Attention mechanisms in large language models (LLMs) enable the model to dynamically focus on different parts of the input sequence when processing each token. Unlike older architectures like RNNs, which process tokens sequentially and struggle with long-range dependencies, attention assigns weights to input tokens to determine their relevance to the current position. This is achieved through three components: queries, keys, and values. The query represents the current token being processed, keys are vectors for all input tokens, and values contain the actual information from those tokens. By comparing the query to all keys, the model calculates attention scores, which determine how much focus each value receives. This allows the model to prioritize contextually important tokens, even if they are far apart in the sequence.
A key implementation is the self-attention mechanism in Transformers. Here, the input is split into queries, keys, and values using learned linear transformations. For example, in the sentence “The cat sat on the mat because it was tired,” when processing the word “it,” the model computes attention scores between “it” (query) and all other words (keys). The scores are normalized via softmax to create a probability distribution, and the resulting weights are applied to the values (the word embeddings) to produce a context-aware representation. This weighted sum helps the model recognize that “it” refers to “cat” rather than “mat.” The process is repeated for every token, allowing the model to build rich, context-sensitive representations by aggregating relevant information across the entire sequence.
Practical variations include multi-head attention, which runs multiple parallel self-attention layers to capture different types of relationships (e.g., syntactic and semantic). For efficiency, masked attention is used in decoder layers to prevent the model from “peeking” at future tokens during training. However, the quadratic computational cost of attention (scaling with sequence length²) remains a challenge. Techniques like sparse attention (limiting the tokens each position can attend to) or chunking sequences into smaller blocks are common optimizations. These mechanisms collectively enable LLMs to handle complex tasks like translation or summarization by dynamically aligning inputs and outputs while maintaining computational feasibility for real-world applications.