Unverified50% confidenceBenchmarkExact time
在lx工具场景中，3B模型在JSON Schema上容易产生幻觉，无法稳定输出结构化结果
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/20/2026
First Seen
Valid until: 10/18/2026
Sources
Related Claims
UnverifiedUnder the tool-calling paradigm, the LLM outputs a structured function-call intent (in JSON format), which is actually executed by an external system that returns the result to the model.72% similarUnverified大语言模型（LLM）启用JSON Mode获取结构化输出时，输出本质上仍是文本，可能出现字段缺失、数据类型错误或模型幻觉导致的非预期字段71% similarUnverifiedThe JSON structure returned by LLM APIs may vary across model versions.70% similarUnverifiedStructured output constrains format via JSON Schema, which can eliminate randomness at the format level and confine uncertainty to the scope of semantic content.67% similarUnverifiedThe randomness introduced by 4-bit quantization may cause occasional formatting errors in Function Calling JSON output; it is recommended to explicitly require strict JSON output in the System Prompt and to add a retry mechanism at the framework layer.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/573253API
curl https://kongchang.com/api/v1/knowledge/claims/573253MCP
get_claim(id=573253)