Alibaba TerminalHorizon-27B Hits 55.8% on Terminal-Bench 2.0

Fewer than three thousand training trajectories beat thirty-two thousand on Terminal-Bench 2.0, if the trajectories are long enough. That is the finding in a paper posted October 12 by researchers at AMAP (Alibaba Group), the University of Science and Technology of China, and Southern University of Science and Technology.

Their system, TerminalHorizon, builds 1,535 task environments and 2,852 training trajectories averaging 52.8 interaction turns each (the 90th percentile is 98 turns). Supervised fine-tuning of Qwen3.5-27B on this set produces TerminalHorizon-27B, which scores 55.8% on Terminal-Bench 2.0, up 14.2 points from the base model and three points above Terminal-Universe-27B, which used the same base and harness but roughly 32,000 trajectories. The model also posts 60.67% on Terminal-Bench 2.1 (+13.85) and gains on all 12 benchmarks the authors tested, including +11.7 on AutomationBench pass rate and +23.3 on partial score.

The paper's premise is that short trajectories miss the point. 'Short trajectories can demonstrate commands and tool calls,' the abstract reads, '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.' Rather than scale volume, the pipeline uses an uninvolved agent to probe a candidate environment, surface the parts a lazy construction skipped over, and reconstruct the environment so the real work has to happen.

Training ran on 144 H20-141GB GPUs across 18 nodes for three epochs at a 262,144-token context. The authors also report 'multiple RSI case studies, each with a 12-hour research budget,' where the model continued experimenting past its first solution. The paper does not say whether the trajectories or the synthesis engine will be released.