Unverified50% confidenceBenchmarkExact time
On the two tasks requiring global planning—syntactic tree and syntactic span—Diffusion-LM surpasses the specially fine-tuned Oracle model, while autoregressive baselines perform notably worse.
1
Sources
50%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified对结构清晰的字段（如日期、金额、编号）用规则或正则处理，仅将语义模糊字段交给模型的混合管线比纯 LLM 方案更稳定高效68% 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67% similarUnverified扩散语言模型在复杂推理、指令遵循和思维链等方面面临结构性挑战，可控性弱于自回归模型66% similarUnverified提示词与模型版本存在深度绑定关系，不同版本模型在RLHF训练数据、指令微调策略上的差异会导致相同提示词产生不同输出，这种现象被称为「提示词漂移」66% similarUnverifiedSmaller models can handle fewer instructions than larger models, and non-reasoning models can handle fewer instructions than reasoning models.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843375API
curl https://kongchang.com/api/v1/knowledge/claims/843375MCP
get_claim(id=843375)