Self-Supervised Scaling of Terminal Environments for Scientific Domains

Self-Supervised Scaling of Terminal Environments for Scientific Domains. By Zhongzhi Li and ten others, from Tencent HY LLM Frontier, the University of Georgia, and five other institutions. Published on the second of October. No code repository is linked to it.

Terminal agents are now being deployed beyond software engineering, into science and other specialised domains. Training them needs environments, and an environment needs two components. One is executable reference behaviour. The other is a verifier specific to the domain, able to tell an output that is semantically correct from an artifact that merely looks plausible. Authoring both of those by hand, task after task, takes repeated engineering and leaves little that can be reused.

Software-in-the-loop reconstruction, shortened to SWR, gets both components from software that already exists. A scientific software workflow is an executable program that maps structured inputs to outputs. Run it, and it supplies the reference outputs and the verification targets for free. The framework is self-supervised in that sense: the supervision comes from the workflow, not from someone writing a solution.

For each workflow, many input configurations are executed, and the cases are split into public observations and hidden evaluations. The agent receives the instruction, the input schema, and the public input and output pairs. It never sees the source workflow. From those it builds an editable program, which is then run on the hidden configurations and compared with what the workflow produced.

Checking is done by a hierarchical verifier. It combines semantic comparison specific to the domain, a test of structural validity, and checks designed to catch shortcuts. Feedback on the public cases lets the agent revise iteratively. Because nothing in this requires a reference solution per task, new workflows and new configurations can be added freely.

The instantiation uses five hundred workflows from forty-six software families, across six domains. Given three attempts per task, Qwen3.8-Max solves eight hundred and thirty-eight tasks and produces one thousand four hundred and twenty-two verified trajectories. Those are oversampled to three thousand training examples built only from reconstruction.

Supervised fine-tuning of Qwen3.8-27B on that data lifts mean Terminal-Bench 2 performance from forty-seven point nine four percent to fifty-three point five six percent, averaged over three seeds. Against four control corpora matched for token count, it achieves the highest mean on all four of the evaluations reported.

The authors' conclusion is that existing scientific software can serve as a scalable source of supervision for terminal agents, and supervision whose correctness is checked against behaviour rather than appearance. Two choices carry that. The software itself supplies the answers, so adding a task means adding a workflow and not writing a solution. And the agent is judged on hidden inputs, against a workflow it never saw, with checks aimed at shortcuts, so a plausible-looking imitation of the public examples is not enough to pass.