HarnessEval-W：将世界模型评估代理化

AI 摘要

HarnessEval-W 是一个用于世界模型评估的代理化流水线，通过分层子代理将评估分解为可验证的推理链，并以透明证据证明评分。该流水线应用于18个世界模型的330个评估案例，其判断与人类偏好高度一致，同时提供可验证的细粒度诊断。作者已开源完整流水线作为实时基准，邀请社区贡献新的技能和评估案例。

正文节选

HarnessEval-W: Agentifying the Evaluation of Visual Worlds Abstract HarnessEval-W uses hierarchical sub-agents to decompose world-model evaluations into verifiable reasoning chains that justify scores with transparent evidence. A benchmark should deliver more than a scalar score: what makes an evaluation trustworthy is the reasoning that justifies the score. This is especially critical for world models, where judging a rollout requires understanding whether physics, causality, and world state ev

发布时间：—

抓取时间：2026-08-18 16:56

来源机构：Hugging Face