A new preprint on arXiv tackles a core question in mechanistic interpretability: when do neural networks learn representations that correspond to the true latent variables behind the data? The authors frame this as an identifiability problem in supervised learning, asking under what conditions the training process can recover those underlying factors.
The abstract introduces two distinct notions: complete identifiability and sparse identifiability. The distinction appears to be central to the paper's analysis, though the truncated abstract does not spell out the exact conditions or results for each. The work is positioned as a contribution to understanding when interpretable representations emerge from supervised training.
As the abstract is brief, the specific theorems and experimental evidence are not detailed in the summary. The preprint is available on arXiv under identifier 2610.20851.