Sparse autoencoders solve cross-seed feature alignment in BERT models
Researchers tackle a core bottleneck in mechanistic interpretability: when neural networks train independently, they learn feature spaces that don't align across runs, making it impossible to identify which learned patterns are truly universal versus noise. This work combines Procrustes rotation with sparse autoencoders to extract consistent features across multiple BERT seeds, validated on standard NLP benchmarks. The advance matters because interpretability research has struggled to distinguish fundamental model behaviors from initialization artifacts. Success here could accelerate efforts to reverse-engineer what large models actually learn, moving mechanistic interpretability from single-model analysis toward reproducible, cross-model understanding.
Modelwire context
ExplainerThe paper doesn't just apply Procrustes rotation to sparse autoencoders; it conditions the alignment on Top-K sparsity patterns, meaning features are matched only when they're actually active. This prevents false alignment of noise dimensions that happen to correlate across seeds but carry no semantic content.
This directly extends the convergence theory from the Contravariance paper (arXiv cs.LG, early July), which showed that weak alignment across independently trained networks guarantees strong alignment of core computational axes. Where that work proved the theory, this paper operationalizes it: by anchoring Procrustes matching to sparse features rather than full representations, it isolates the signal that theory predicts should be universal. The same day's Procrustes matching breakthrough in high-dimensional point clouds provides the algorithmic foundation, though applied here to feature space rather than geometric correspondence.
If the same Procrustes-sparse autoencoder pipeline recovers identical features when applied to GPT-2 or other architectures beyond BERT, that confirms the method finds genuinely universal patterns rather than BERT-specific artifacts. If results degrade significantly when applied to models trained on different datasets or objectives, the approach is architecture-portable but not task-portable, which would narrow its mechanistic interpretability claims.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsBERT · Sparse Autoencoders · Procrustes rotation · SST-2
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Cross-seed explainability using Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoders”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.