Unverified50% confidenceFactExact time
Models tend to give high scores to their own outputs (or outputs of similar style); Panickssery et al. found that LLM evaluators can recognize their own generated content and favor it.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBy constructing semantically equivalent but differently phrased prompt variants and observing the output distribution, one can distinguish whether sensitivity originates from the model or the evaluator.75% similarUnverifiedLLM模型在「生成」和「评估」时所激活的推理路径不同，同一模型生成代码时可能忽略某个边界条件，但在评估另一份代码时却能识别该问题73% similarUnverifiedIn LLM post-training, the model directly spews intermediate reasoning steps that should have been hidden to the user, or during RL reward computation, the reward model that should only evaluate the final answer sees the reasoning process, thereby distorting the optimization direction.72% similarUnverifiedScoring functions often require careful design and may even rely on another LLM for scoring, known as the LLM-as-judge pattern.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1066482API
curl https://kongchang.com/api/v1/knowledge/claims/1066482MCP
get_claim(id=1066482)