StateM：透過框架擴展在 Terminal-Bench 2.1 上達到 95.3% 原始準確率，或實現 15 美元的前沿級運行

摘要

長視野代理即使其底層模型能解決構成步驟，仍可能失敗。它們可能失去對可變狀態的追蹤，未能重新啟用先前執行中的教訓，跳過已知程序，或過早停止。我們押注於框架擴展，以改善代理周遭的執行系統，而不改變其模型權重。我們引入 StateM，一個代理原生運行時，其圍繞持久狀態、階段局部上下文、受檢轉換、可恢復的運行手冊，以及可供代理與使用者共同檢視的版本化程序化實踐來組織執行。在 Terminal-Bench 2.1 上，StateM 將 GPT-5.5 xhigh 提升至 92.1%，而參考基準為 83.1%，GPT-5.6 Sol Ultra 則為 91.9%。該運行手冊在不作修改的情況下轉移至 GPT-5.6。使用 GPT-5.6 Sol xhigh，StateM 在 445 次試驗中達到 95.3% 的原始準確率，並在 89 個任務中每一項至少成功一次。凍結的設定檔將 GPT-5.6 Luna 從 76.7% 提升至 85.4%，高於 84.9% 的 Sol xhigh 參考。使用相同的運行時、運行手冊結構與黃金規則，少於 38 美元的調整使 DeepSeek-V4 Flash 在標準逾時下從 82.7% 提升至 88.1%，並在 88 個任務的公共核心上達到 89.1%。僅延長剩餘的延遲敏感任務，即可匹配已報告的 88.8% GPT-5.6 Sol 最高結果。最終評分的 API 使用量約為 15 美元，相比之下 GPT 參考為 574.68 美元；DeepSeek 的總支出為 52.22 美元。在 BusinessBench 上，基於開發集建構的家族特定運行手冊，在保留集上分別產生 0.55 個宏觀點數與 1.34 個微觀點數的增益；兩個機制匹配的家族提升了 10.04 個點數。當任務共享執行結構時，具體規則得以泛化，而控制方法則廣泛適用。StateM 將選定的事後檢討發現轉化為持久、可執行的前置條件與實踐，透過狀態化控制使學得的控制明確且可強制執行。程式碼位於 github.com/henryqin1997/statem。

English

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 procedures, or stop prematurely. We bet on harness scaling to improve the execution system around an agent without changing its model weights. We introduce StateM, an agent-native runtime that organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices that agents and users can inspect together. On Terminal-Bench 2.1, StateM raises GPT-5.5 xhigh to 92.1\%, versus 83.1\% reference and GPT-5.6 Sol Ultra at 91.9\%. The runbook transfers unchanged to GPT-5.6. With GPT-5.6 Sol xhigh, StateM reaches 95.3\% raw accuracy across 445 trials and succeeds on all 89 tasks at least once. The frozen profile raises GPT-5.6 Luna from 76.7 to 85.4\%, above the 84.9\% Sol xhigh reference. Using the same runtime, runbook structure, and golden rules, less than \38 of adaptation raises DeepSeek-V4 Flash from 82.7 to 88.1% under standard timeouts and to 89.1% on an 88-task common core. Extending only the remaining latency-sensitive task matches the reported 88.8% GPT-5.6 Sol max result. Final-score API usage is about 15 versus \574.68 for the GPT reference; total DeepSeek expenditure is 52.22. On BusinessBench, family-specific runbooks built on development sets yield held-out gains of 0.55 macro and 1.34 micro points; two mechanism-matched families improve by 10.04 points. Concrete rules generalize when tasks share execution structure, while the control methodology applies broadly. StateM turns selected postmortem findings into persistent, executable preconditions and practices, making learned controls explicit and enforceable through stateful controls. Code at github.com/henryqin1997/statem.