Researchers have developed a new approach to help artificial intelligence (AI) systems better handle complex tasks that take hours or even days to complete. These “long-horizon” tasks are crucial for applications like automated scientific discovery and improving AI’s own capabilities over time. However, current open-source AI models often struggle with such challenges because of limited training data and the difficulty of providing clear feedback during learning. A newly published research paper introduces a solution combining a large, carefully designed training dataset with a novel method to give AI more detailed guidance as it learns, making it easier for AI to master long, complicated tasks.
Key Takeaways
- The research introduces Tev-Gym, the first large-scale dataset specifically designed for training AI on long-horizon tasks, featuring 1,170 verified task environments across seven different domains.
- Tev-Gym uses a “latent-world-first” approach, creating a hidden, detailed blueprint of each task’s workflow and expected outcomes before presenting it to the AI, ensuring accuracy and controllable complexity.
- A new training technique called Progress-Potential Reward Shaping (PPRS) turns progress into frequent, step-by-step rewards, helping AI learn more efficiently from long, sparse tasks.
- Using this approach, the team trained Tev-27B, an AI model that matches or surpasses top AI systems like GPT-6 Astra and Claude Opus 5 on challenging computing and math science tasks, and generalizes well to other test environments.
Long-horizon tasks are those that require an AI to plan and act over extended periods, often involving many steps with complex dependencies. For example, automating a scientific experiment or managing a multi-day project involves numerous decisions where success depends on earlier actions. Training AI to handle these tasks is difficult because the feedback (or “rewards”) it receives often only comes at the end, making it hard to know which steps were good or bad. This “sparse reward” problem slows down learning and makes many tasks nearly impossible to train on effectively.
To tackle this, the researchers created Tev-Gym, a comprehensive training environment that offers “dense” rewards—meaning the AI gets feedback regularly throughout the task, not just at the end. They achieved this by building a “latent world” for each task: a hidden, detailed map that specifies the task’s workflow and the expected state after every step. This latent world serves as a ground truth that can be progressively rendered into a form the AI interacts with, allowing the researchers to measure progress accurately and provide continuous rewards.
The key innovation in training is the Progress-Potential Reward Shaping (PPRS) technique. Instead of waiting for final success or failure, PPRS assigns a reward based on how much progress the AI makes toward completing the task at each step. This approach turns a previously difficult training problem into a more manageable one by giving the AI more consistent and informative signals, accelerating learning.
Using these tools, the team trained an AI model named Tev-27B. Despite being trained primarily on data from the Agents’ Last Exam computing and math sciences domains, Tev-27B performed impressively, matching or exceeding the performance of advanced commercial AI models like GPT-6 Astra and Claude Opus 5 in certain benchmarks. Moreover, Tev-27B showed strong generalization, performing well on related but distinct test environments without additional training.
This research marks a significant step toward more capable AI systems that can handle real-world tasks requiring extended focus and multiple stages. By providing detailed, verifiable training environments and a way to reward incremental progress, the Tev framework could help unlock applications in automated research, long-term planning, and recursive self-improvement—where AI iteratively enhances its own abilities. Future work will likely explore expanding the variety of task domains and further refining reward mechanisms to improve AI autonomy and reliability on even more complex challenges.
Based on research published on arXiv by Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh et al..