Unverified50% confidenceFactExact time
Wei et al. (2023) pointed out that small language models ignore flipped labels in the context and rely mainly on pretrained semantic priors, and that overriding these priors is an emergent capability brought by model scale.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAn uncensored model is a model version in which the safety alignment restrictions injected during the training phase of a large language model are removed or weakened through fine-tuning.74% similarUnverified预训练阶段的下一词预测目标让模型倾向生成自然流畅序列，而JSON等结构化格式在通用网络语料中占比有限，导致模型偶发格式违规74% similarUnverified主流LLM的预训练采用下一个token预测作为目标，优化的是语言建模损失而非事实准确性73% similarUnverified2022 年 Wei 等人在《Emergent Abilities of Large Language Models》中记录，当模型规模跨越某一参数阈值时性能会出现类似物理相变的非线性跃升72% similarVerifiedLarge language models rely solely on parameter weights fixed during training for autoregressive generation at inference time, and have two inherent limitations: a knowledge cutoff and capability boundaries72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1029244API
curl https://kongchang.com/api/v1/knowledge/claims/1029244MCP
get_claim(id=1029244)