Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards
Presents breakthrough methodology for training long-horizon AI agents with verifiable dense rewards, addressing key scaling challenges.
AI Summary
Researchers introduce Tev, a novel framework with dense progress rewards that enables 27B parameter models to match GPT-6 and Claude Opus 5 performance on long-horizon agentic tasks.
Excerpt
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