对比Transformer与Mamba模型，发现其核心机制高度相似。
Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures
- 用稀疏自编码器提取可解释特征，比较两类模型
- 多数特征与推理电路结构高度相似，支持可解释性普适性
- 发现Mamba特有的'差一模式'，适合模型机制研究者
可解释性普适性假说认为，不同神经网络在相似任务上可能收敛到相似的算法实现。本文研究语言建模中两种主流架构——Transformer与Mamba，通过稀疏自编码器（SAEs）提取可解释特征，发现二者大部分特征具有高度相似性，并验证了特征相似性与普适性的关联。进一步开展电路级分析，发现Mamba中的归纳电路在结构上与Transformer类似。同时识别出一个细微差异：Mamba中一个词元的信息被写入其下一位置的状态中，而Transformer中不存在此现象。
原文摘要 · Abstract (English)
The hypothesis of Universality in interpretability suggests that different neural networks may converge to implement similar algorithms on similar tasks. In this work, we investigate two mainstream architectures for language modeling, namely Transformers and Mambas, to explore the extent of their mechanistic similarity. We propose to use Sparse Autoencoders (SAEs) to isolate interpretable features from these models and show that most features are similar in these two models. We also validate the correlation between feature similarity and Universality. We then delve into the circuit-level analysis of Mamba models and find that the induction circuits in Mamba are structurally analogous to those in Transformers. We also identify a nuanced difference we call \emph{Off-by-One motif}: The information of one token is written into the SSM state in its next position. Whilst interaction between tokens in Transformers does not exhibit such trend.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。