Gespeichert in:
| Hauptverfasser: | , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | https://arxiv.org/abs/2505.24244 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866909628415082496 |
|---|---|
| author | Endy, Nir Grosbard, Idan Daniel Ran-Milo, Yuval Slutzky, Yonatan Tshuva, Itay Giryes, Raja |
| author_facet | Endy, Nir Grosbard, Idan Daniel Ran-Milo, Yuval Slutzky, Yonatan Tshuva, Itay Giryes, Raja |
| contents | This paper investigates the flow of factual information in Mamba State-Space Model (SSM)-based language models. We rely on theoretical and empirical connections to Transformer-based architectures and their attention mechanisms. Exploiting this relationship, we adapt attentional interpretability techniques originally developed for Transformers--specifically, the Attention Knockout methodology--to both Mamba-1 and Mamba-2. Using them we trace how information is transmitted and localized across tokens and layers, revealing patterns of subject-token information emergence and layer-wise dynamics. Notably, some phenomena vary between mamba models and Transformer based models, while others appear universally across all models inspected--hinting that these may be inherent to LLMs in general. By further leveraging Mamba's structured factorization, we disentangle how distinct "features" either enable token-to-token information exchange or enrich individual tokens, thus offering a unified lens to understand Mamba internal operations. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_24244 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Mamba Knockout for Unraveling Factual Information Flow Endy, Nir Grosbard, Idan Daniel Ran-Milo, Yuval Slutzky, Yonatan Tshuva, Itay Giryes, Raja Computation and Language Machine Learning This paper investigates the flow of factual information in Mamba State-Space Model (SSM)-based language models. We rely on theoretical and empirical connections to Transformer-based architectures and their attention mechanisms. Exploiting this relationship, we adapt attentional interpretability techniques originally developed for Transformers--specifically, the Attention Knockout methodology--to both Mamba-1 and Mamba-2. Using them we trace how information is transmitted and localized across tokens and layers, revealing patterns of subject-token information emergence and layer-wise dynamics. Notably, some phenomena vary between mamba models and Transformer based models, while others appear universally across all models inspected--hinting that these may be inherent to LLMs in general. By further leveraging Mamba's structured factorization, we disentangle how distinct "features" either enable token-to-token information exchange or enrich individual tokens, thus offering a unified lens to understand Mamba internal operations. |
| title | Mamba Knockout for Unraveling Factual Information Flow |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2505.24244 |