Finding Belief Geometries with Sparse Autoencoders

📰 ArXiv cs.AI

Researchers use sparse autoencoders to find belief geometries in internal representations of large language models

advanced Published 6 Apr 2026

Action Steps

Identify the geometric structure of internal representations in large language models
Use sparse autoencoders to encode probabilistic belief states
Analyze the residual stream to find simplex-shaped geometries
Interpret the vertices of the geometries as latent generative states

Who Needs to Know This

ML researchers and AI engineers can benefit from this research to improve model interpretability and understand how language models represent probabilistic beliefs

Key Insight

💡 Sparse autoencoders can be used to identify geometric structures in internal representations of large language models

Key Takeaways

Researchers use sparse autoencoders to find belief geometries in internal representations of large language models

Full Article

Title: Finding Belief Geometries with Sparse Autoencoders

Abstract:
arXiv:2604.02685v1 Announce Type: cross Abstract: Understanding the geometric structure of internal representations is a central goal of mechanistic interpretability. Prior work has shown that transformers trained on sequences generated by hidden Markov models encode probabilistic belief states as simplex-shaped geometries in their residual stream, with vertices corresponding to latent generative states. Whether large language models trained on naturalistic text develop analogous geometric repre

Read full paper → ← Back to Reads