Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
Tev-Gym gives long-horizon agents step-by-step rewards across 1,170 verified task environments.
The paper says its latent-world-first setup builds hidden task specifications, then renders them into agent-facing environments with measurable progress. That progress becomes dense per-step reward through progress-potential reward shaping. The authors report that Tev-27B improves on ALE-CM, matches GPT-6 Astra and Claude Opus 5 at Pass@5, and transfers to Terminal-Bench subsets despite training only on ALE-related data. Source: ArXiv · AI/CL/LG's note.
The paper says its latent-world-first setup builds hidden task specifications, then renders them into agent-facing environments with measurable progress. That progress becomes dense per-step reward through progress-potential reward shaping. The authors report that Tev-27B improves on ALE-CM, matches GPT-6 Astra and Claude Opus 5 at Pass@5, and transfers to Terminal-Bench subsets despite training only on ALE-related data. Source: ArXiv · AI/CL/LG's note.
score 6