Voxel-Steered Autoregressive Generation: Overcoming the Quadratic Output Wall

Autoregressive language models face a strict quadratic wall: generation costs O(N²) in attention and O(N) in Key-Value (KV) cache memory. We present Voxel-Steered Autoregressive Generation, a geometric inversion of WP-20's spatial index. By implementing a “Rolling Output Eviction” loop, the model evicts recently generated chunks from the dense KV cache into a persistent HNSW index over the TERA cube, querying its own past geometry to maintain coherence. Furthermore, by predicting coarse semantic cells (Meji/Odu) before decoding discrete tokens, the model generates a trajectory through meaning-space rather than a sequence of characters. We demonstrate that this architecture allows infinite-length coherent generation with constant O(1) attention overhead and significant deduplication via Greedy Meshing. WP-21 extends WP-20 (Voxel-Addressable Memory) by inverting the retrieval mechanism from the input context space to the generative output space.