Agents
Environment-free Synthetic Data Generation for API-Calling Agents
Apple ML shows LLM world-model simulation can generate API-agent training trajectories from specs alone—no executable environments required.
Environment-free Synthetic Data Generation for API-Calling Agents
What the paper does
Apple ML Research generates synthetic trajectories for API-calling agents using LLMs as on-the-fly digital world models. Given only API specifications, an LLM creates tasks; a teacher agent solves them while an LLM simulator emits coherent stateful API responses; an LLM judge filters quality.
Key findings
- Collecting real trajectories usually needs full environments and populated backends—a scalability bottleneck.
- Environment-free simulation from specs alone produced useful supervision on AppWorld and OfficeBench (IR and state-changing tasks).
- Fine-tuning on the synthetic data yielded significant gains, establishing LLM-based API simulation as a practical training path.
Why it matters
Decouples agent post-training data from environment engineering cost—important for MCP/tool ecosystems that change faster than sandboxes can be built.
Related
- environment-free-synthetic-data-generation-for-api-calling-a
- uncertainty-quantification-for-llm-function-calling
- apple-machine-learning-research-at-iclr-2026
Source: https://machinelearning.apple.com/research/environment-free | 2026-07-23