A faster way to locate the circuits inside large language models that govern refusal behavior is now on the table.
Researchers have adapted an algorithm called the Recursive Feature Machine (RFM) to identify what are known as "refusal subspaces" - the multi-dimensional regions in a model's activations that determine whether it refuses a harmful query. Earlier work assumed refusal behavior lived along a single linear direction in a model's internal state, which made it easy to study. More recent findings complicated that picture: refusal turns out to occupy higher-dimensional structures. Existing tools for mapping those structures are slow enough to become impractical on reasoning models, which generate lengthy internal traces before responding. The new approach, using probe-informed initialization to bootstrap RFM, completes that mapping in seconds on both reasoning (Qwen 3) and non-reasoning (Qwen 2.5) models.
The speed gain matters beyond academic convenience. Safety and interpretability researchers increasingly rely on activation steering - nudging a model's internal state to change its behavior - and on monitoring activations to detect when a model is about to do something it shouldn't. Both techniques depend on knowing where the relevant signal lives. If RFM can find those subspaces cheaply and at scale, it becomes a practical tool rather than a research curiosity. The paper also reports that RFM outperformed competing methods on an ablation task, suggesting the speed gain does not come at a quality cost.
The authors are careful to frame this as preliminary - more work is needed to understand how subspaces found by different methods relate to each other. But the direction is clear: the harder it becomes to map what models are actually doing internally, the more pressure there will be to find shortcuts that do not sacrifice rigor.