Qwen3.8-27B-Weight-Space-Programming (WSP v2.4 — 4-Subspace Mid + Late Architecture)

Training-free, closed-form Multi-Subspace Weight-Space Programming (WSP) for Qwen/Qwen3.8-27B (64 layers, hidden_size = 5120, intermediate_size = 17408, $16 \times [3\times\text{Gated DeltaNet} + 1\times\text{Gated Full Attention}]$) — zero gradient descent, zero fine-tuning, synthesized in 47 seconds on a single workstation.

---

1. Key Highlights & 4-Way Factorial Ablation (123 Live Benchmark Tasks)

1. Root-Cause Spectral Discovery (probe_qwen38_subspaces.py): Across the upper Full-Attention and Pre-Terminal layers ([47, 51, 55, 59, 62, 63]) of Qwen/Qwen3.8-27B, the top-64 singular directions of mlp.down_proj exhibit a 28.1% to 45.0% stronger cosine alignment with the self-doubt/hesitation subspace (Wait, Should, Or, Maybe, Alternatively, verify, test) than with the </think> termination token (248069). This single spectral bias explains both primary baseline failure modes:
   • Circular <think> deliberation loops that exhaust token budgets on InterCode-NL2Bash, GSM8K, HumanEval, and Anti-Sycophancy (Suite D).
   • Post-solution self-doubt ("Should I verify it? Let me run a simple test") triggering unnecessary exec tool calls on pure coding tasks (Suite E: restraint_1_reverse_ll).
2. Exact SwiGLU Receptor & Woodbury Dual-Space Ridge Surgery (build_qwen38_subspaces_v2.py, build_midlayer_subspace4.py): Using exact SwiGLU receptor states $\hat{x}j = \text{unit}(\text{SiLU}(W{\text{gate}} \hat{h}j) \odot (W{\text{up}} \hat{h}j))$ and the Woodbury Dual-Space Nullspace-Anchored Ridge Identity: $$\Delta W_s^{(l)} = \tilde{Y}{\text{aug}}^\top \left(X_{\text{aug}} X_{\text{aug}}^\top + \lambda_{\text{reg}} I_m\right)^{-1} X_{\text{aug}}, \qquad X_{\text{aug}} = \begin{bmatrix} X_s \ \sqrt{\lambda_{\text{null}}}, X_{\text{null}, s} \end{bmatrix} \in \mathbb{R}^{m \times 17408}$$ we invert an $m \times m$ ($160 \times 160$) Gram matrix instead of a dense $17408 \times 17408$ matrix, cutting synthesis time by 30× (1419.0s $\rightarrow$ 30.8s–47.7s) and reducing cross-subspace/nullspace leakage by 25× (to 0.11%–0.31% across all 13 edited layers).
3. 2×2 Factorial Ablation (None vs. Dynamic Late-Layer Routing Only vs. Mid-Layer Full-Attention Edit Only vs. Both):

*Evaluated with normalized JSON Schema ("type": "float" $\rightarrow$ "number" in BFCL v3 Irrelevance); raw unnormalized baseline is 102/123 (82.9%).

---

2. Granular Per-Suite Benchmark Breakdown (123 Tasks)

---

3. Repository Contents & Artifacts

---

4. Quickstart Usage (llama.cpp / llama-server)

Launch llama-server with all 4 orthogonal subspace GGUF LoRA adapters:

llama-server -m Qwen3.8-27B-UD-Q4_K_M.gguf -c 4096 -ngl 46 --port 8088 --host 127.0.0.1 \
  --lora wsp_subspace1_think.gguf \
  --lora wsp_subspace2_restraint.gguf \
  --lora wsp_subspace3_bash.gguf \
  --lora wsp_subspace4_mid.gguf

Then dynamically route subspace scales in 1 ms via POST http://127.0.0.1:8088/lora-adapters (see scripts/run_4way_ablation.py for the <0.1 ms dynamic_route_scales() router):

curl -X POST http://127.0.0.1:8088/lora-adapters \
  -H "Content-Type: application/json" \
  -d '[{"id":0,"scale":1.0},{"id":1,"scale":1.0},{"id":2,"scale":1.0},{"id":3,"scale":0.85}]'