Long agent tasks gain verified progress feedback
In plain words
Tev-Gym is a collection of long computer and math tasks for training software agents. Each task is built from a known-correct answer, so the authors can check both whether an agent finishes and how much correct work it has done along the way. Training rewards progress as well as finishing.
On 18 computing-and-math tasks, the trained 27B model fully solved a larger fraction in at least one of five tries than its starting model: 0.28 compared with 0.17. This result applies to that task set.
The research
When groups of attempts at long computer tasks all fail, a pass-or-fail score treats them alike even if some have done more useful work. Can training distinguish those attempts while keeping an exact check for whether the task was completed?
Tev-Gym builds tasks from verified hidden answers, then uses each agent’s progress through them as extra training feedback alongside final success.
The authors generate task files, strict pass-or-fail checks, and intermediate progress scores from one known-correct description of each task.
They train a 27B model with those scores and compare it with its starting model on 18 computing-and-math tasks from Agents’ Last Exam, with five attempts per task.
The best-of-five solve rate, the fraction of tasks fully solved by at least one attempt, rises from 0.17 to 0.28. This supports the training recipe on that task split; a separate 9B ablation also finds a higher mean graded score with progress rewards when the training data is fixed.
The study focuses on synthetic terminal environments and selected task sets; the models still trail leading frontier agents on the hardest tests.
Results do not establish performance across all long-horizon workflows, and supporting new workflows requires more environment components to be authored and validated.
Analysis
Deep
- Original paper authors
- Wenbo Hu, Shraman Pramanick, Krishna Kumar Singh, Nanyun Peng, Kai-Wei Chang
- Submitted to arXiv
- License
- CC BY 4.0