Sparse Concept Anchoring for Interpretable and Controllable Neural Representations

2025/12/13 by Fraser, Sandy, Patryk Wielopolski, Wielopolski, Patryk
Computer Science · #Domain Adaptation and Few-Shot Learning #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG)

paper · pdf · doi:10.48550/arxiv.2512.12469

openalex publication_date 2025/12/13 · openalex created_date 2025/12/17 · openalex updated_date 2026/07/28

Abstract

We introduce Sparse Concept Anchoring, a method that biases latent space to position a targeted subset of concepts while allowing others to self-organize, using only minimal supervision (labels for <0.1% of examples per anchored concept). Training combines activation normalization, a separation regularizer, and anchor or subspace regularizers that attract rare labeled examples to predefined directions or axis-aligned subspaces. The anchored geometry enables two practical interventions: reversible behavioral steering that projects out a concept's latent component at inference, and permanent removal via targeted weight ablation of anchored dimensions. Experiments on structured autoencoders show selective attenuation of targeted concepts with negligible impact on orthogonal features, and complete elimination with reconstruction error approaching theoretical bounds. Sparse Concept Anchoring therefore provides a practical pathway to interpretable, steerable behavior in learned representations.

Citations

• Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors
• Ready2Unlearn: A Learning-Time Approach for Preparing Models with Future Unlearning Readiness
• Zero-shot Concept Bottleneck Models
• nGPT: Normalized Transformer with Representation Learning on the Hypersphere
• Optimal ablation for interpretability
• Analyzing the Generalization and Reliability of Steering Vectors
• External Invariants: A Cryptographic Trust Architecture for Institutional AI Inference
• Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization
• Not All Language Model Features Are One-Dimensionally Linear
• Mechanistic Interpretability for AI Safety -- A Review
• Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
• Sparse Concept Bottleneck Models: Gumbel Tricks in Contrastive Learning
• The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning
• Machine Unlearning of Pre-trained Large Language Models
• Steering Llama 2 via Contrastive Activation Addition
• Auxiliary Losses for Learning Generalizable Concept-based Models
• Representation Engineering: A Top-Down Approach to AI Transparency
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Label-Free Concept Bottleneck Models
• Knowledge Unlearning for Mitigating Privacy Risks in Language Models
• Concept Bottleneck Model with Additional Unsupervised Concepts
• Do Concept Bottleneck Models Learn as Intended?
• Concept Bottleneck Models
• Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere
• Machine Unlearning
• Ablation Studies in Artificial Neural Networks
• SphereFace: Deep Hypersphere Embedding for Face Recognition
• Distributed Representations of Words and Phrases and their Compositionality
• Representation Learning: A Review and New Perspectives
• Representation Learning: A Review and New Perspectives

Related

• A Geometric Unification of Concept Learning with Concept Cones
• Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry
• Interpretable and Steerable Concept Bottleneck Sparse Autoencoders
• Evaluating Sparse Autoencoders for Monosemantic Representation
• Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment