Multi-Token Attention
This addresses a fundamental limitation in attention mechanisms for LLMs, offering a novel approach to improve information retrieval in contexts, though it appears incremental as an enhancement to existing attention methods.
The paper tackles the bottleneck in LLMs where attention weights rely on single token vectors, proposing Multi-Token Attention (MTA) to condition weights on multiple vectors via convolution, resulting in enhanced performance on language modeling and long-context tasks compared to Transformer baselines.
Soft attention is a critical mechanism powering LLMs to locate relevant parts within a given context. However, individual attention weights are determined by the similarity of only a single query and key token vector. This "single token attention" bottlenecks the amount of information used in distinguishing a relevant part from the rest of the context. To address this issue, we propose a new attention method, Multi-Token Attention (MTA), which allows LLMs to condition their attention weights on multiple query and key vectors simultaneously. This is achieved by applying convolution operations over queries, keys and heads, allowing nearby queries and keys to affect each other's attention weights for more precise attention. As a result, our method can locate relevant context using richer, more nuanced information that can exceed a single vector's capacity. Through extensive evaluations, we demonstrate that MTA achieves enhanced performance on a range of popular benchmarks. Notably, it outperforms Transformer baseline models on standard language modeling tasks, and on tasks that require searching for information within long contexts, where our method's ability to leverage richer information proves particularly beneficial.