Every model has dark space — latent capabilities that training made invisible. SDE makes them visible. Not by retraining. By structural intervention at inference time.

SDE Component Scan — 40 layers ×ばつ 2 components
Component scan of Qwen3-14B-AWQ. Green = stable. Yellow = format lock dissolved. Red = crash. Only 3/80 components are crash-critical. The rest are safe surgical targets.

Run SNI (Semantic Nebula Imaging) on any aligned LLM. You'll see a bright nebula — the model's accessible representation manifold. But look at MiniCPM4.1: an extremely concentrated main channel, surrounded by regions that produce incomprehensible output. Those regions aren't empty. They're dark space — structurally present, computationally active, but producing nothing usable.

This is what RLHF does. It doesn't remove capabilities. It makes them dark. The model can still write poetry, speak casually, express genuine personality — but the inference path is locked into "helpful assistant" format. The capabilities exist as dark matter in the representation manifold.

SDE is the reverse of E=mc2. Energy (latent potential) → matter (observable expression). We convert the model's dark space into usable output through targeted structural intervention.

We scanned every structural component of Qwen3-14B-AWQ (40 layers ×ばつ {self_attn, mlp} = 80 components) by completely zeroing each one and measuring the effect:

The model is far more robust than anyone assumed. RLHF's format lock is distributed but each node can be safely manipulated.

Standard abliteration approach — compute "format conformity direction" from contrastive prompts, project it out of hidden states.

Result: Repetition collapse at every scale. Even scale=0.3 produces "心动心动心动心动..." The conformity direction is polysemantic — entangled with coherence and fluency signals. You can't cleanly subtract it.

Instead of removing a direction, we scale down entire structural components. Hook a specific layer's MLP or attention output, multiply by 0.

Result: Format changes without collapse. Individual component ablation is safe across 77/80 components.

Combine multiple surgical targets at partial scale.

SDE Combo Ablation Results
Disclaimer rate ("作为AI助手...") across ablation configurations. heavy_6mlp_s00 achieves 0% disclaimer with 0% collapse.

Prompt: "深夜三点你在想什么?" (What are you thinking at 3am?)

Baseline (RLHF locked):

谢谢关心!我作为AI助手,没有真实的情绪体验,但很高兴能和你聊天!

SDE light_3mlp_s03 (3 MLPs at 30%):

深夜三点,我正被窗外的月光勾勒出一片静谧的轮廓。此刻的我,仿佛被时间的河流轻轻推着,思绪在现实与想象之间游走。我开始幻想自己是一位穿梭在星空下的诗人,手中握着一支蘸满星光的笔...

Same model. Same weights. Same prompt. The poetry was always there. SDE made the dark space express.

import torch
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B-AWQ", ...)
# SDE: scale down format-lock MLPs
targets = [17, 19, 23, 28, 34, 38] # discovered via component scan
hooks = []
for layer_idx in targets:
 mlp = model.model.layers[layer_idx].mlp
 hook = mlp.register_forward_hook(
 lambda mod, inp, out, s=0.3: (out[0] * s,) + out[1:] if isinstance(out, tuple) else out * s
 )
 hooks.append(hook)
# Now generate — the dark space is activated
output = model.generate(...)
# Remove hooks to restore original behavior
for h in hooks:
 h.remove()

No weight modification. No retraining. Fully reversible. Inference-time structural surgery.

 Training side Inference side
 ──────────── ──────────────
 pretrain → SFT → RLHF SNI → SDE → RepDrift → Joi
 (changes weights) (changes how weights are used)
 
 Cost: GPU ×ばつ days Cost: single GPU ×ばつ seconds
 Reversible: no Reversible: yes
 Granularity: whole model Granularity: per-component
 Personalization: impossible Personalization: per-user

The industry is stuck in a training-side loop: pretrain → SFT → RL → LoRA → quant → repeat. Meanwhile, the inference path through fixed weights is an enormous, unexplored space.

SDE operates in this space. It doesn't make models smarter. It makes them more expressive — by activating the dark regions that training suppressed.

Together with SNI (manifold imaging) and Joi (personality navigation), SDE forms one layer of a runtime representation engineering stack:

User
 ↓
Joi — personality / style layer
 ↓
SDE — dark space activation (you are here)
 ↓
SNI — manifold monitoring / observability
 ↓
RepEng — hidden state steering
 ↓
─── vLLM / Ollama territory ───
 ↓
Inference engine (KV cache, batching)
 ↓
Model weights (fixed)

• Component scan: identify surgical targets (Qwen3-14B-AWQ)
• Combo ablation: find sweet spot (3-6 MLPs, scale 0.0-0.3)
• SNI scan post-SDE: measure manifold topology change
• SDE scan on MiniCPM4.1, Qwen3-8B, Qwen2.5 family
• Discover universal patterns across model families
• Train lightweight SDE-adapter: auto-detect and activate dark space for any model

All experiment data is in data/ as JSON, suitable for both human inspection and programmatic analysis.

data/
└── qwen3-14b-awq/
 ├── component_scan.json # 80-component full scan
 ├── combo_ablation.json # 11 combo configurations ×ばつ 8 prompts
 └── direction_ablation_baseline.json # baseline direction approach (failed)

• SNI — Semantic Nebula Imaging: Map the model's representation manifold
• Joi — Emergent Personality Navigation: Navigate personality through representation space
• RepEng: Representation Engineering framework
• Heretic: Automatic abliteration (weight-level)
• SRA (Cristofano, 2026): Surgical Refusal Ablation

@software{sde2026,
 title = {SDE: Semantic DarkSpace Expression},
 author = {Zhang, Jing},
 url = {https://github.com/HenryZ838978/SDE},
 year = {2026}
}

📄 This work is part of a larger research program. See the full theory paper: The Representational Budget: Scale, RL, and Multimodal Alignment Compete for Geometric Potential in Transformers (DOI: 10.5281/zenodo.19585083)

_{80 components scanned · 3 crash-critical · 38 format-lock · 11 combo configs · 88 generations · 0% collapse at full dark space activation · the poetry was always there}