AgentMercury: Your Agent Can Synthesize Verifiable Environments for Business Scenarios at scale
AgentMercury generates executable business “worlds” first, then lets tasks emerge from them.
The paper reports 4,783 synthesized environments across 14 industries and 50 countries.
Those environments include entities, services, tools, state, and executable cross-service invariants.
Training on them improved Qwen3.5-4B on EnterpriseOps-GYM and AIME26, even though the environments were not built for those benchmarks.
The authors also report that fine-tuning a larger Qwen model on construction traces raised executable-world authoring success from 3.3% to 83.3%.
Source: HF Daily Papers' note
The paper reports 4,783 synthesized environments across 14 industries and 50 countries.
Those environments include entities, services, tools, state, and executable cross-service invariants.
Training on them improved Qwen3.5-4B on EnterpriseOps-GYM and AIME26, even though the environments were not built for those benchmarks.
The authors also report that fine-tuning a larger Qwen model on construction traces raised executable-world authoring success from 3.3% to 83.3%.
Source: HF Daily Papers' note
score 5