Researchers introduced Tev-Gym, a training dataset for long-horizon AI agents, with 1,170 verified task environments across 7 domains. The dataset provides dense, verifiable rewards that measure progress at each step, addressing the sparse rewards and limited training data that the paper identifies as obstacles to training open-source models on long tasks. The paper was published on October 13 and submitted to Hugging Face on October 14.

Tev-Gym uses a “latent-world-first” approach: researchers first create a complete hidden specification of a task domain, workflow and expected state after each step. They then render tasks from that specification into environments that agents can interact with. The method allows task complexity to be controlled and gives each task a natural progress metric.

The researchers use progress-potential reward shaping (PPRS) to turn that metric into a dense reward for each step, which they say improves training efficiency and makes otherwise intractable tasks trainable. Training with this method produced Tev-27B, which matched GPT-6 Astra and Claude Opus 5 on computing and math-science tasks in the Agents’ Last Exam benchmark (ALE-CM) at Pass@5. The model was trained only on ALE-related data.

The paper reports that Tev-27B also generalized to the Terminal-Bench 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash. On middle-horizon tasks such as TB 2.1, it reached performance comparable to Claude Opus 4.8. The Hugging Face page also reports that Tev-320B topped open models on Terminal-Bench 4.0-CM.

Researchers introduced Tev-Gym, a training dataset for long-horizon AI agents, with 1,170 verified task environments across 7 domains. The dataset provides dense, verifiable rewards that measure progress at each step, addressing the sparse rewards and limited training data that the paper identifies as obstacles to training open-source models on long tasks. The paper was published on October 13 and submitted to Hugging Face on October 14.

Tev-Gym uses a “latent-world-first” approach: researchers first create a complete hidden specification of a task domain, workflow and expected state after each step. They then render tasks from that specification into environments that agents can interact with. The method allows task complexity to be controlled and gives each task a natural progress metric.

The researchers use progress-potential reward shaping (PPRS) to turn that metric into a dense reward for each step, which they say improves training efficiency and makes otherwise intractable tasks trainable. Training with this method produced Tev-27B, which matched GPT-6 Astra and Claude Opus 5 on computing and math-science tasks in the Agents’ Last Exam benchmark (ALE-CM) at Pass@5. The model was trained only on ALE-related data.

The paper reports that Tev-27B also generalized to the Terminal-Bench 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash. On middle-horizon tasks such as TB 2.1, it reached performance comparable to Claude Opus 4.8. The Hugging Face page also reports that Tev-320B topped open models on Terminal-Bench 4.0-CM.