Cross-seed explainability using Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoders
arXiv:2607.08499v1 Announce Type: new Abstract: We present a Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoder (SAE) for extracting cross-seed universal features from independently trained BERT models. Cross-seed feature universality is a fundamental challenge in mechanistic interpretability: because dictionary learning is non-convex, independently trained networks learn misaligned feature spaces, so apparently identical features may differ by random initialization. We address this by computing an orthogonal Procrustes rotation between seeds' activation spaces before joint SAE t
The increasing complexity and scale of AI models necessitate advanced interpretability techniques to understand their internal workings and ensure robust performance.
This research addresses a fundamental challenge in mechanistic interpretability, enabling more reliable and universal feature extraction from independent AI models, which is crucial for model debugging and trustworthiness.
The ability to identify and align 'universal features' across independently trained AI models will improve the transferability of AI insights and potentially accelerate safer AI development.
- · AI researchers
- · AI safety organizations
- · Developers of large language models
- · AI interpretability tool providers
- · Those relying on 'black box' AI solutions
- · Less rigorous interpretability methodologies
Improved understanding and debugging of complex AI systems, leading to more reliable and predictable AI behavior.
Reduced barriers to combining insights from different AI models and research efforts, fostering collaborative AI development.
Accelerated progress toward truly generalizable and interpretable AI, potentially impacting applications in critical sectors where explainability is paramount.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL