ProgAgent: Learning Tool Orchestration through Programmatic Training
ProgAgent trains LLMs to orchestrate tools via code, beating direct tool-calling baselines including at 64k-token states.
ProgAgent is a training stack for programmatic tool calling, where agents invoke tools through executable code and keep intermediate results in runtime variables rather than the context. It includes a task-synthesis pipeline, a lightweight runtime, and a two-stage recipe using 215 tasks at four environment scales. ProgAgent-14B reaches a 30.7% pass rate on LOCA-bench at a 64k-token environment state, where compared same-size baselines score zero, and PTC-trained models outperform direct-tool-calling models under both interfaces.
- Programmatic tool calling stores intermediates in variables instead of context.
- Training uses 215 synthetic verifiable tasks at four environment scales.
- ProgAgent-14B scores 30.7% on LOCA-bench at a 64k-token state.
- PTC-trained models beat DTC-trained models under both evaluation interfaces.
From the original article
1 min read at huggingface.co
Large language models (LLMs) are increasingly expected to orchestrate diverse tools over long-horizon tasks. Existing agents typically use Direct Tool Calling (DTC), where the model invokes tools sequentially, and tool responses accumulate in the model context, which can lead to context rot and orchestration breakdown.
By Hao Li, Keqing He, Longtao Zheng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.