A new preprint from researchers at Southeast University and Huawei Noah’s Ark Lab proposes a way to train command-line AI agents that connects their learning signal to the commands that contribute to a task’s verified result. On two versions of a terminal-agent benchmark, the method scored above the strongest competing reinforcement-learning approach across all eight combinations of model size, training dataset, and benchmark version.
When an agent completes a multi-step terminal task, a verifier usually checks the final files or processes, not every command along the way. Standard outcome-based training can give the whole sequence a shared signal, even when some commands were useful and others did not contribute to the result.
The proposed method, Dependency-Aware Group Policy Optimization (DepGPO), uses the execution record to trace backward from the resources the verifier inspected. It gives more credit to writes whose contents feed into those resources, including an intermediate script that later creates a checked file, and some credit to earlier reads that guided a useful write. The task’s final reward still determines whether a run is reinforced or penalized; the trace changes how that signal is distributed across steps.
The authors tested two Qwen model sizes, 9 billion and 27 billion parameters, trained on either the SETA or TMAX terminal-agent data. DepGPO led the other reinforcement-learning methods on both Terminal-Bench 2.0 and 2.1. Pass@1, the single-attempt solve rate estimated from five attempts per task and three evaluation rounds, was 3.22 to 10.03 percentage points above the strongest competing method in each setting. For example, with the 9B model trained on TMAX, it scored 20.45% versus 17.23% on version 2.0 and 22.17% versus 18.95% on version 2.1. The benchmark maintainers say version 2.1 fixes issues in 28 of version 2.0’s 89 tasks, so the two editions are related but not identical tests.
In the paper’s Qwen3.5-9B/SETA cost comparison, trace processing added 7% to wall-clock time per training update and 1% to peak GPU memory, with no extra model forward passes. Those measurements come from that one configuration.
The results remain a preprint’s benchmark evidence: the paper tests two Qwen sizes and two training datasets, and its reported variation comes from repeating evaluation on fixed trained checkpoints, not from independent training runs. It does not test deployment or establish that the method transfers to other models or real-world workloads.
Sources
- DepGPO research paper (arXiv preprint, submitted Oct. 2, 2026)
- Terminal-Bench paper: benchmark design and task scope
- Terminal-Bench maintainers on the 2.1 task revisions
- Illustrative stock photograph of code on a monitor, not an image of the study: Markus Spiske via Wikimedia Commons, CC0 1.0.
.jpg){kind=link}