Query (Attention)

Key (Attention)

Value (Attention)

State Function from Previous Outputs

Value Weight Matrix Formula

Set of Sequential Key-Value Pairs

Query Vector

Key Vector

Value Vector

Implicit Relative Position Modeling in Self-Attention with RoPE

Value Weight Matrix Definition ($\mathbf{W}_j^v \in \mathbb{R}^{d \times \frac{d}{\tau}}$)

Imagine a system translating the sentence 'The quick brown fox jumps'. When the system is generating the output word corresponding to 'jumps', it needs to determine which words in the input sentence are most relevant. To do this, a vector representing the current translation context (i.e., 'what information do I need to produce the next word?') is compared against a set of searchable 'label' vectors, one for each word in the input sentence. This comparison generates a relevance score for each input word. Finally, a new vector is created by taking a weighted average of the 'content' vectors of the input words, using the relevance scores as weights. How do the three main vector types in this process correspond to their roles?

In a system designed to answer questions based on a provided document, the model first creates a representation of the user's question. It then compares this representation against a set of searchable representations, one for each sentence in the document, to determine relevance scores. Finally, it constructs an answer by creating a weighted combination of the informational content from each sentence, using the relevance scores as weights. Which option correctly assigns the roles of Query, Key, and Value vectors in this scenario?

Context Window of Key Vectors Notation

Key-Value Cache

In a computational mechanism designed to selectively focus on different parts of an input sequence, information is represented by three distinct types of vectors that interact to produce a context-aware output. Match each vector type to its specific role in this process.

Masked QKV Attention Formula

Set of Sequential Key-Value Pairs

Let a sequence of vectors be constructed where the first element is $a_1 = [0.1, 0.9]$ and the second element is $a_2 = [0.4, 0.6]$. The third element has multiple potential versions, and the 5th version is given as $a_3^{(5)} = [0.8, 0.2]$. According to the notational definition $\mathbf{a}_{\leq i+1} = (a_1, ..., a_i, a_{i+1}^{(j)})$, what is the specific sequence represented by $\mathbf{a}_{\leq 3}$ when using the 5th version of the 3rd element?

Key Matrix for Causal Attention (K_≤i)

Deconstructing Vector Prefix Notation

Key-Value Cache

Consider a sequence of vectors represented as $\mathbf{Z} = (\mathbf{z}_1, \mathbf{z}_2, \mathbf{z}_3, \mathbf{z}_4, \mathbf{z}_5)$. The notation $\mathbf{Z}_{\leq 3}$ represents the subsequence containing only the first two vectors, $(\mathbf{z}_1, \mathbf{z}_2)$.