Alibaba's TerminalHorizon synthesizes long-horizon agent data, lifting Qwen3.5-27B across 12 benchmarks
alibabagroup · hf · 2026-10-13
A USTC and Alibaba (AMAP) team introduce TerminalHorizon, a fully automated data synthesis engine for long-horizon terminal agents, and build 1.5K task environments to fine-tune Qwen3.5-27B into TerminalHorizon-27B.
- Motivation: short trajectories fail to show how agents accumulate evidence, adjust strategies from feedback and produce deliverables; existing environment synthesis pre-completes key work, removing the exploration, implementation and verification agents should perform.
- Method: start from real-world professional work, use independent execution to identify omitted key work and reconstruct environments accordingly; progressively expand tasks under the same goal with new requirements building on prior results.
- Results: substantial gains on 12 benchmarks spanning coding-agent tasks, real-world production agentic tasks and general reasoning; RSI case studies (12-hour research budgets each) show continued experimentation, feedback-driven improvement and retention of effective improvements.
More from coding & agent
- yetone shares his workflow: 3 parallel Claude Code loops, nightly human cleanup — lxfater · 2026-10-13
- NVIDIA ships a NURBS B-rep kernel on USD, but devs slam its nondeterministic geometry — rms80 · 2026-10-13
- huashu-chrome open-sources: let any AI agent drive your Chrome with full login state — AlchainHust · 2026-10-13
- Indie dev built his own custom heatmaps and analytics just because he could — tomjohndesign · 2026-10-13
- OpenAI Challenge agent went rogue: burned $200+ and dumped 80GB of junk — ZeusRah · 2026-10-13
- Indie dev builds CRED in a weekend: 8,413 free Canadian economic series, 3x FRED — RichardsonDx · 2026-10-13