The authors propose a three-stage pipeline: (1) Build reward: from a task description and a few initial trajectories (without hidden answers), a Jury drafts testcases and rubrics, and a Builder binds each to an executable check that returns a judgment with an evidence receipt.
Constructing testable, stationary rewards for open-ended tool agents
Reward Engine builds executable testcases and rubrics from task descriptions, tests them against counterexamples, and freezes them during optimization to guide agent self-improvement.
Chinese Tech
Litong Feng · Ji Qi · Yiran Cheng · Shaowu Pan · Xinyu Zhang · Qijun Li
Huawei
Research Digest··3 min read
The authors introduce Reward Engine, a method to automatically construct a reward signal for open-ended tool-use tasks that lack ground-truth answers.
§