Unverified50% confidenceFactExact time
The average evaluation metrics of a new LLM may be better than those of an older model, but regressions can occur in niche critical scenarios.
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
从生产事故到评估用例：LLM回归测试的实战方法论
Related Claims
UnverifiedResearch shows that the same base LLM model can vary by 30-50 percentage points in performance under different system prompts.70% similarUnverified大模型的性能提升已经明显放缓，最近新版本的评分甚至不一定超越之前的模型，能力基本已定型69% similarUnverifiedFor small models with limited parameter counts, data quality can have an even greater impact on final performance than data quantity.68% similarUnverified传统 LLM 推理中因需按最大长度预留显存，KV Cache 的实际利用率通常不到 30%68% similarUnverifiedWei等人在2022年系统记录了大模型涌现能力现象，发现某些能力在小模型上几乎为零而在大模型上急剧提升，呈现非连续的相变特征68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/705565API
curl https://kongchang.com/api/v1/knowledge/claims/705565MCP
get_claim(id=705565)