Google's ToolGrad reverses tool-use data generation, hits 83.1 BFCL with 12B Gemma-3
dl_weekly · x · 2026-09-23
Google Research published ToolGrad (ACL 2026), a data generation framework that reverses the traditional paradigm: it first generates a ground-truth tool-use chain, then annotates the matching user query, instead of generating instructions first and searching solutions via DFS agents (ToolBench/ToolACE style).
- Motivation: distilling trajectories from agent exploration is inefficient; manual annotation (as in InstructPipe) doesn't scale
- Results: building data backwards from valid tool chains, a 12B Gemma-3 student model reaches a 99.8% pass rate and an 83.1 BFCL score
- Introduces textual "gradients" for efficient tool-use dataset generation
More from coding & agent
- Cursor cuts token costs 7% via tighter prompts, caching, without quality loss — mattyp · 2026-09-23
- Everyone builds AI agents, almost nobody builds the harness around them — alex_verem · 2026-09-23
- Claude generates a stunning Three.js Japanese boat scene, and people can't believe it — nateliason · 2026-09-23
- The Agentic AI Storage Shock: enterprise agents turn storage into the next bottleneck — BenBajarin · 2026-09-23
- Claude Opus 5.5 impresses at building animated maps in hands-on test — DavidmComfort · 2026-09-23
- Self-host Firecrawl and give agents a Firecrawl skill for cleaner, cheaper web reading — gregmushen · 2026-09-23