Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression

2026-05-27 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

A new method, "Semantic Optimal Transport," addresses key challenges in interpreting language models using Sparse Autoencoders (SAEs). Specifically, it tackles the difficulty of matching semantically similar features across multi-layers and compressing large feature circuits into interpretable supernodes. The authors, Tue M. Cao, Nguyen Do, and My T. Thai, unify these problems as estimating semantic distances between SAE features on different activation manifolds. Their distributional framework represents each feature as an activation-weighted distribution over hidden states, rather than a single decoder vector. By projecting these distributions into a shared reference space and comparing them with Wasserstein distance, the method offers a unified semantic metric for cross-layer feature comparison. This approach is proven invariant to activation rescaling, stable under perturbations, and recovers true matches. Empirically, it outperforms decoder-vector and LLM-based baselines, capturing subtle functional distinctions and automatically compressing large feature circuits.

Key takeaway

For AI Scientists and Machine Learning Engineers focused on interpreting large language models, this Semantic Optimal Transport method offers a robust approach to analyze Sparse Autoencoder features. You can now reliably match semantically similar features across different layers and automatically compress complex feature circuits into interpretable supernodes. This improves your ability to understand model internals and debug emergent behaviors more effectively. Consider integrating this distributional framework for deeper mechanistic interpretability.

Key insights

Semantic Optimal Transport unifies SAE feature matching and circuit compression via activation-weighted distributions and Wasserstein distance.

Principles

• Features are activation-weighted distributions.
• Wasserstein distance measures semantic similarity.
• Representation is invariant to activation rescaling.

Method

The method represents SAE features as activation-weighted distributions over hidden states, projects them into a shared reference space, and compares them using Wasserstein distance for unified semantic metrics.

In practice

• Match semantically similar SAE features across layers.
• Compress large SAE feature circuits into supernodes.
• Evaluate feature quality beyond decoder vectors.

Topics

• Sparse Autoencoders
• Language Model Interpretability
• Optimal Transport
• Wasserstein Distance
• Feature Matching
• Circuit Compression

Code references

• marianlupascu/OT-RF

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Improving Sparse Autoencoder with Dynamic Attention — Dynamic sparse attention using sparsemax improves SAEs by automatically balancing sparsity and reconstruction quality.
• UNBOX: Unveiling Black-box visual models with Natural-language — UNBOX enables black-box visual model interpretability by using LLMs and diffusion models for semantic search.
• Sparsity as a Key: Unlocking New Insights from Latent Structures for Out-of-Distribution Detection — Sparse Autoencoders can disentangle Vision Transformer features, revealing structural invariants for robust OOD detection.
• $L^2$ over Wasserstein: Statistical Analysis for Optimal Transport — The $L^{2}$ over Wasserstein space unifies optimal transport under statistical uncertainty, extending its geometric and flow properties to random measures.
• Towards Intrinsic Interpretability of Large Language Models:A Survey of Design Principles and Architectures — Intrinsic interpretability builds transparency directly into LLM architectures to enhance trustworthiness and safety.
• MonoLoss: A Training Objective for Interpretable Monosemantic Representations — MonoLoss transforms a quadratic monosemanticity metric into a linear-time training objective, significantly improving model interpretability and performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.