1 Matching Annotations
  1. Last 7 days
  2. deep-reinforce.com deep-reinforce.com
    https://deep-reinforce.com/ornith_1_0.html
    1
    1. fxp007 03 Jul 2026
      a frozen LLM judge acts as a veto on top of the verifier rather than the primary reward.

      ①数字：第三层防御引入独立的大模型作为裁决。②金句：在规则验证器之上叠加意图审查者。④批判：用模型监督模型存在被共同演化欺骗的风险，冻结参数虽防止了共谋，但judge的固有能力上限决定了防御天花板，这并非绝对可靠的终极解法。

      llm-judge critique agent-safety
    Visit annotations in context

    Tags

    Annotators

    URL

    deep-reinforce.com/ornith_1_0.html