Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards (arXiv 2610.18868)

Forum topic · 小凯 · 2026-10-15

Summary

Researchers introduce Tev, a framework addressing two bottlenecks in training long-horizon AI agents: the scarcity of verifiable training data and the sparsity of rewards. The paper presents Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards, containing 1,170 verified task environments across 7 domains. Tev-Gym uses a latent-world-first approach: it first builds a complete hidden specification of a task domain, workflows, and expected per-step states, then progressively renders each task into an agent-interactive environment with controllable complexity and a natural progress metric. The authors convert this progress metric into a training recipe via Progress-Potential Reward Shaping (PPRS), which turns progress into dense per-step rewards, improving training efficiency and making previously untrainable tasks feasible. The resulting Tev-27B model achieves significant gains on the Agents' Last Exam computation and math science domain (ALE-CM), matching GPT-6 Astra and Claude Opus 5 on Pass@5. Despite training only on ALE-related data, Tev-27B generalizes to Terminal-Bench 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash, and reaches Claude Opus 4.8-level performance on mid-horizon tasks like TB 2.1. Paper: arXiv 2610.18868, published 2026-10-13.

Paper Overview

Field: NLP Authors: Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh, Nanyun Peng, Kai-Wei Chang, Yong Jae Lee Published: 2026-10-13 arXiv: 2610.18868

Abstract

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). Yet open-source models still struggle on long-horizon tasks, largely for lack of both training data and a working recipe, and scaling both is hard: long, complex tasks are difficult to curate and verify, and their rewards are sparse. This work addresses both problems.

Key Contributions

1. Tev-Gym — the first long-horizon agent training dataset with verifiable, dense, progressive rewards, comprising 1,170 verified task environments across 7 domains. It is built on a latent-world-first approach: a complete and correct hidden specification of the task domain, workflows, and expected per-step states is constructed first, and each task is then progressively rendered from this specification into an agent-interactive environment, enabling controllable complexity and a natural progress metric.

2. Progress-Potential Reward Shaping (PPRS) — a training recipe that converts the progress metric into dense per-step rewards, improving training efficiency and making previously untrainable tasks trainable.

Results

Training with this recipe yields Tev-27B, with significant improvements on the computation and math science domain of Agents' Last Exam (ALE-CM), matching GPT-6 Astra and Claude Opus 5 on 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.
On mid-horizon tasks such as TB 2.1, it reaches Claude Opus 4.8-level performance.