MetaSAEs: Joint Training with a Decomposability Penalty Produces More Atomic Sparse Autoencoder Latents

• URL: http://arxiv.org/abs/2604.03436v1
• Date: Fri, 03 Apr 2026 20:20:28 GMT
• Title: MetaSAEs: Joint Training with a Decomposability Penalty Produces More Atomic Sparse Autoencoder Latents
• Abstract summary: Sparse autoencoders (SAEs) are increasingly used for safety-relevant applications including alignment detection and model steering.<n>In practice, SAE latents blend representational subspaces together.<n>A single feature can activate across semantically distinct contexts that share no true common representation.<n>We introduce a joint training objective that directly penalizes this subspace blending.
• Score: 0.0
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Sparse autoencoders (SAEs) are increasingly used for safety-relevant applications including alignment detection and model steering. These use cases require SAE latents to be as atomic as possible. Each latent should represent a single coherent concept drawn from a single underlying representational subspace. In practice, SAE latents blend representational subspaces together. A single feature can activate across semantically distinct contexts that share no true common representation, muddying an already complex picture of model computation. We introduce a joint training objective that directly penalizes this subspace blending. A small meta SAE is trained alongside the primary SAE to sparsely reconstruct the primary SAE's decoder columns; the primary SAE is penalized whenever its decoder directions are easy to reconstruct from the meta dictionary. This occurs whenever latent directions lie in a subspace spanned by other primary directions. This creates gradient pressure toward more mutually independent decoder directions that resist sparse meta-compression. On GPT-2 large (layer 20), the selected configuration reduces mean $|\varphi|$ by 7.5% relative to an identical solo SAE trained on the same data. Automated interpretability (fuzzing) scores improve by 7.6%, providing external validation of the atomicity gain independent of the training and co-occurrence metrics. Reconstruction overhead is modest. Results on Gemma 2 9B are directional. On not-fully-converged SAEs, the same parameterization yields the best results, a $+8.6\%$ $Δ$Fuzz. Though directional, this is an encouraging sign that the method transfers to a larger model. Qualitative analysis confirms that features firing on polysemantic tokens are split into semantically distinct sub-features, each specializing in a distinct representational subspace.

Related papers

• SPARE: Structural Parameter-Free Affinity Regularization for Flow Matching [30.43836162456455]
  Regularizing internal representations has emerged as an effective accelerator, yet existing methods split into two families with complementary costs.<n>We propose Structural.<n>free Affinity Regularization (SPARE), a regularizer that matches the pairwise affinities of intermediate tokens to those of the clean latents.<n>On ImageNet $256 times 256$ with SiT backbones under matched 400K-iteration budgets, SPARE adds no encoder, head, or parameters and only 0.08 GB of training memory.
  arXiv  Detail & Related papers  (2026-08-03T09:50:51Z)
• Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability [11.543771846135021]
  Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models.<n>We show that this assumption mismatches with the multi-dimensional structure of model features.<n>We introduce Subspace-Aware Sparse Autoencoders (SASA), which replace single-vector decoders with learned decoder subspaces.
  arXiv  Detail & Related papers  (2026-06-04T16:08:25Z)
• Label-Free Cross-Task LoRA Merging with Null-Space Compression [50.63908869296697]
  We introduce Null-Space Compression (NSC) Merging, a label-free, output-agnostic method that sets merge weights from adapter geometry.<n>NSC achieves state-of-the-art performance across twenty heterogeneous vision tasks with balanced gains where prior methods overfit subsets of tasks.
  arXiv  Detail & Related papers  (2026-03-27T11:34:41Z)
• Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? [10.871959954490217]
  Sparse Autoencoders (SAEs) have emerged as a promising tool for interpreting neural networks by decomposing their activations into sparse sets of human-interpretable features.<n>Recent work has introduced multiple SAE variants and successfully scaled them to frontier models.<n>Despite much excitement, a growing number of negative results in downstream tasks casts doubt on whether SAEs recover meaningful features.
  arXiv  Detail & Related papers  (2026-02-15T11:53:55Z)
• Sparse Semantic Dimension as a Generalization Certificate for LLMs [53.681678236115836]
  We introduce the Sparse Semantic Dimension (SSD), a complexity measure derived from the active feature vocabulary of a Sparse Autoencoder (SAE) trained on the model's layers.<n>We validate this framework on GPT-2 Small and Gemma-2B, demonstrating that our bound provides non-vacuous certificates at realistic sample sizes.
  arXiv  Detail & Related papers  (2026-02-11T21:45:18Z)
• What matters for Representation Alignment: Global Information or Spatial Structure? [64.67092609921816]
  Representation alignment (REPA) guides generative training by distilling representations from a strong, pretrained vision encoder to intermediate diffusion features.<n>We investigate a fundamental question: what aspect of the target representation matters for generation, its textitglobal revisionsemantic information.<n>We replace the standard projection layer in REPA with a simple convolution layer and introduce a spatial normalization layer for the external representation.
  arXiv  Detail & Related papers  (2025-12-11T16:39:53Z)
• AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features [19.58274892471746]
  Sparse autoencoders (SAEs) have emerged as powerful techniques for interpretability of large language models.<n>We introduce such a framework by unrolling the proximal gradient method for sparse coding.<n>We show that a single-step update naturally recovers common SAE variants, including ReLU, JumpReLU, and TopK.
  arXiv  Detail & Related papers  (2025-10-01T01:29:31Z)
• Sparse Autoencoders Do Not Find Canonical Units of Analysis [6.0188420022822955]
  A common goal of mechanistic interpretability is to decompose the activations of neural networks into features.<n>Sparse autoencoders (SAEs) are a popular method for finding these features.<n>We use two techniques: SAE stitching to show they are incomplete, and meta-SAEs to show they are not atomic.
  arXiv  Detail & Related papers  (2025-02-07T12:33:08Z)
• AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders [73.37603699731329]
  We introduce AxBench, a large-scale benchmark for steering and concept detection.<n>For steering, we find that prompting outperforms all existing methods, followed by finetuning.<n>For concept detection, representation-based methods such as difference-in-means, perform the best.
  arXiv  Detail & Related papers  (2025-01-28T18:51:24Z)
• SA-DVAE: Improving Zero-Shot Skeleton-Based Action Recognition by Disentangled Variational Autoencoders [7.618223798662929]
  We propose SA-DVAE -- Semantic Alignment via Disentangled Variational Autoencoders. We implement this idea via a pair of modality-specific variational autoencoders coupled with a total correction penalty. Experiments show that SA-DAVE produces improved performance over existing methods.
  arXiv  Detail & Related papers  (2024-07-18T12:35:46Z)
• Symmetric Equilibrium Learning of VAEs [56.56929742714685]
  We view variational autoencoders (VAEs) as decoder-encoder pairs, which map distributions in the data space to distributions in the latent space and vice versa. We propose a Nash equilibrium learning approach, which is symmetric with respect to the encoder and decoder and allows learning VAEs in situations where both the data and the latent distributions are accessible only by sampling.
  arXiv  Detail & Related papers  (2023-07-19T10:27:34Z)
• Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition [62.83832841523525]
  We propose a fast and accurate parallel transformer, termed Paraformer. It accurately predicts the number of output tokens and extract hidden variables. It can attain comparable performance to the state-of-the-art AR transformer, with more than 10x speedup.
  arXiv  Detail & Related papers  (2022-06-16T17:24:14Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.