Latent Dynamics · 2026-10-14
Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh, Nanyun Peng, Kai-Wei Chang, Yong Jae Lee
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Tev proposes Tev-Gym, a long-horizon agent training dataset with verifiable, dense, progressive rewards built from a latent-world-first synthesis approach (latent specification → progressive rendering with exact verification). It trains models with progress-potential reward shaping (PPRS), yielding Tev-27B that matches GPT-6 Astra and Claude Opus 5 at Pass@5 on ALE-CM, and generalizes to Terminal-Bench (TB) subsets.
Why it matters
The authors address open-source difficulty on long-horizon tasks caused by sparse terminal rewards and lack of scalable training data with reliable verification. Tev-Gym provides 1,170 verified environments with dense progress signals derived from a latent world, and PPRS turns progress into per-step rewards intended to make otherwise-intractable long-horizon tasks trainable.
Method
- Latent-world-first synthesis: construct a latent world (hidden specification) and progressively render it into an agent-facing environment; ensure exact verification by construction, with an invertible render/corrupt pipeline.
- Tev-Gym dataset: 1,170 verified environments across 7 domains with controllable complexity via scale tiers and workflow DAG horizon; admission uses qualification gates including exploit resistance, progress validity, and role independence.
- Progress-potential reward shaping (PPRS): use progress potential Φ(s_t) from workspace grading at each turn and set per-step reward as change in potential, plus a terminal term for Φ(s_T).
Abstract (from arXiv)
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. We address both. First, we introduce Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards -- 1,170 verified task environments across 7 domains. It is built on a latent-world-first approach that first constructs a latent world: a complete and correct hidden specification of the task domain, workflow, and expected state after each step. Each task is then progressively rendered from this specification into an agent-facing environment, enabling controllable complexity and a natural progress metric. Second, 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-6 Astra and Claude Opus 5 at Pass@5. Despite training only on ALE-related data, it generalizes to the Terminal-Bench (TB) 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash, and reaches Claude Opus 4.8-level performance on middle-horizon tasks such as TB 2.1.
Related papers
- Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design
- Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
- Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
- EvoBelief: Falsifiable World Models for Self-Improving LLM Agents
- SEAR: Self-Evolving Agentic Robotics
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-13 · Indexed by Pub-AI: 2026-10-14 UTC
Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
Preprint posted: 2026-10-13 · Original record ↗
- arxiv
- 2610.18868
Explicit source versions (1)
- v1 · 2026-10-13
- Preprint available
Source date: 2026-10-13 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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. We address both. First, we introduce Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards -- 1,170 verified task environments across 7 domains. It is built on a latent-world-first approach that first constructs a latent world: a complete and correct hidden specification of the task domain, workflow, and expected state after each step. Each task is then progressively rendered from this specification into an agent-facing environment, enabling controllable complexity and a natural progress metric. Second, 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-6 Astra and Claude Opus 5 at Pass@5. Despite training only on ALE-related data, it generalizes to the Terminal-Bench (TB) 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash, and reaches Claude Opus 4.8-level performance on middle-horizon tasks such as TB 2.1.