超越合理预测：网页世界模型中的安全充分性
原标题：Beyond Plausible Prediction: Safety Sufficiency in Web World Models
AI 摘要
该论文提出「安全充分性」概念，指出网页世界模型即使能生成语义上合理的预测（3000条预测中86.1%被判定为合理），也未必包含做出正确安全授权决策所需的关键信息。作者构建了SafeWebWorld基准，覆盖五个网页安全领域，联合评估动作条件未来、策略决定性证据以及Execute/Ask/Block三分类决策。研究发现，在安全对齐监督微调下，证据保留能力提升的同时授权敏感决策反而恶化（称为「授权-策略坍缩」），并引入Latent Authorisation Recovery方法分析该现象，结论是合理预测、证据保留与聚合恢复各自都不足以建立稳健的授权行为。
正文节选
Beyond Plausible Prediction: Safety Sufficiency in Web World Models Abstract Web world models increasingly predict the consequences of candidate actions before agents act, yet a plausible predicted future does not necessarily support the correct safety decision. In a matched human evaluation of 3,000 predictions from ten frozen open models, 86.1% are judged semantically plausible. However, several models exhibit highly asymmetric authorisation behaviour, combining low false-safe execution with p