Unverified50% confidenceFactExact time
Directly asking a large model to score candidates often results in scores concentrated in a certain range (e.g., 7–9 points), lacking differentiation.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified研究数据表明，在数学竞赛级别问题上，启用深度推理的模型准确率可从不足30%提升至80%以上67% similarUnverified一个61分的模型和一个59分的模型在多数实际应用中的差异可能远小于分数暗示的程度66% similarUnverified靶向数据工程中，针对性数据占总微调数据的 5%-15% 通常能取得良好效果，过高比例可能导致模型在其他维度退化65% similarUnverifiedOn mainstream evaluations such as MMLU, HumanEval, and GSM8K, the score gaps among leading large models are often only a few percentage points.64% similarVerifiedIn scenarios with extreme class imbalance, even an accuracy as high as 95% may indicate that the model is completely unable to identify the minority class.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/842013API
curl https://kongchang.com/api/v1/knowledge/claims/842013MCP
get_claim(id=842013)