A new training technique aims to make AI interpretability tools more reliable by forcing them to stay consistent across data samples.
Sparse autoencoders are a popular method for peering inside large language models: they decompose the model's internal activations into simpler, supposedly human-readable features. The problem is that at scale, these tools develop two ugly habits. Feature splitting breaks a single coherent concept into several redundant fragments. Feature absorption does the opposite - it hides exceptions to a general rule inside a separate, overlapping feature, making both unreliable. The root cause, according to the researchers, is that each training sample gets optimized in isolation, so the same underlying concept can land in different places depending on which sample the model happens to be processing.
The proposed fix, called Cross-sample Consistency Regularization (C2R), adds a penalty during training that discourages multiple directionally similar features from firing together. The idea is to pressure the autoencoder into picking one stable slot for each concept and sticking with it across the whole batch. The researchers report that C2R reduces both splitting and absorption without meaningfully hurting reconstruction accuracy - the metric that measures how well the autoencoder still captures what the original model was doing.
Interpretability research has gained urgency as labs push models into higher-stakes settings, but the tooling has quietly lagged behind the models themselves. If sparse autoencoders cannot reliably isolate concepts, the explanations they produce are closer to impressionism than engineering. A consistency fix that does not cost reconstruction quality is a meaningful step - though whether it holds up across model families larger than those tested here remains an open question.