behind Perplexity Computer — perplexity_ai · 2026-10-16
- Blackstone goes all-in on Factory for autonomous software development — matanSF · 2026-10-16
- Longtime Windsurf user cancels $10/month legacy plan, citing AI subscription fatigue — tristanbob · 2026-10-16