analiza aiqbench
This paper is clearly about AI interpretability: sparse autoencoder features in language models, so it passes the topic gate. Its primary contribution is a Wasserstein-based distributional metric for matching SAE features across layers and compressing circuits into supernodes, which is a methods contribution rather than a product or policy story. Readers working on mechanistic interpretability should weigh the reported gains over decoder-vector baselines against the usual open questions about how well such metrics hold up on real models and whether the "interpretable supernodes" are validated beyond automated compression.
Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression
arXiv:2605.28567v2 Announce Type: replace Abstract: Sparse autoencoders (SAEs) have become a central tool for interpreting language models. However, two key SAE analyses that remain difficult to scale are (1) matching semantically similar features across multi-layers and (2) compressing large feature circuits into interpretable supernodes. Although these have been treated as separate problems, we show that both are instances of a more fundamental challenge, which we frame as the estimation of semantic distances between SAE features that lie on different activation manifolds. We introduce a distributional framework for this problem, in which each feature is represented not by a single decoder vector like in the literature, but by an activation-weighted distribution over the hidden states that express it. By projecting these distributions into a shared reference space and comparing them with Wasserstein distance, our method provides a unified semantic metric for cross-layer feature comparison. We prove that our representation is invariant to activation rescaling, stable under perturbations, and recovers true matches under finite-sample margin conditions. Empirically, our method outperforms decoder-vector and LLM-based baselines and captures subtle functional distinctions between related features. Notably, our method compresses large feature circuits into interpretable supernodes automatically.
Abstract:Sparse autoencoders (SAEs) have become a central tool for interpreting language models. However, two key SAE analyses that remain difficult to scale are (1) matching semantically similar features across multi-layers and (2) compressing large feature circuits into interpretable supernodes. Although these have been treated as separate problems, we show that both are instances of a more fundamental challenge, which we frame as the estimation of semantic distances between SAE features that lie on different activation manifolds. We introduce a distributional framework for this problem, in which each feature is represented not by a single decoder vector like in the literature, but by an activation-weighted distribution over the hidden states that express it. By projecting these distributions into a shared reference space and comparing them with Wasserstein distance, our method provides a unified semantic metric for cross-layer feature comparison. We prove that our representation is invariant to activation rescaling, stable under perturbations, and recovers true matches under finite-sample margin conditions. Empirically, our method outperforms decoder-vector and LLM-based baselines and captures subtle functional distinctions between related features. Notably, our method compresses large feature circuits into interpretable supernodes automatically.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2605.28567 [cs.LG]
(or arXiv:2605.28567v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2605.28567 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 281 cuvinte
Subiecte conexe
Decoder-Preserving Sparse Autoencoders: Which Readouts Survive Sparse Compression?
arXiv · 21 iul. 2026
Hierarchical Token Merging with Recursive Autoencoders for Unsupervised Grammar Induction
arXiv · 21 oct. 2026
AutoEncoder-Compressed Parallel Split Learning for Pre-trained Model Fine-Tuning
arXiv · 21 iul. 2026
Manifolds for Five-Year Olds
LessWrong · 12 oct. 2026