Researchers have introduced Geometric Evolution Maps (GEMs), a new approach to extracting stable concept probes from transformer residual streams. A concept probe is considered reliable only if it's taken from a layer where its direction has settled, rather than during assembly or at the peak of a separation curve. GEMs record a concept's directional trajectory across one segment and identify the handoff layer beyond it, where the direction can be evaluated outside the window it was estimated on. This approach is more precise than probing during assembly and provides a unit for downstream use. The extracted direction is causal and ablating it suppresses separation more than ablating a random direction. --- Why it matters: This matters to AI researchers because GEMs provide a more accurate way to extract stable concept probes, which can improve the performance of downstream tasks such as natural language processing and computer vision. Source: https://arxiv.org/abs/2605.25848