TerminalHorizon: preserving the work in synthetic tasks

TL;DR

TerminalHorizon constructs long-horizon terminal environments while checking that environment generation has not already solved the work the agent should perform. (source)

Authors & Org

Shun Zou, Ziyu Ma, Yi Zou, Yong Wang, Lin Chen, Zehui Chen, Guanhua Chen, Xiangxiang Chu and Feng Zhao. Affiliations were not established from the abstract page. Submitted 2026-10-09. (source)

Method

Independent execution identifies essential work omitted or simplified during environment construction and guides reconstruction. Tasks progressively add requirements under the same goal, preserving the need for exploration, implementation and verification. The authors build 1.5K task environments and fine-tune Qwen3.5-27B to obtain TerminalHorizon-27B. (source)

Results

The authors report gains on 12 benchmarks spanning coding, production agent tasks and reasoning. Recursive self-improvement case studies use a 12-hour research budget each and report continued experimentation and retained improvements. The abstract supplies neither per-benchmark scores nor a cost-matched comparison. (source)

Significance

For Agentic Reinforcement Learning and Agents (LLM Agents), task difficulty depends on what the environment leaves undone. This result concerns supervised fine-tuning on synthesized trajectories; it should not be described as an RL gain merely because the tasks are agentic. (source)

Open Questions

Do gains survive independent tasks, and how much compute does synthesis require? How well does reconstruction detect subtle shortcuts? The abstract does not answer these questions, and full-paper and code review remain outstanding. (source)

Cite

• Scaling Agentic Data for Long-Horizon Terminal Intelligence, arXiv:2610.13166.
• Captured abstract page.