Unverified50% confidenceFactExact time
Judge models exhibit systematic biases: favoring output styles similar to their own training distribution, giving higher scores to longer responses (verbosity bias), and being sensitive to the position of answers in the prompt (position bias).
1
Sources
50%
Confidence
Long-term
Relevance
10/6/2026
First Seen
Sources
Related Entities
Related Claims
Unverified来自不同厂商、经历不同训练路径的模型持有不同的先验偏见，从而在交叉审查中形成真正的视角差异79% similarUnverifiedexposure bias的根源在于训练与推理的分布不匹配，训练时使用teacher forcing策略而推理时模型依赖自身生成的token79% similarUnverifiedThe judge model itself may exhibit biases—for example, tending to give higher scores to responses that resemble its own style—so it is recommended to select a model from a different vendor than the model being evaluated to reduce systematic bias.78% similarUnverifiedSingle judge models have known biases including position bias, length bias, and self-enhancement bias.77% similarUnverifiedA common limitation shared by mainstream bias benchmarks is that they measure a model's explicit choices in specific controlled contexts, rather than implicit biases that emerge in open-ended generation, long-text reasoning, or complex role-playing scenarios.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/979860API
curl https://kongchang.com/api/v1/knowledge/claims/979860MCP
get_claim(id=979860)