Unverified50% confidenceFactExact time
ROME 的因果追踪方法通过在 Transformer 前向传播的不同位置注入噪声，观察哪些层的激活状态对最终输出影响最大，从而定位存储特定事实的关键层
1
Sources
50%
Confidence
Long-term
Relevance
7/7/2026
First Seen
Sources
反向微调实验：AI如何被训练相信东京不是日本首都
hackernewshackernews7/5/2026
Related Claims
UnverifiedROME的核心洞见是Transformer模型的MLP层在事实知识存储中扮演关键角色，功能类似键值记忆库68% similarUnverifiedIn the Transformer architecture, factual knowledge is considered to be primarily distributed in the weight matrices of MLP (feed-forward network) layers, while the attention mechanism is responsible for dynamically retrieving and integrating this knowledge during inference.62% similarVerifiedThe self-attention mechanism of the Transformer allows the model to directly compute dependencies between any two positions in a sequence, breaking through the parallelization bottleneck caused by the sequential computation of RNNs/LSTMs.60% similarUnverifiedTransformer 的自注意力机制让模型能捕捉代码 token 之间的长程依赖关系，但其本质是条件概率分布估计器，学到的是代码看起来像好代码而非在特定业务场景下正确60% similarUnverified基于Transformer的SAKT、AKT等模型引入注意力机制来建模历史练习对当前知识状态的影响，提升了预测精度59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/179094API
curl https://kongchang.com/api/v1/knowledge/claims/179094MCP
get_claim(id=179094)