Tev-Gym brings dense progress rewards to long-horizon AI agents

Tev-Gym introduced PPRS, a dense per-step reward method for multi-day agentic tasks. Its Tev-27B model, trained only on ALE computing and math domains, matches GPT-6 Astra and Claude Opus 5 at Pass@5 on ALE-CM and beats GLM 5.2 and Gemini 3.7 Flash on Terminal-Bench 3.0/4.0.

• Tev-Gym built 1,170 verified task environments from latent world specs
• Task-aware checkpointing cuts state storage by 90%
• Tev-27B trained only on ALE-CM and generalizes to Terminal-Bench
• PPRS keeps optimal policies under both dense and sparse rewards