StateM Improves Agent Runtime Accuracy
- •StateM reaches 95.3% raw accuracy across 445 Terminal-Bench 2.1 trials
- •Harness scaling raises GPT-5.5 xhigh to 92.1%, above 83.1% reference
- •BusinessBench runbooks add 0.55 macro and 1.34 micro held-out gains
Ziheng Qin, Yaxin Lu, Zhangyang Atlas Wang and Kai Wang published StateM, a runtime system for long-horizon agents, on Aug 15, with the paper submitted to Hugging Face by Kai Wang on Aug 18. StateM aims to improve agent execution by scaling the harness around a model rather than altering model weights, using durable states, phase-local context, checked transitions, recoverable runbooks and versioned procedural practices that agents and users can inspect together.
The paper says long-horizon agents can fail even when their models can solve individual steps because they may lose mutable state, fail to reuse lessons from earlier runs, skip known procedures or stop too early. StateM turns selected postmortem findings into persistent, executable preconditions and practices, making learned controls explicit through stateful controls.
On Terminal-Bench 2.1, StateM raises GPT-5.5 xhigh to 92.1% from an 83.1% reference, ahead of GPT-5.6 Sol Ultra at 91.9%. The same 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. A 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, while total DeepSeek expenditure is 52.22.
On BusinessBench, family-specific runbooks built on development sets produce held-out gains of 0.55 macro and 1.34 micro points, while two mechanism-matched families improve by 10.04 points. The authors say concrete rules generalize when tasks share execution structure, and code is available at github.com/henryqin1997/statem.
The fastest, most reliable AI news, straight to your inbox