Updated
Updated · Google Research · Sep 10
Google XR Unveils ToolGrad, Claiming Nearly 100% Pass Rate for 16,000-API LLM Training
Updated
Updated · Google Research · Sep 10

Google XR Unveils ToolGrad, Claiming Nearly 100% Pass Rate for 16,000-API LLM Training

1 articles · Updated · Google Research · Sep 10

Summary

  • ACL 2026 paper ToolGrad flips tool-use data generation by creating a verified API workflow first and only then writing the user query, which Google XR says cuts cost and improves model performance.
  • 16,000-plus ToolBench APIs were used to test the framework against query-first depth-first-search methods; Google says ToolGrad produced more complex long-horizon examples with a higher pass rate at lower generation cost.
  • ToolGrad-500 datasets were used to fine-tune Gemma-3 1B, 4B and 12B models, and Google reports the resulting models beat their base versions and matched state-of-the-art proprietary systems on out-of-distribution benchmarks with unseen tools.
  • Nearly 100% pass rate is the headline claim, with the framework using textual gradients to iteratively propose, execute, select and update API chains—an approach aimed at easing a key scalability bottleneck for agentic AI training.

Insights

Why are Google's top XR researchers spearheading an AI tool-use breakthrough that outperforms industry giants like GPT and Claude?
Could Google's backward approach to AI training fundamentally alter how language models learn to solve complex real-world problems?