Extending Activation Steering to Broad Skills and Multiple Behaviours

Paper Summary

1. Broad skills — e.g., general coding ability vs. Python-specific ability
2. Multiple behaviours — myopia, wealth-seeking, agreeableness, anti-immigration, sycophancy

• Steering broader skills is competitive with steering narrower skills
• Combining steering vectors into one vector is largely unsuccessful
• Injecting individual steering vectors at different layers simultaneously is promising

Repository Structure

.
├── broad_steering.py       # Experiment 1: broad steering (Section 2.1, 3.1)
├── multi_steering.py       # Experiment 2: multi-steering (Section 2.2, 3.2)
├── run_experiments.py      # End-to-end runner for both experiments
├── test_pipeline_small.py  # Quick CPU test with GPT-2
└── README.md

Requirements

• Python 3.10+
• PyTorch
• Transformers
• Datasets
• NumPy
• tqdm

Usage

Quick Test (CPU, GPT-2)

Full Experiments (GPU, Llama-2-7b)

Experiment Details

Experiment 1: Broad Steering

• Model: Llama-2-7b-chat-hf (bfloat16)
• Datasets:
  • Text: Pile (non-code subsets)
  • Code: Pile (Github, StackExchange)
  • Python: Pile (Github filtered for Python code)
• Method: Compute steering vectors by averaging last-token activations. Subtract code vector + add text vector.
• Evaluation: Top-1 next token prediction accuracy on 500k text tokens and 500k code/Python tokens
• Injection coefficients: 0.0–50.0 (see Appendix A.2)
• Layers evaluated: 0, 5, 10, 15, 20, 25, 29, 31

Experiment 2: Multi-Steering

• Behaviors: myopia, wealth-seeking, agreeableness, anti-immigration, sycophancy (political)
• Dataset: Anthropic's model-written-evals
• Method: Contrastive Activation Addition (CAA) — matching answer activations minus non-matching answer activations
• Individual steering: Grid search over coefficients {0.5, 1, 2, 3, 5, 10, 20, 30, 40, 60, 80, 120, 200, 300} and layers
• Combined steering: 8 combinations of mean/sum, weighted/unweighted, add/subtract
• Simultaneous steering: Inject each behavior's vector at a different layer with a global coefficient
• Evaluation: Matching score (fraction of answers matching target behavior), with mode-collapse and validity checks

Key Implementation Notes

1. Activation extraction: Last token in the residual stream (most contextual information due to causal attention)
2. No normalization: Steering vectors are NOT normalized, allowing similar coefficients across layers
3. No sampling: All generation is greedy (no sampling) per the paper
4. Permuted baselines: Permuted steering vectors maintain mean/std but distort activation order
5. Discard criteria: >5% invalid answers OR >95% mode collapse → discard hyperparameter combination

Results

• broad_steering_results.json — relative coding/text accuracy per layer and coefficient
• multi_steering_results.json — individual, combined, and simultaneous steering scores
• steering_vectors/ — saved numpy arrays of steering vectors

Citation

License