SkillTV-Bench
What it is
SkillTV-Bench is a benchmark for evaluating how well LLM judges can verify long-horizon agent executions by inspecting trajectories, skills, and artifacts. It includes a reusable JudgeSkill that guides agents to plan targeted inspections and issue evidence-grounded verdicts, improving verification accuracy by 14.8 percentage points.
The problem it solves
Standard LLM judges often fail to detect subtle failures in long-horizon agent executions because they only check final outputs rather than inspecting the full trajectory and available skills.
What you could use it for
- Evaluating the reliability of LLM judges on complex agentic tasks
- Selecting the best trajectory from multiple agent rollouts
- Developing and refining judge skills for automated verification
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.