The State of Mechanistic Interpretability
for AI Safety

Methods · Representation Analysis: Comparison, Localisation, and Control

Representation Engineering and Steering (Steering Vectors)

Survey sections: §9.3 Representation Engineering and Steering · §9.3 Steering as a Causal Test for Feature Interpretability · §13.3.1 Steering Vectors: Practical Behavioural Control

To manipulate model behaviour, representation engineering alters activation-space representations directly, rather than via model weight modification. Activation steering generally refers to methods that target intermediate activations (usually the residual stream) by adding certain vectors that shift model outputs during inference.

Assumptions and requirements

Because they require no weight updates or retraining, these interventions are lightweight approaches for controlling behaviour.
Another popular method is difference-in-means, which obtains the steering vector by taking the difference between activations associated with two behaviours (Rimsky et al., 2024)

Evaluation considerations

If intervening along a direction consistently changes a target behaviour, this provides evidence that the corresponding representation causally influences that behaviour.
Supporting evidence can be obtained by steering along the identified feature direction and evaluating the resulting change in behaviour. These evaluations have been used to compare interpretable direction-identification approaches, including SAE features, difference-in-means directions, and DAS (Huang et al., 2024; Wu et al., 2025b).
Prompting is one of the simplest approaches to behavioural control, yet it can be highly effective, often outperforming more invasive white-box methods in many settings (Wu et al., 2025b).

Intended applications

While these approaches appear simple, they reliably change model behaviours such as refusal, helpfulness, and stylistic tendencies (Rimsky et al., 2024; Zhao et al., 2025; Arditi et al., 2024).
Researchers identified and applied a steering vector that suppresses this awareness, allowing evaluation of its actual behaviour.
Steering vectors, introduced in Section 9.3 as activation-level interventions, have been shown to be effective for targeted behavioural modification: adding a specific direction α⋅vsteer\alpha \cdot v_{\text{steer}} to the residual stream can reduce sycophancy, increase honesty, and prevent certain jailbreaks (Rimsky et al., 2024; Eshuijs et al., 2026). Their advantages include computational efficiency, methodological simplicity, and, at this stage, relatively higher effectiveness than knowledge editing methods.

Limitations

While steering methods effectively control behaviours and support causal testing, they have limitations. Notably, they can be brittle, as a direction identified for one distribution, model, or task may not generalise reliably.
Another limitation is that successful behavioural steering does not, by itself, establish mechanistic transparency, because an intervention may simultaneously impact multiple features, interact nonlinearly with existing computations, or exploit unintended pathways that produce the desired behaviour without corresponding to a clean mechanism (Makelov et al., 2024). Thus, steering should generally be interpreted as evidence for causal influence, but it should not, by itself, be taken as a mechanistic explanation of the behaviour.
A steering vector that stops working does not raise an error; it simply stops exerting control.
Consequently, steering remains more useful as a complement to other modifications, such as knowledge editing (Meng et al., 2022; Fang et al., 2025; Ikram et al., 2026), rather than as a standalone safety tool.

Related methods

Compare side by side →

Used in safety applications

Run it

Key papers

Cited in §9.3.

Also discussed in: §3.2 Concept-Based Approaches · §7.5.1 The Need for Standardised Evaluation · §8.3 Jacobian Lens: Decoding Computational Dispositions · §12.1 Pragmatic Interpretability: Proxy-Task-Oriented Safety Interventions · §13.5 Choosing Appropriate Methods for Different Safety Applications · §13.4 Automated Interpretability Agents · §13.7.2 Distribution Shift Vulnerability · §13.7.2 Silent Failure Modes · §13.7.3 Dual-Use and Adversarial Exploitation · §14.4 Mechanistic Understanding and Alignment

Other methods in Representation Analysis: Comparison, Localisation, and Control

Model Diffing · Representation Alignment and Convergent Learning (SVCCA, PWCCA, CKA) · Causal Tracing (Knowledge Localization) · Optimisation-Based Steering