Scaling Agentic Data for Long-Horizon Terminal Intelligence
Researchers fine-tune Qwen3.5-27B into TerminalHorizon-27B using 1.5K long-horizon terminal environments.
The paper introduces TerminalHorizon, an automated engine that builds long-horizon terminal-agent environments from real professional work and reconstructs tasks so agents must explore, implement, and verify. It yields 1.5K environments used to supervised-fine-tune Qwen3.5-27B into TerminalHorizon-27B. The authors report substantial gains on all 12 benchmarks covering coding agents, production agentic tasks, and general reasoning. Case studies with a 12-hour research budget show the model keeps experimenting, uses feedback, and retains useful improvements.
- TerminalHorizon synthesizes long-horizon terminal environments from real professional work.
- Independent execution finds omitted key work and guides reconstruction.
- Supervised fine-tuning of Qwen3.5-27B produces TerminalHorizon-27B on 1.5K environments.
- The model gains on all 12 coding, agentic, and reasoning benchmarks.
- Twelve-hour RSI studies show continued experimentation and retained improvements.
- AI models
- Qwen3.5-27BTerminalHorizon-27B
From the original article
1 min read at huggingface.co
Terminal agents are moving beyond local operations toward completing real-world professional tasks. Short trajectories can demonstrate commands and tool calls, but they do not fully show how agents accumulate evidence through exploration, use it to inform judgments, continually adjust strategies based on feedback, and turn intermediate results into final deliverables.
By Shun Zou, Ziyu Ma, Yi Zou · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.