A paper out of arXiv proposes a label-free way to see inside a language model's decision-making.
The technique, called Mining via Activation Geometry (MAG), works by prepending the same natural-language question to every input — something like "Is this prompt malicious?" — and measuring how that question shifts the model's internal activations. The difference between the model's state with and without the question becomes a signal for a specific reasoning feature. The researchers tested eight variations of the method and found that the extracted features can be compressed into a single vector direction, which can then be injected back into the model to steer its outputs. On the practical side, using these "reasoning feature directions" to select training data for prompt-injection classifiers hit 94.7% Top-1 and 100% Top-2 accuracy — far better than using standard activation similarity.
Most interpretability research starts with human-labeled examples of a concept, which means it inherits whatever assumptions the labelers brought in. MAG sidesteps that entirely, letting the model's own geometry define the feature. That matters for safety work especially: prompt-injection detection is a real and unsolved problem, and a method that can bootstrap a classifier probe without curated labels is genuinely useful.
The caveat worth noting: the paper acknowledges that some features are more linearly represented than others, so the single-vector approximation won't always hold. Vector steering has also been around long enough that the field knows its limits — it can nudge a model, but it is not a precision instrument.