Long-horizon agents that can work autonomously for hours or days on economically valuable tasks are seen as a key capability for automated discovery and recursive self-improvement. But open-source models still lag on such tasks, in large part because both training data and a working training recipe are scarce: long, complex tasks are hard to curate and verify, and the rewards they produce are sparse.
A new Hugging Face paper, Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards, tackles both problems. The authors introduce Tev-Gym, described as the first long-horizon agent training dataset with verifiable, dense, progressive rewards, spanning 1,170 verified task environments across 7 domains.
Tev-Gym is built on a "latent-world-first" approach: a complete and correct hidden specification of the task domain, workflow, and expected state after each step is constructed first, and each task is then progressively rendered from that specification into an agent-facing environment. This design enables controllable complexity and yields a natural progress metric.
That metric becomes a training recipe through progress-potential reward shaping (PPRS), which converts progress into a dense per-step reward. According to the paper, this improves training efficiency and makes otherwise-intractable tasks trainable.
Training with this recipe produces Tev-27B, which the authors report achieves substantial gains on the Agents' Last Exam computing and math sciences domains (ALE-CM), matching GPT-6 Astra and Claude Opus 5 at Pass@5. Despite training only on ALE-related data, Tev-27B generalizes to Terminal-Bench (TB) 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash, and reaches Claude Opus 4.8-level performance on middle-horizon tasks such as TB 2.1. The paper also notes that Tev-320B tops open models on Terminal-Bench 4.0-CM.