Tool-use agents improve when training diversifies tools and database states

Across 849 trained models, the authors show that synthetic environment composition predicts generalization better than task counts alone.

Big Tech

Andrew Szot · Samira Abnar · Omar Attia · Abbas Kazerouni · Michael Kirchhof · Cheng-Yu Hsieh · +2 more

Apple

Research Digest·28 October 2026·2 min read

Illustration: Tool-use agents improve when training diversifies tools and database states

In one line

To scale tool-use agent training data, grow tool and database diversity, not just task count.

Before you read this

Three ideas this paper builds on

scaling lawA predictable power-law relationship between compute or data and model performance.16 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.41 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.266 papers →

Read it as

Teaching a large language model (LLM), the AI behind chatbots, to use tools like calculators, databases, and application programming interfaces (APIs) requires many examples. Those examples are rare in ordinary text, so researchers create synthetic training environments.

The researchers built SynthAPI, a pipeline that generates such environments while varying three things independently: tools, tasks, and initial database states. Training many models on controlled mixes revealed how each kind of variety affects learning.

The key finding: as the training budget grows, adding more tools and more varied starting states helps more than adding more tasks. A mix chosen by this rule outperformed agents trained on prior large-scale data by 30.5 percentage points on average, showing where to put generation effort.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

scaling law synthetic data generation tool invocation verification large language model agentic evaluation