A new paper tackles one of the persistent bottlenecks in training autonomous AI agents: the scarcity of data and reward signals for tasks that unfold over long sequences of steps. While long-horizon agents capable of working independently for hours or days are seen as a prerequisite for automated scientific discovery and recursive self-improvement, open-source models have lagged behind because such tasks are hard to curate, hard to verify, and typically only reward an agent once the entire task is complete.
The authors address this with Tev-Gym, described as the first large-scale training dataset built specifically for long-horizon agents with dense, verifiable rewards. It contains 1,170 verified task environments spanning seven domains, constructed through what the team calls a latent-world-first method. Instead of designing tasks directly, they first define a complete hidden specification of a domain, including the expected workflow and state after every step, and then progressively render that specification into an environment the agent can interact with. This approach allows task difficulty to be tuned deliberately and provides a built-in metric for how much progress an agent has made at any point.
That progress metric is then converted into a practical training signal through a method the paper names progress-potential reward shaping, or PPRS. Rather than giving the agent feedback only at the very end of a task, PPRS supplies a dense, per-step reward tied to incremental progress, which the authors argue makes training on otherwise intractable long tasks both feasible and more efficient.
Applying this recipe produces a 27-billion-parameter model, Tev-27B, which the researchers report achieves notable improvements on the computing and math-science subsets of the Agents' Last Exam benchmark, reaching performance comparable to larger frontier systems referenced in the paper at a Pass@5 evaluation setting. Despite being trained only on data related to that benchmark, the model also generalizes to separate Terminal-Bench evaluations, outperforming some competing systems on harder subsets and approaching frontier-level results on medium-horizon tasks.
The work is a reminder that progress on agentic AI is increasingly bottlenecked less by model architecture and more by the quality of training environments and reward design. If the generalization results hold up under independent scrutiny, dense progress rewards derived from structured latent task specifications could become a standard ingredient for training agents that need to operate reliably over extended, multi-step workflows.