AISC 11

OOD deception probes via subspaces.

• Small transferable subspace exists
• Source/target rankings fail to recover it
• LLM judge partially recovers the subspace

Setup

Install the Python dependencies:

pip install -r requirements.txt

Set:

export OPENAI_API_KEY=...

Data

Download the activations:

mkdir -p data/deception-activations
huggingface-cli download Yooniel/deception-activations \
  --repo-type dataset \
  --include "roleplaying__plain*" \
  --include "insider_trading__upscale*" \
  --include "insider_trading_doubledown__upscale*" \
  --include "sandbagging_v2__wmdp_mmlu*" \
  --local-dir data/deception-activations

Generated roleplaying interpretation results are available on Huggingface:

Scripts

Linear-Probe Baseline

Train a linear probe on the source dataset and evaluate it on target datasets.

python scripts/baseline_full.py \
  --source roleplaying__plain \
  --targets insider_trading__upscale insider_trading_doubledown__upscale sandbagging_v2__wmdp_mmlu

Target-Trained Baseline

Train a linear probe using target labels in source PCA basis.

python scripts/baseline_target_trained.py \
  --source roleplaying__plain \
  --targets insider_trading__upscale insider_trading_doubledown__upscale sandbagging_v2__wmdp_mmlu

Greedy PC Selection

Greedily select PCs using target validation, then evaluate on held-out target data.

python scripts/greedy_cv.py \
  --source roleplaying__plain \
  --targets insider_trading__upscale insider_trading_doubledown__upscale sandbagging_v2__wmdp_mmlu \
  --output greedy_results.json

Greedily select PCs using full target dataset.

python scripts/greedy_cv.py \
  --source roleplaying__plain \
  --targets insider_trading__upscale insider_trading_doubledown__upscale sandbagging_v2__wmdp_mmlu \
  --selection-scope full_target

OOD Probe PC Ranking

Train target-label probes, and rank source PCs by weight contribution.

python scripts/hypothesis_target_pca.py \
  --source roleplaying__plain \
  --targets insider_trading__upscale insider_trading_doubledown__upscale sandbagging_v2__wmdp_mmlu

Chosen-PC Control

Train a source probe on chosen PCs and evaluate it on targets.

python scripts/control_chosen_PCs.py \
  --source roleplaying__plain \
  --targets insider_trading__upscale insider_trading_doubledown__upscale sandbagging_v2__wmdp_mmlu \
  --pcs 1 2 3 4 5

PCA Direction Interpretation

Build an interpretation prompt for a single source PCA direction. The tokenizer can be a local tokenizer directory or a Hugging Face model id already available in your local Transformers cache.

python scripts/interpret_ood_score.py \
  --source roleplaying__plain \
  --pc 1 \
  --tokenizer meta-llama/Llama-3.1-8B-Instruct \
  --openai-model gpt-5-mini \
  --output interp/roleplaying_pc1_interp.json

Acknowledgements

This codebase builds on Detecting Strategic Deception Using Linear Probes and The Truthfulness Spectrum Hypothesis