AI TrackerFrontier desk · AEST
Back to news
normalaitech

Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards

arXiv cs.AI · 14 Oct 2026, 03:59 AEST

Tev introduces a new long-horizon agent training dataset called Tev-Gym, featuring 1,170 verified task environments with dense, progressive rewards. This approach enhances training efficiency and enables agents to tackle complex tasks more effectively.

Key facts

Read aloud

Tev introduces a new long-horizon agent training dataset called Tev-Gym, featuring one thousand one hundred seventy verified task environments with dense, progressive rewards. This approach enhances training efficiency and enables agents to tackle complex tasks more effectively. Long-horizon agents have drawn growing attention for their ability to work autonomously -- from hours to days -- on economically valuable tasks, a capability essential to automated discovery and recursive self-improvement RSI. We introduce Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards -- one thousand one hundred seventy verified task environments across seven domains. We turn this metric into a training recipe: via progress-potential reward shaping PPRS, we convert progress into a dense per-step reward that improves training efficiency and makes otherwise-intractable tasks trainable. Training with this recipe yields Tev-27B, which achieves substantial gains on the Agents' Last Exam computing and math sciences domains ALE-CM, matching GPT-six Astra and Claude Opus five at Pass@five.

Sources in this cluster