Reproduce “Salesforce Koa: An Enterprise Language Model for Agentic Tool Use” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zixiang Chen, Sufeng Niu, Yingchi Liu, Wenting Zhao, Akshara Prabhakar, Shubham Mehrotra, Bin Bi, Zhujun Lan and 18 more

Koa uses declarative enterprise-agent specifications to generate simulated tasks and tool-grounded rewards for reinforcement learning. It is post-trained from Nemotron-3-Super-120B with GRPO using public or synthetic interaction data. On CRMAgentBench, Koa reaches 87% task success, versus 82% for GPT-4.1 and 79% for its untuned base. In-domain RL raises tool-call success on the human-labeled set, while BFCL accuracy remains near base.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.