Karthik Yaddanapudi

The Hierarchical Latent Workspace Model: Root-Preserving Variable-Depth Expertise for Parallel Diffusion Reasoning

The full technical report on a latent-workspace language model built on a frozen Qwen3-0.6B decoder, and the ten preregistered prototype experiments that falsified it at this scale. Routing failed causal liveness on 10 of 10 seed-runs; ablating the read-out's 34 memory positions changed graded accuracy by 0.000 on one seed and improved it by 0.022 on the other. Parity was reached by irrelevance.