Unverified50% confidenceFactExact time
Two widely cited direct-probing benchmarks exhibited saturation, as frontier models generally gave neutral responses to straightforward questions, causing the benchmarks to detect almost no bias.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/16/2026
First Seen
Valid until: 12/15/2026
Sources
Related Entities
Related Claims
Unverified当前主流评测基准很难客观衡量顶级模型之间的真实差距，benchmark分数参考价值有限68% similarUnverified先前的难负样本挖掘工作大多忽略非干扰项，只关注与查询的相似度，因而更易误选假负样本67% similarUnverifiedLossless compression conclusions reported on contaminated benchmarks are meaningless, because the model can directly retrieve answers from parametric memory rather than relying on attention to retrieve from context.65% similarUnverified端到端模型在代表性充足与代表性不足口音之间的WER差距并未按比例缩小，属于数据分布问题而非架构问题65% similarUnverifiedThe core idea of hard negative mining is to introduce samples near the classification boundary to drive refinement of the decision boundary, enabling the model to learn deep semantic features rather than surface-level keyword patterns.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/924827API
curl https://kongchang.com/api/v1/knowledge/claims/924827MCP
get_claim(id=924827)