Research map / Architectures and efficiency / RNNs and normalization
A Curvature Phase Transition Governs Coherence Penalty Efficiency Against Feature Absorption in SAEs
Hak Hyun Kim, Yash Raj, Peter Chin, Soroush Vosoughi
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies the geometry governing whether a coherence penalty can escape feature absorption in sparse autoencoders, reporting a curvature-driven phase transition in penalty efficiency.
Problem
Feature absorption in sparse autoencoders, where one dictionary atom captures multiple distinct concepts, reduces interpretability; it is unclear what geometric mechanisms determine whether a coherence penalty can escape absorption spurious local minima.
Approach
Frames escape from absorption configurations as a geometric question and identifies a curvature phase transition that determines when a coherence penalty is efficient, building on prior work showing absorption configurations are spurious local minima of the sparse dictionary learning loss.
Experimental setup
Builds on: sparse dictionary learning feature absorption analysis
Closest papers by meaning
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.89)
- Task-Induced Riemannian Metrics for Vision Transformer Feature Spaces (NeurIPS 2026 · similarity 0.85)
- Avoiding spurious sharpness minimization broadens applicability of SAM (ICML 2025 · similarity 0.85)
- Deconstructing the Goldilocks Zone of Neural Network Initialization (ICML 2024 · similarity 0.84)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.84)
- Sparse Autoencoders Do Not Find Canonical Units of Analysis (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
