StateM 通过运行时扩展在 Terminal-Bench 2.1 上实现 95.3% 准确率
原标题：Paper page - StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling
AI 摘要
StateM 是一个通过持久状态、可恢复运行手册和程序化控制来提升长时程智能体执行能力的运行时系统，无需修改模型权重。在 Terminal-Bench 2.1 上，StateM 将 GPT-5.5 xhigh 的准确率从 83.1% 提升至 92.1%，并将 GPT-5.6 Sol xhigh 提升至 95.3%，同时将 DeepSeek-V4 Flash 从 82.7% 提升至 88.1%。该方法通过低成本（约 15 美元）实现了高性能，展示了“harness scaling”作为模型缩放之外的另一条提升路径。
正文节选
StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling Abstract StateM is a runtime system that improves long-horizon agent execution through durable states, recoverable runbooks, and enforceable procedural controls without altering model weights. Long-horizon agents can fail even when their underlying models can solve the constituent steps. They may lose track of mutable state, fail to reactivate lessons from earlier executions, skip known procedu