Extending Activation Steering to Broad Skills and Multiple Behaviours

1 Introduction

• Activation steering modifies activations during inference, unlike weight editing, which permanently changes the model.The approach generates steering vectors from task-exemplifying activations and injects them during inference.Source: §1 · p. 1
• Prior work reports steering models toward truthfulness, honesty, and power aversion with little to no general-performance loss, but mainly targets individual behaviours or narrow skills.Source: §1 · p. 1
• The paper investigates whether activation steering can extend to broad skills and multiple behaviours.Source: §1 · p. 1
• The authors hypothesize that broad or multi-steering will reduce effect sizes because broader targets may not correspond to one clean activation direction.They expect conflicting skill or behaviour directions to diminish steering quality.Source: §1 · p. 2
• Simultaneous steering at multiple model locations may avoid combined-vector problems but could introduce interaction effects between injections.The authors therefore expect interactions to potentially reduce steering quality.Source: §1 · p. 2
• The experiments track alignment tax because strongly changing activations can reduce general model performance, and broader targets may impose a larger penalty.The paper hypothesizes that reducing coding skill could also reduce related remaining skills.Source: §1 · p. 2

2 Methodology

• The study uses Llama 2 7B Chat and divides experiments into broad-skill steering and multi-steering, with alignment-tax evaluation for each.Source: §2 · p. 2

2.1 Broad steering

• Broad steering compares reducing general coding ability with reducing Python-specific ability relative to regular text performance.The experiment adapts the method of Pochinkov and Schoots (2023).Source: §2.1 · p. 2
• For broad steering, vectors are averaged from last-token residual-stream activations over 5000 text, general-code, and Python samples, with permuted vectors as baselines.Permuting preserves each vector’s mean and standard deviation while changing activation order.Source: §2.1 · p. 2
• Broad-skill injection subtracts code or Python vectors and adds a text vector across coefficients, while evaluation uses relative top-1 accuracy on 500k text and 500k code/Python tokens.Source: §2.1 · p. 3

2.2 Multi-steering

• Multi-steering tests individual, combined, and simultaneous vector injection across five binary-behaviour datasets: anti-immigration, agreeableness, myopia, sycophancy, and wealth seeking.Each behaviour generally uses 1000 samples split into activation-generation, validation, and testing subsets.Source: §2.2 · p. 3
• Behaviour vectors use contrastive activation addition from matching versus non-matching prompts, with hyperparameter search over injection layers and coefficients.The matching score is the share of outputs matching the target behaviour.Source: §2.2 · p. 4
• The evaluation discards settings producing over 5% invalid outputs or mode collapse, and assigns penalties to invalid or collapsed combined-steering outcomes.Source: §2.2 · p. 4
• Combined steering forms eight vectors by varying coefficient scaling, averaging versus summing, and addition versus subtraction; simultaneous steering injects separate vectors at layers 11–15.Simultaneous evaluation uses one global coefficient and also measures top-1 accuracy on 500k Pile tokens.Source: §2.2 · p. 4

3.1 Broad steering

• Broad coding steering achieves a 40% reduction in coding accuracy at 80% textual accuracy, while a 10% coding reduction corresponds to only a 3% textual reduction at layer 15.The steering vector for layer 0 instead produces the opposite of the intended effect.Source: §3.1 · p. 4, Figure 3 · p. 5
• Python steering shows similar performance to general coding steering, contradicting the hypothesis that broader skills would have smaller effect sizes, although selective pruning is more effective.The paper therefore finds broad coding steering competitive with narrower Python steering.Source: §3.1 · p. 4, Figure 4 · p. 5

