🗂️ AI & Agents · View mindmap

Cross-Attention

Definition

Cross-attention is an attention mechanism in Transformer architectures where query ($Q$) tensors are derived from a distinct sequence compared to the key ($K$) and value ($V$) tensors. Enables information exchange between modalities or across encoder-decoder boundaries.

Mechanism

• Formula: $\text{Attention}(Q,K,V)=\text{softmax}\left(\frac{Q{K}^T}{\sqrt{d_k}}\right)V$.
• Complexity: $O(N\cdot M\cdot d)$, where $N$ is query length and $M$ is source sequence length.
• Contrast: Differs from self-attention where $Q,K,V$ originate from the same input sequence.

Applications

• Encoder-Decoder models: Decoder layers query encoder representations (e.g., T5, BART).
• Multimodal alignment: Text queries attend to image/video patches (e.g., CLIP, Flamingo, llava).
• Diffusion models: UNet blocks condition generation via cross-attention with text embeddings.
• ControlNet/Adapter: Inject auxiliary signals through cross-attention pathways.

Optimizations

• Sparse attention: Prunes or subsets attention interactions to reduce memory/compute.
• Linear attention: Kernelized approximations for near-linear scaling.
• FlashAttention: I/O-aware tiling for efficient dense attention computation.

Recent Developments

• SubQ (Subquadratic, 2026) implements a sparse attention architecture supporting a 12M token context window.
• Reports claim 52x efficiency improvements over standard dense attention baselines.
• Sparse cross-attention strategies may resolve quadratic bottlenecks in ultra-long context or high-resolution multimodal fusion.
• Verification concerns raised regarding context fidelity and efficiency metric reproducibility.
• Reference: SubQ AI: 12M Token Context, Sparse Attention Architecture, and Verification Concerns.