(1/3) A Primer on the Inner Workings of Transformer-based LMs

https://arxiv.org/pdf/2405.00208v3

(Oct 2024)

Figure 1: Survey overview.  Section 2  introduces the Transformer language model and its components.  Section 3  and  Section 4  present interpretability techniques used to analyze models’ inner workings. Finally,  Section 5  presents known inner workings of Transformer language models.

---

---

Figure 2: Unrolled Transformer LM with expanded views of the Attention and Feedforward network blocks, including model weights (gray) and residual stream states (green). Based on figures from  (Ferrando & Voita, 2024; Voita et al., 2023) .

---

Figure 3: Forward pass decomposition in a simplified Transformer LM. The direct path (red), full OV circuits (yellow) and virtual attention heads (grey) expressed in  Equation 11  are highlighted.

---

---

---

---

---