BizBench：揭示长程智能体在真实商业部署中的隐藏挑战
原标题：Paper page - BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
AI 摘要
研究团队发布 BizBench，一个基于某领先支付公司真实业务流程构建的多轮长程智能体基准，包含 789 份知识文档、601 个可发现业务工具和 50 个任务，用于评估智能体在支付等高风险场景中的中间执行合规性。实验显示，前沿模型表现有限：Claude Opus 5 与 GPT-6 Sol 的严格成功率仅为 64.0% 和 61.3%，Qwen3.8-27B 目标达成率 42.0% 但严格成功率降至 32.7%，DeepSeek-V4.1-Flash 平均每次运行调用工具 191.3 次，比参考执行多约 70%。该基准揭示了仅看最终结果会掩盖无效执行的问题，推动长程智能体评估从任务完成转向约束合规的可靠执行。
正文节选
BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business Abstract BizBench evaluates long-horizon business agents on intermediate execution validity using real-world payment procedures, revealing that frontier models frequently violate constraints despite achieving partial goal success. Existing long-horizon agent benchmarks often emphasize final task success while providing limited assessment of intermediate execution validity. Yet real-world business opera