Decomposing MLP Activations into Interpretable Features via Semi-Nonnegative Matrix Factorization

• URL: http://arxiv.org/abs/2506.10920v1
• Date: Thu, 12 Jun 2025 17:33:29 GMT
• Title: Decomposing MLP Activations into Interpretable Features via Semi-Nonnegative Matrix Factorization
• Authors: Or Shafran, Atticus Geiger, Mor Geva,
• Abstract summary: Current methods rely on dictionary learning with sparse autoencoders (SAEs)<n>Here, we tackle these limitations by directly decomposing activations with semi-nonnegative matrix factorization (SNMF)<n>Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering.
• Score: 17.101290138120564
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: A central goal for mechanistic interpretability has been to identify the right units of analysis in large language models (LLMs) that causally explain their outputs. While early work focused on individual neurons, evidence that neurons often encode multiple concepts has motivated a shift toward analyzing directions in activation space. A key question is how to find directions that capture interpretable features in an unsupervised manner. Current methods rely on dictionary learning with sparse autoencoders (SAEs), commonly trained over residual stream activations to learn directions from scratch. However, SAEs often struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. Here, we tackle these limitations by directly decomposing MLP activations with semi-nonnegative matrix factorization (SNMF), such that the learned features are (a) sparse linear combinations of co-activated neurons, and (b) mapped to their activating inputs, making them directly interpretable. Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline (difference-in-means) on causal steering, while aligning with human-interpretable concepts. Further analysis reveals that specific neuron combinations are reused across semantically-related features, exposing a hierarchical structure in the MLP's activation space. Together, these results position SNMF as a simple and effective tool for identifying interpretable features and dissecting concept representations in LLMs.

Related papers

• Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling [56.26834106704781]
  Factual incorrectness in generated content is one of the primary concerns in ubiquitous deployment of large language models (LLMs)<n>We provide evidence supporting the presence of LLMs' internal compass that dictate the correctness of factual recall at the time of generation.<n>Scaling experiments across model sizes and training dynamics highlight that self-awareness emerges rapidly during training and peaks in intermediate layers.
  arXiv  Detail & Related papers  (2025-05-27T16:24:02Z)
• Hybrid Latent Reasoning via Reinforcement Learning [51.06635386903026]
  We explore latent reasoning by leveraging the capabilities of large language models (LLMs) via reinforcement learning (RL)<n>We introduce hybrid reasoning policy optimization (HRPO), an RL-based hybrid latent reasoning approach that integrates prior hidden states into sampled tokens with a learnable gating mechanism.<n>HRPO-trained LLMs remain interpretable and exhibit intriguing behaviors like cross-lingual patterns and shorter completion lengths.
  arXiv  Detail & Related papers  (2025-05-24T01:26:16Z)
• Sparse Autoencoders Learn Monosemantic Features in Vision-Language Models [50.587868616659826]
  We introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in vision representations.<n>Our experimental results reveal that SAEs trained on Vision-Language Models significantly enhance the monosemanticity of individual neurons.
  arXiv  Detail & Related papers  (2025-04-03T17:58:35Z)
• Computation Mechanism Behind LLM Position Generalization [59.013857707250814]
  Large language models (LLMs) exhibit flexibility in handling textual positions.<n>They can understand texts with position perturbations and generalize to longer texts.<n>This work connects the linguistic phenomenon with LLMs' computational mechanisms.
  arXiv  Detail & Related papers  (2025-03-17T15:47:37Z)
• I Predict Therefore I Am: Is Next Token Prediction Enough to Learn Human-Interpretable Concepts from Data? [76.15163242945813]
  Large language models (LLMs) have led many to conclude that they exhibit a form of intelligence.<n>We introduce a novel generative model that generates tokens on the basis of human-interpretable concepts represented as latent discrete variables.
  arXiv  Detail & Related papers  (2025-03-12T01:21:17Z)
• SAFR: Neuron Redistribution for Interpretability [7.756342860929851]
  Superposition refers to encoding representations of multiple features within a single neuron.<n>Despite promising performance, the model's interpretability has been diminished.<n>This paper presents a novel approach to enhance model interpretability by regularizing feature superposition.
  arXiv  Detail & Related papers  (2025-01-23T06:20:33Z)
• Automatically Interpreting Millions of Features in Large Language Models [1.8035046415192353]
  sparse autoencoders (SAEs) can be used to transform activations into a higher-dimensional latent space.<n>We build an open-source pipeline to generate and evaluate natural language explanations for SAE features.<n>Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons.
  arXiv  Detail & Related papers  (2024-10-17T17:56:01Z)
• The Local Interaction Basis: Identifying Computationally-Relevant and Sparsely Interacting Features in Neural Networks [0.0]
  Local Interaction Basis aims to identify computational features by removing irrelevant activations and interactions. We evaluate the effectiveness of LIB on modular addition and CIFAR-10 models. We conclude that LIB is a promising theory-driven approach for analyzing neural networks, but in its current form is not applicable to large language models.
  arXiv  Detail & Related papers  (2024-05-17T17:27:19Z)
• Understanding Large Language Model Behaviors through Interactive Counterfactual Generation and Analysis [22.755345889167934]
  We present an interactive visualization system that enables exploration of large language models (LLMs) through counterfactual analysis.<n>Our system features a novel algorithm that generates fluent and semantically meaningful counterfactuals.<n>A user study with LLM practitioners and interviews with experts demonstrate the system's usability and effectiveness.
  arXiv  Detail & Related papers  (2024-04-23T19:57:03Z)
• Self-Supervised Neuron Segmentation with Multi-Agent Reinforcement Learning [53.00683059396803]
  Mask image model (MIM) has been widely used due to its simplicity and effectiveness in recovering original information from masked images. We propose a decision-based MIM that utilizes reinforcement learning (RL) to automatically search for optimal image masking ratio and masking strategy. Our approach has a significant advantage over alternative self-supervised methods on the task of neuron segmentation.
  arXiv  Detail & Related papers  (2023-10-06T10:40:46Z)
• Sparse Autoencoders Find Highly Interpretable Features in Language Models [0.0]
  Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. We use sparse autoencoders to reconstruct the internal activations of a language model. Our method may serve as a foundation for future mechanistic interpretability work.
  arXiv  Detail & Related papers  (2023-09-15T17:56:55Z)

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.