Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

C$^{2}$R: Cross-sample Consistency Regularization Mitigates Feature Splitting and Absorption in Sparse Autoencoders

Source

arXiv

arxiv.org/abs/2606.30609v1

Read the source

Source

Plain-Language Brief

The paper introduces C$^2$R, a method to improve Sparse Autoencoders (SAEs) used for interpreting large language models. The main challenge with SAEs is scaling to large dictionaries, which leads to issues like feature splitting and absorption that negatively impact the reliability of latent features. C$^2$R addresses this by encouraging a consistent representation of semantic features across multiple samples in the batch, without sacrificing reconstruction fidelity.

What the Document Says

  • Sparse Autoencoders (SAEs) are used for interpreting large language models.
  • Scaling SAEs to large dictionaries exposes challenges like feature splitting and absorption.
  • Feature splitting fragments coherent concepts into non-atomic latents, while feature absorption creates arbitrary exceptions in general features.
  • These issues stem from inconsistent latent assignment across samples.
  • C$^2$R (Cross-sample Consistency Regularization) encourages consistent representation of semantic features across the batch.
  • C$^2$R penalizes co-activation of directionally similar latents to address inconsistent latent assignment.
  • Comprehensive evaluation shows that C$^2$R mitigates both splitting and absorption while preserving reconstruction fidelity.
  • The source code for C$^2$R is available at https://github.com/hr-jin/Cross-sample-Consistency-Regularization.

Key Concepts Mentioned

  • Sparse Autoencoders
  • Feature splitting
  • Feature absorption
  • Cross-sample Consistency Regularization
  • Latent features
  • Large language models

Claims Made by the Document

  • SAEs are used for interpreting large language models.
  • Scaling SAEs to large dictionaries exposes challenges like feature splitting and absorption.
  • C$^2$R addresses the inconsistent latent assignment issue in SAEs.
  • C$^2$R mitigates both feature splitting and absorption while preserving reconstruction fidelity.

Evidence Used

  • The text describes the use of SAEs for interpreting large language models, the issues of feature splitting and absorption, and how C$^2$R addresses these issues.

Methods / Approach

  • Cross-sample Consistency Regularization (C$^2$R)

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Sparse Autoencoders (SAEs) are widely used to interpret large language models by decomposing activations into sparse, human-understandable features, but scaling to large dictionaries exposes fundamental challenges. Systematic studies reveal pervasive feature splitting that fragments coherent concepts into non-atomic latents and widespread feature absorption that creates arbitrary exceptions in general features, severely compromising latent reliability. These issues stem from inconsistent latent assignment across samples: without cross-sample constraints, per-sample optimization often allows a single underlying concept to be inconsistently distributed across multiple redundant or interfering latents. To address this, we introduce C$^2$R (\underline{\textbf{C}}ross-sample \underline{\textbf{C}}onsistency \underline{\textbf{R}}egularization). C$^2$R explicitly encourages that each semantic feature is consistently represented by a unified latent across the batch by penalizing the co-activation of directionally similar latents. Comprehensive evaluation demonstrates that C$^2$R effectively mitigates both splitting and absorption while, crucially, preserving reconstruction fidelity, providing a principled solution that enhances latent interpretability without degrading model performance. Source code is available at https://github.com/hr-jin/Cross-sample-Consistency-Regularization.