Does the Model Use the Feature? Separating Steering from Mechanism in LLMs

arXiv:2610.07270 · cs.AI · Submitted 2026-10-05 · Read on arXiv

cs.AI

Submitted: 2026-10-05

Updated: 2026-10-05

Terminology

• Steering — 1× in this paper · explained in 2 episodes

Sources

• Faithful Bi-Directional Model Steering via Distribution Matching and Distributed Interchange Interventions
• Persona Vectors: Monitoring and Controlling Character Traits in Language Models
• SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior
• Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals
• Causal Analysis of Syntactic Agreement Mechanisms in Neural Language Models
• Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
• Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability
• Does Localization Inform Editing? Surprising Differences in Causality-Based Localization vs. Knowledge Editing in Language Models
• How to use and interpret activation patching
• RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representations
• There Is More to Refusal in Large Language Models than a Single Direction
• Probing for the Usage of Grammatical Number
• Bucketing the Good Apples: A Method for Diagnosing and Improving Causal Abstraction
• Position: Mechanistic Interpretability Must Disclose Identification Assumptions for Causal Claims
• The Hydra Effect: Emergent Self-repair in Language Model Computations
• Transformer Circuit Faithfulness Metrics are not Robust
• Steered LLM Activations are Non-Surjective
• Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders
• Counterfactual Interventions Reveal the Causal Effect of Relative Clause Representations on Agreement Prediction
• The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence

Related papers

• MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
• Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
• Agents' Last Exam
• The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
• MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
• Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems