Saccade Attention Networks: Using Transfer Learning of Attention to Reduce Network Sizes

Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n^2 d)$ cost of self-attention. Existing methods reduce this cost along the sequence axis through local windows, kernel approximations, or token-level sparsity, but these approaches consistently degrade accuracy. In this paper, we instead explore an orthogonal axis: feature sparsity. We propose Sparse Feature Attention (SFA), where queries and keys are represented as $k$-sparse codes that preserve high-dimensional expressivity while reducing the cost of attention from $Θ(n^2 d)$ to $Θ(n^2 k^2/d)$. To make this efficient at scale, we introduce FlashSFA, an IO-aware kernel that extends FlashAttention to operate directly on sparse overlaps without materializing dense score matrices. Across GPT-2 and Qwen3 pretraining, SFA matches dense baselines while improving speed by up to $2.5\times$ and reducing FLOPs and KV-cache by nearly 50%. On synthetic and downstream benchmarks, SFA preserves retrieval accuracy and robustness at long contexts, outperforming short-embedding baselines that collapse feature diversity. These results establish feature-level sparsity as a complementary and underexplored axis for efficient attention, enabling Transformers to scale to orders-of-magnitude longer contexts with minimal quality loss. Code is available at https://github.com/YannX1e/Sparse-Feature-Attention.

Feed-forward networks hold two thirds of a transformer's non-embedding parameters, yet the architecture has not received a necessity test that controls parameters, compute, and depth at once. We pretrain attention-only decoder transformers (Simple Attention Networks, SANs) against standard transformers matched separately for parameter count, training FLOPs, and depth (2 to 48 layers), for up to 105B tokens at 6M to 87M parameters. Deleting feed-forward layers in place is costly: the standard transformer leads by 0.47 nats at matched depth and 0.26 nats at matched FLOPs. Reallocating the freed budget into attention depth closes the gap: at matched parameters the difference is 0.006 nats (0.27 percent of loss), reproducible to one part in ten thousand across seed pairs, shrinking across 5B, 30B, and 105B budgets, and holding near 0.02 nats across a 29x size range. Three measurements localize the remaining gap to parametric recall: attention-only models are better on context-grounded answers and worse where knowledge must come from weights. Weight spectra show why: routing matrices (Q/K) crystallize early, content matrices accumulate rank slowly, and removing feed-forward layers relocates this accumulation to the attention output projection. QK-normalization, not feed-forward layers or residual gating, keeps 48-layer attention-only stacks trainable. The deficit concentrates on low-context query prediction and localizes there entirely by the largest budget. A pre-registered test confirms the account: it predicts a 0.02 to 0.05 nat gap on knowledge-dense web text; a matched pair trained on fineweb-edu measures 0.040. Within the tested regime, attention does the rest.

Standard attention scales quadratically with sequence length. Efficient attention methods reduce this O(n^2) cost, but when retrofitted into pretrained models, they often degrade perplexity, downstream accuracy, or both. We introduce Focus, a method that learns which token pairs matter. Focus adds a small set of learnable centroids--as few as 148K parameters per layer--that act as gates: only token pairs belonging to the same centroid group attend to each other over long ranges. Focus is composable: it can be added to any pretrained model by training only the centroids while keeping all original weights frozen. Experiments show that composing Focus onto pretrained models yields zero degradation on downstream benchmarks across model sizes from 124M to 70B parameters and five attention architectures. Surprisingly, sparse Focus attention outperforms full attention at 124M scale (30.3 vs. 31.4 perplexity) and matches full attention when trained from scratch at 7B scale (13.82 vs. 13.89). Focus is also fast: top-k group membership gives a 2x speedup with better quality than the original pretrained model. Using our FlashAttention decomposition, Focus achieves an 8.6x speedup at 1M tokens without custom kernels.