Scaling Laws for Tool-Use Agents via Controllable Environment Synthesis
As training data budgets grow, tool-use agents gain more from diverse tools and databases than from additional tasks.
The researchers built SynthAPI to control the tools, tasks and databases in synthetic training environments. Across 458,752 tasks and 849 trained models of four sizes, they derived a scaling law predicting how agents generalize to downstream tasks. More complex training data also delivered larger gains. Agents trained using these findings outperformed models trained on data from prior work across four evaluation benchmarks.
ArXiv · AI/CL/LG's note
The researchers built SynthAPI to control the tools, tasks and databases in synthetic training environments. Across 458,752 tasks and 849 trained models of four sizes, they derived a scaling law predicting how agents generalize to downstream tasks. More complex training data also delivered larger gains. Agents trained using these findings outperformed models trained on data from prior work across four evaluation benchmarks.
ArXiv · AI/CL/LG's note
score 6