PhD student, EPFL - EPF Lausanne
2 papers at NeurIPS 2025
We show in theory and practice that by allowing non-linear transformations in causal abstraction, any neural network (even random ones) can be perfectly aligned to any algorithm, rendering this interpretability approach meaningless if unconstrained.
Using crosscoders (SAE variant) for chat-tuning concept identification, we diagnose spurious chat-only concepts arising from L1 loss artifacts and show BatchTopK robustly reveals genuine, interpretable ones.