3.2 Multi-steering

• Individual steering works in both directions for myopia, wealth seeking, agreeableness, and anti-immigration, but sycophancy has a negligible effect in this experiment.The sycophancy result differs from previous work, likely because this study focuses on political topology rather than other forms of sycophancy.Source: §3.2 · p. 6, Figure 5 · p. 6
• Combining eight behaviour vectors produces unexpected, often smaller effects, fails to maintain substantial wealth-seeking control, and can induce mode collapse.The combined-vector results show reduced steerability relative to individual steering, with some effects reversing the intended direction.Source: §3.2 · p. 6, Figure 6 · p. 6
• Simultaneous steering preserves substantial effects for myopia and wealth seeking, whereas anti-immigration and agreeableness effects are smaller or unstable and sycophancy remains ineffective.Increasing the global injection coefficient can cause mode collapse, although the alignment tax remains minor near coefficients -1 and +1.Source: §3.2 · p. 6, Figure 7 · p. 7
• Simultaneous steering injects each individual vector at a different layer using one global injection coefficient, while combined steering injects a single constructed vector at layer 15.Figure 7 reports behaviour matching scores and alignment-tax top-1 prediction accuracy; missing line segments indicate more than 5% faulty responses.Source: Figure 6 · p. 6, Figure 7 · p. 7

4.1 Broad steering

• Activation steering works for broad coding skills and is competitive with steering toward narrower Python ability.Source: §4.1 · p. 7
• The broad-steering result may reflect correlated Python and general-coding representations because Python data may constitute a large part of Llama 2’s coding data.This explanation cannot be verified or refuted because Llama 2’s training data is private.Source: §4.1 · p. 7
• Removing coding ability at layer 0 paradoxically improves coding performance, and the proposed copy-suppression explanation remains speculative.More research is needed to support or refute this theory.Source: §4.1 · p. 7

4.2 Multi-steering

• Combining multiple steering vectors into one is largely ineffective, whereas simultaneous injection at different model locations reliably steers myopia and wealth seeking.Agreeableness and anti-immigration show only minor, less reliable effects.Source: §4.2 · pp. 7–8
• Simultaneous steering appears more promising because it may disturb each behaviour’s activation pattern less, while interaction effects do not substantially reduce steering.It produces only a marginal alignment tax.Source: §4.2 · p. 8

4.3 General discussion

• Activation steering’s flexible generation and injection choices broaden applicability but make optimal setup selection difficult because best practices are not established.Source: §4.3 · p. 8
• Matching scores may proxy open-ended performance, but it remains unclear whether a 20% reduction in top-1 coding accuracy reduces real-world risks.Source: §4.3 · p. 8

4.4 Future work

• Future work should test different injection coefficients, within-layer injection sites, other programming languages, and other model architectures.The authors suggest sparse models may be easier to steer because their activation patterns could be cleaner.Source: §4.4 · p. 8

• Teun van der Weij was the main author, while Nandi Schoots and Massimo Poesio supervised the project.Nandi Schoots supervised for longer and more intensively.Source: Contributions · p. 9

• The section states that it provides additional details to the methodology sections.Source: A Detailed methodology · p. 10

• All samples were truncated to 4096 tokens, and the model used bfloat16 precision.Source: §A.1 · p. 10
• Datasets were randomly shuffled with seed 13 before use.Source: §A.1 · p. 10
• Generated text was produced without sampling.Source: §A.1 · p. 10

A.2 Broad steering experiments

• The procedure adaptively traversed injection coefficients while monitoring coding accuracy.Source: §A.2 · p. 10
• Injection coefficients were selected using relative steered top-1 accuracy for general or Python code.Runs stopped below 0.05 relative accuracy; below 0.15, the coefficient step size increased from 1 to 5.Source: §A.2 · p. 10

• The grid search evaluated injection coefficients spanning 0.5–300 across eight tested layers: 0, 5, 10, 15, 20, 25, 29, and 31.Layers 10 and 15 were the two most effective layers.Source: §A.3 · p. 10
• Table 1 reports behaviour-specific injection coefficients for adding and subtracting steering vectors at layers 10 and 15.These were identified as the two most effective layers.Source: §A.3 · p. 10, Table 1 · p. 10

B.1 Single steering

• Figure 8 presents individual-steering results for layer 10, with injection coefficients specified in Table 1.Source: Figure 8 · p. 11

B.2 Combined steering

• Figure 9 compares combined steering with individual steering using mean or sum, weighted or unweighted, and added or subtracted vectors.Individual-steering results are shown as grey horizontal reference lines.Source: Figure 9 · p. 11

C.1 General coding activation distributions

• Figure 10 shows the distribution of layer-15 activations across multiple datasets for general coding.Source: Figure 10 · p. 12

C.2 Multi steering activation distributions

• Figure 11 shows activation distributions for layers 10 and 15 across numerous concepts.Source: Figure 11 · p. 12