Agent-ValueBench: First Benchmark for Autonomous Agent Values
Agent-ValueBench has been launched by researchers as the inaugural benchmark aimed at assessing the values of autonomous agents. Previous value benchmarks have primarily focused on large language models (LLMs); however, this study reveals that the values of agents differ from those of their foundational LLMs due to their agentic nature. This benchmark tackles challenges related to datasets, evaluations, and systems that are not present in text-only frameworks. It encompasses 394 executable environments across 16 domains and includes 4,335 value-conflict tasks that span 28 value systems and 332 dimensions. Each task is meticulously co-synthesized through a specialized end-to-end pipeline and individually curated by expert annotators. The findings are available on arXiv with the identifier 2605.10365.
Key facts
- Agent-ValueBench is the first benchmark dedicated to agent values.
- It features 394 executable environments across 16 domains.
- It offers 4,335 value-conflict tasks covering 28 value systems and 332 dimensions.
- The benchmark addresses dataset-, evaluation-, and system-level challenges.
- Agent values diverge from those of their underlying LLMs.
- Each instance is co-synthesized through an end-to-end pipeline.
- Instances are curated per-instance by professional annotators.
- The research is published on arXiv under identifier 2605.10365.
Entities
Institutions
- arXiv