Unverified50% confidenceFactExact time
Large language models' understanding of JSON structure is probabilistic, and when processing longer structured outputs, type degradation occurs due to attention mechanism weight decay on earlier context, such as serializing arrays as strings or writing numbers as strings.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified提示词工程通过角色定义、输出格式约束（如JSON Schema）、少样本示例等手段约束模型概率性输出71% 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71% similarVerifiedWhen large language models process long contexts, the computational complexity of the attention mechanism grows proportionally to the square of the sequence length.67% similarUnverified大语言模型采用自回归生成机制，在生成token时使用单向前向注意力，无法进行全局类型校验，易产生API幻觉67% similarUnverified语言的语法结构越简单、惯用模式越统一，大模型在每一步预测中面临的困惑度越低，生成正确序列的概率越高65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/830402API
curl https://kongchang.com/api/v1/knowledge/claims/830402MCP
get_claim(id=830402)