Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
Tev-Gym turns hidden task specifications into dense progress rewards for long-horizon agent training.
The paper introduces 1,170 verified task environments across seven domains.
Its method builds a latent “correct” workflow first, then renders agent-facing tasks from it.
Progress-potential reward shaping converts that workflow progress into per-step rewards.
The authors report Tev-27B gains on ALE-CM and generalization to Terminal-Bench subsets.
HF Daily Papers' note
The paper introduces 1,170 verified task environments across seven domains.
Its method builds a latent “correct” workflow first, then renders agent-facing tasks from it.
Progress-potential reward shaping converts that workflow progress into per-step rewards.
The authors report Tev-27B gains on ALE-CM and generalization to Terminal-Bench subsets.
HF Daily Papers' note
score 5