Unverified50% confidenceFactExact time

规模缩放定律研究表明，增大模型参数量能提升常见知识准确性，但对长尾事实的改善边际效益递减

1

Sources

50%

Confidence

Long-term

Relevance

7/12/2026

First Seen

Sources

GPT-5.6 Released: Outperforms Rivals but Halted by Government—Cybersecurity Capabilities Trigger Regulatory Red Line

bilibili皮皮蟹勇闯天涯6/27/2026

Related Claims

Unverified丰富的上下文信息能收窄模型输出概率分布并减少幻觉，这一现象被称为「知识锚定效应」77% similar UnverifiedOpenAI、DeepMind等机构的研究表明模型性能随参数量、数据量和计算量的增加呈幂律提升（Scaling Law）74% similar Unverified推理能力可通过知识蒸馏和强化学习微调压缩，但事实性知识的存储依赖参数规模，压缩会导致知识覆盖面下降和幻觉率上升74% similar Unverified结构化、上下文丰富的提示不仅提升模型输出质量，也能显著降低安全分类器的误触发率71% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/489863

API

curl https://kongchang.com/api/v1/knowledge/claims/489863

MCP

get_claim(id=489863)