SeekJudge: A New Reward Framework for Computer-Use Agents
There's a new research paper out that introduces SeekJudge, a groundbreaking reward system designed for reinforcement learning in AI agents. This system addresses the challenge of evaluating if a path aligns with instructions in long-term GUI tasks. Traditional rule-based methods often miss the mark on human intent and can become outdated with app updates. Current model-based judges haven't quite reached the effectiveness of rule-based ones. SeekJudge employs four unique agents—Condense, Ground, Seek, and Analyze—that work together in a Seek-Analyze loop to assess outcomes. A specially calibrated distillation process trains a combined 9B model to serve as the foundation for these agents. Remarkably, SeekJudge is the first model-based reward system to match or exceed rule-based evaluations based on downstream success rates. You can find the paper on arXiv under the ID 2607.23263.
Key facts
- SeekJudge is a reward framework for reinforcement learning in computer-use agents.
- It uses four role-specialized agents: Condense, Ground, Seek, and Analyze.
- The agents reach a verdict through a Seek-Analyze loop over the trajectory.
- A seed-calibrated distillation pipeline trains a single 9B model as shared backbone.
- SeekJudge matches or exceeds rule-based evaluation on held-out RL test goals.
- It is the first practical model-based reward to achieve this.
- Rule-based evaluation struggles with aligning to human intention and becomes outdated.
- Existing model-based judges still lag behind rule-based evaluation.
Entities
Institutions
- arXiv