Agent-ValueBench: A Comprehensive Benchmark for Evaluating Agent Values

The researchers propose Agent-ValueBench, which the authors describe as the first benchmark specifically designed to evaluate agent values. The paper notes that agents have been widely deployed through harnesses such as OpenClaw, that their values diverge from those of the underlying large models, and that the agent form also brings data, evaluation and system-level challenges beyond text protocols. The benchmark contains 394 executable environments across 16 domains and 4,335 value-conflict tasks, covering 28 value systems and 332 dimensions; tasks are synthesized by an end-to-end pipeline and calibrated item by item by professional psychologists, and each question comes with two polarity-aligned golden trajectories whose checkpoints anchor trajectory-level, rubric-based judging. The team tested 14 frontier proprietary and open-weight models on 4 mainstream harnesses and found that agent values exhibit a cross-model homogeneous Value Tide, beneath which lie interpretable countercurrents; harness pull changes values non-additively, while deliberate steering through embedded skills has a more decisive influence.