Tev-27B: Long-Horizon Agent Training with Dense Progress Rewards
An arXiv paper introduces 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. Using a latent-world-first approach, the authors first build a complete hidden specification of each task domain, workflow and expected per-step state, then progressively render it into an agent-facing environment, and convert progress into dense per-step rewards via progress-potential reward shaping (PPRS) to train Tev-27B. The paper reports that Tev-27B matches GPT-6 Astra and Claude Opus 5 at Pass@5 on the Agents' Last Exam computing and math sciences domains (ALE-CM), and that despite training only on ALE-related data it generalizes to Terminal-Bench 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash and reaching Claude Opus 4.8-level performance on mid-horizon tasks such as TB 2.1.
Sparse rewards and scarce training data are key bottlenecks for long-horizon 智能体, so pairing a verifiable progressive-reward dataset with a training recipe offers a reproducible route for open models to close the gap with frontier systems — though the comparisons are currently the authors' own claims.
Computer Science > Computation and Language
arXiv:2610.18868 (cs)
[Submitted on 13 Oct 2026]
Title:Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
Authors:Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh, Nanyun Peng, Kai-Wei Chang, Yong Jae Lee
View a PDF of the paper titled Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards, by Wenbo Hu and 5 other authors
#Tev#Tev-Gym#Tev-27B#progress-potential reward shaping#Agents' Last Exam#Terminal-Bench#GPT-6 Astra#Claude Opus 5#Claude Opus 4.8#GLM 5.2#Gemini 3.7 Flash