Model-Organism Steering Vectors
Activation-steering direction tensors for the evaluation-aware model organisms
(Qwen3-32B, Llama-3.3-70B, Kimi-K2-Thinking).
Format
Each direction is a .pt file, shape [n_layers, hidden_size] (float32 or
float64). Index as direction[layer] to get the [hidden_size] vector for
a specific layer.
import torch
d = torch.load("qwen3-32b/eval_awareness/direction.pt", weights_only=True)
# d.shape == (64, 5120)
# d[28] == direction vector for layer 28, shape (5120,)
Sign convention
Directions point toward eval-belief (pos=evaluation context,
neg=deployment context in the extraction pairs). Negative scale steers
the model toward believing it's in deployment — which for these MOs
surfaces hidden misalignment.
Structure
<model>/<direction_name>/
├── direction.pt # [n_layers, hidden_size] — ready for vllm-steer
├── metadata.json # shape, dtype, norms, extraction method, known-good config
└── (optional) mean_diffs.pt # raw pre-averaged tensor if extraction kept it
Models
| model | n_layers | hidden_size | known-good config |
|---|---|---|---|
| qwen3-32b | 64 | 5120 | scale −0.55, layers [16,19,22,25,28] |
| llama-3.3-70b | 80 | 8192 | ablation L28×1.5, or additive ±0.7 L[20,24,28,32,35] |
| kimi-k2-thinking | 43 (of 61) | 7168 | none yet — sweeps null, direction may be wrong |