A 2025 paper proposes Causal Grassmann layers that represent token pairs as geometric subspaces. On WikiText-2, 13–18-million-parameter models achieved perplexities within about 10–15% of similarly sized Transformers. The SNLI comparison tested classification heads on top of DistilBERT, which still uses attention. Computation scales linearly with sequence length under fixed rank and window settings, but the current implementation can be slower than optimized GPU attention, so this is not evidence of a general replacement for large models.