MBZUAI's CRH Reveals Why AI Steering Vectors Randomly Fail on Specific Inputs

If you have ever tried to steer a large language model's behavior by injecting a vector into its activations, you have probably noticed something frustrating: it works brilliantly on some inputs and completely fails on others, even when the steering vector looks perfectly well-crafted. A new paper accepted at ICML 2026 finally offers a principled geometric explanation for why this happens, and it challenges one of the most foundational assumptions in the interpretability field.

The promise and the problem with steering

Controlling LLM behavior through directly intervening on internal activations is an appealing idea, and various methods for doing so have been proposed. Most steering methods add a "steering vector" to the model's activations at a given layer and token position during inference. This approach leverages the hypothesis that many human-interpretable concepts like truthfulness, refusal, and sentiment are represented as directions in activation space. Steering interventions are appealing because they use much less data than fine-tuning and do not require changes to model parameters.

Much existing work on steering implicitly assumes the Linear Representation Hypothesis (LRH): the notion that models tend to represent atomic features as directions in latent space. However, this is unlikely to be true in full generality, given that nonlinearities are what make neural networks universal approximators. The practical consequence is stark: steerability is highly variable across different inputs, and depending on the concept, spurious biases can substantially contribute to how effective steering is for each input, presenting a challenge for the widespread use of steering vectors.

Why LRH breaks down in practice

The linear representation hypothesis suggests that human-interpretable concepts are encoded as linear directions or subspaces within an LLM's representation space, implying that LLM behavior can be understood and controlled by steering residual stream activations along these directions. The critical additional assumption is causal separability: the idea that logically independent concepts can be made orthogonal to each other, so steering one concept does not accidentally activate another.

The new paper shows this orthogonality assumption is mathematically impossible to satisfy in practice. In a d-dimensional space, at most d directions can be mutually orthogonal. Once the number of concepts encoded by a model exceeds the dimensionality of the representation space, overlap between concept directions is unavoidable. This is not noise or a failure of vector quality. It is a hard geometric constraint.