Unverified50% confidenceFactExact time
扩散语言模型天然支持双向注意力，每个位置可以同时看到左侧和右侧的上下文，而GPT类模型受因果掩码限制只能看到左侧
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
UnverifiedTransformer注意力层存在位置偏置，模型对序列起始位置(primacy bias)和末尾位置(recency bias)的关注天然高于中间段落65% similarUnverifiedThe self-attention mechanism is capable of capturing dependencies between any two positions in a sequence.63% similarUnverifiedUltra采用宽树范式，多个独立推理上下文并行广度优先探索问题空间的不同子域，彼此不共享中间状态，最终由主代理聚合61% similarUnverified外部分类器依赖表面特征匹配，与语言模型缺乏深度语义共享，难以真正理解上下文意图60% similarUnverifiedIn the T-shaped talent model, the horizontal bar represents broad cross-domain knowledge and perspective, while the vertical bar represents deep expertise in a specific field.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/542324API
curl https://kongchang.com/api/v1/knowledge/claims/542324MCP
get_claim(id=542324)