landmark attention

**Landmark Attention** is the **efficient transformer attention mechanism that reduces computational complexity by routing all token attention through a sparse set of landmark (anchor) tokens that serve as information hubs — achieving sub-quadratic attention cost while preserving global information flow** — the architecture that demonstrates how strategically placed landmark tokens can serve as a compressed global context, enabling long-sequence processing without the full O(n²) cost of standard self-attention. **What Is Landmark Attention?** - **Definition**: A modified attention mechanism where regular tokens attend only to nearby local tokens and to a set of specially designated landmark tokens, while landmark tokens attend to all other landmarks — creating a two-level attention hierarchy with O(n × k) complexity where k << n is the number of landmarks. - **Landmark Selection**: Landmarks are chosen at fixed intervals (every m-th token), at content boundaries (sentence/paragraph breaks), or through learned prominence scoring — they serve as representative summaries of their local region. - **Two-Level Attention**: (1) Local tokens attend to their neighborhood + all landmarks (sparse), (2) Landmarks attend to all other landmarks (dense but small) — global information propagates through the landmark network while local processing remains efficient. - **Information Bridge**: Landmarks act as bridges between distant sequence regions — a token at position 1 can influence a token at position 10,000 through their respective nearest landmarks, which are connected via landmark-to-landmark attention. **Why Landmark Attention Matters** - **Sub-Quadratic Complexity**: Standard attention is O(n²); Landmark attention is O(n × k + k²) where k << n — for k = √n, this becomes O(n^1.5), dramatically more efficient for long sequences. - **Global Information Preservation**: Unlike local-only attention (which loses distant context), landmark-to-landmark attention maintains a global information pathway — important for tasks requiring full-document understanding. - **Minimal Quality Loss**: Well-placed landmarks preserve 95%+ of full attention's information — the compression through landmarks retains the most important global signals. - **Compatible With Flash Attention**: The local attention windows and landmark attention patterns can be implemented efficiently with existing optimized kernels. - **Configurable Trade-Off**: Adjusting landmark density (k) provides a smooth trade-off between efficiency and information retention — more landmarks = more global information at higher cost. **Landmark Attention Architecture** **Landmark Placement Strategies**: - **Fixed Stride**: Every m-th token is a landmark — simplest, works well for uniform-density text. - **Learned Selection**: A scoring network assigns prominence scores; top-k scoring tokens become landmarks — content-aware, better for heterogeneous inputs. - **Boundary-Based**: Landmarks placed at sentence boundaries, paragraph breaks, or topic transitions — aligns with natural information structure. **Attention Pattern**: - Regular token t attends to: local window [t−w, t+w] UNION all landmarks. - Landmark l attends to: its local region UNION all other landmarks. - This creates a sparse attention pattern with guaranteed global connectivity. **Complexity Comparison** | Method | Attention Complexity | Global Context | Memory | |--------|---------------------|----------------|--------| | **Full Attention** | O(n²) | Complete | O(n²) | | **Local Window** | O(n × w) | None | O(n × w) | | **Landmark Attention** | O(n × k + k²) | Via landmarks | O(n × k) | | **Longformer** | O(n × (w + g)) | Via global tokens | O(n × (w + g)) | Landmark Attention is **the information-routing architecture that proves global context can be maintained through strategic compression** — using a sparse network of landmark tokens as information hubs that connect distant sequence regions at sub-quadratic cost, achieving the practical efficiency of local attention with the semantic capability of global attention.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account