{"total":95,"rows":[{"id":"2501.17148","title":"AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders","created":"2025-01-28","cited_by_count":218,"area":"mechinterp"},{"id":"2502.17420","title":"The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence","created":"2025-02-24","cited_by_count":115,"area":"alignment"},{"id":"2506.07022","title":"AlphaSteer: Learning Refusal Steering with Principled Null-Space Constraint","created":"2025-06-08","cited_by_count":49,"area":"safety"},{"id":"2604.02029","title":"The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook","created":"2026-04-02","cited_by_count":43,"area":null},{"id":"2501.08319","title":"Enhancing Automated Interpretability with Output-Centric Feature Descriptions","created":"2025-01-14","cited_by_count":41,"area":null},{"id":"2503.05021","title":"Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety","created":"2025-03-06","cited_by_count":39,"area":"lr"},{"id":"2602.02132","title":"There Is More to Refusal in Large Language Models than a Single Direction","created":"2026-02-02","cited_by_count":26,"area":"mechinterp"},{"id":"2601.22594","title":"Language Model Circuits Are Sparse in the Neuron Basis","created":"2026-01-30","cited_by_count":25,"area":"mechinterp"},{"id":"2504.08192","title":"SAEs Can Improve Unlearning: Dynamic Sparse Autoencoder Guardrails for Precision Unlearning in LLMs","created":"2025-04-11","cited_by_count":22,"area":"safety"},{"id":"2509.09708","title":"Beyond I'm Sorry, I Can't: Dissecting Large Language Model Refusal","created":"2025-09-07","cited_by_count":19,"area":"safety"},{"id":"2505.20254","title":"Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs","created":"2025-05-26","cited_by_count":14,"area":"mechinterp"},{"id":"2506.01197","title":"Incorporating Hierarchical Semantics in Sparse Autoencoder Architectures","created":"2025-06-01","cited_by_count":14,"area":"mechinterp"},{"id":"2503.14530","title":"SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoders","created":"2025-03-16","cited_by_count":13,"area":"mllm"},{"id":"2508.12535","title":"CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features","created":"2025-08-18","cited_by_count":12,"area":"mechinterp"},{"id":"2601.05679","title":"Do Sparse Autoencoders Identify Reasoning Features in Language Models?","created":"2026-01-09","cited_by_count":12,"area":"lr"},{"id":"2412.19512","title":"Safeguard Fine-Tuned LLMs Through Pre- and Post-Tuning Model Merging","created":"2024-12-27","cited_by_count":11,"area":null},{"id":"2502.18147","title":"Jacobian Sparse Autoencoders: Sparsify Computations, Not Just Activations","created":"2025-02-25","cited_by_count":10,"area":"mechinterp"},{"id":"2503.06269","title":"Using Mechanistic Interpretability to Craft Adversarial Attacks against Large Language Models","created":"2025-03-08","cited_by_count":10,"area":"mechinterp"},{"id":"2510.03659","title":"Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoders","created":"2025-10-04","cited_by_count":10,"area":"mechinterp"},{"id":"2603.01326","title":"Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning","created":"2026-03-01","cited_by_count":10,"area":"lr"},{"id":"2508.05775","title":"Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation of LLM","created":"2025-08-07","cited_by_count":9,"area":"alignment"},{"id":"2604.22167","title":"Estimating Tail Risks in Language Model Output Distributions","created":"2026-04-24","cited_by_count":9,"area":"safety"},{"id":"2507.23220","title":"Model Directions, Not Words: Mechanistic Topic Models Using Sparse Autoencoders","created":"2025-07-31","cited_by_count":8,"area":null},{"id":"2509.25624","title":"STAC: When Innocent Tools Form Dangerous Chains for LLM Agents","created":"2025-09-30","cited_by_count":8,"area":"agents"},{"id":"2602.04735","title":"From Data to Behavior: Predicting Unintended Model Behaviors Before Training","created":"2026-02-04","cited_by_count":8,"area":null},{"id":"2602.05176","title":"Among Us: Measuring and Mitigating Malicious Contributions in Model Collaboration Systems","created":"2026-02-05","cited_by_count":7,"area":"agents"},{"id":"2602.05859","title":"DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders","created":"2026-02-05","cited_by_count":7,"area":"diffusion"},{"id":"2505.24360","title":"Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning","created":"2025-05-30","cited_by_count":6,"area":"diffusion"},{"id":"2506.09967","title":"Resa: Transparent Reasoning Models via SAEs","created":"2025-06-11","cited_by_count":6,"area":"lr"},{"id":"2602.12418","title":"Sparse Autoencoders are Capable LLM Jailbreak Mitigators","created":"2026-02-12","cited_by_count":6,"area":"safety"},{"id":"2508.20766","title":"Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection","created":"2025-08-28","cited_by_count":5,"area":"alignment"},{"id":"2512.23260","title":"Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation","created":"2025-12-29","cited_by_count":5,"area":"alignment"},{"id":"2606.03002","title":"Perplexity Can Miss SAE Feature Damage Under Quantization","created":"2026-06-02","cited_by_count":5,"area":"mechinterp"},{"id":"2510.01246","title":"A Comparative Analysis of Sparse Autoencoder and Activation Difference in Language Model Steering","created":"2025-09-24","cited_by_count":4,"area":"mechinterp"},{"id":"2512.02004","title":"AlignSAE: Concept-Aligned Sparse Autoencoders","created":"2025-12-01","cited_by_count":4,"area":"mechinterp"},{"id":"2512.07667","title":"Depth-Wise Activation Steering for Honest Language Models","created":"2025-12-08","cited_by_count":4,"area":"mechinterp"},{"id":"2602.18505","title":"Suppression or Deletion: A Restoration-Based Representation-Level Analysis of Machine Unlearning","created":"2026-02-18","cited_by_count":4,"area":"safety"},{"id":"2605.23036","title":"Multilingual Steering by Design: Multilingual Sparse Autoencoders and Principled Layer Selection","created":"2026-05-21","cited_by_count":4,"area":"mechinterp"},{"id":"2502.16385","title":"Toward a Flexible Framework for Linear Representation Hypothesis Using Maximum Likelihood Estimation","created":"2025-02-22","cited_by_count":3,"area":null},{"id":"2505.03189","title":"Patterns and Mechanisms of Contrastive Activation Engineering","created":"2025-05-06","cited_by_count":3,"area":"mechinterp"},{"id":"2505.16077","title":"Ensembling Sparse Autoencoders","created":"2025-05-21","cited_by_count":3,"area":"mechinterp"},{"id":"2510.22099","title":"Generalization or Memorization: Dynamic Decoding for Mode Steering","created":"2025-10-25","cited_by_count":3,"area":"mechinterp"},{"id":"2602.04903","title":"Mind the Performance Gap: Capability-Behavior Trade-offs in Feature Steering","created":"2026-02-03","cited_by_count":3,"area":null},{"id":"2602.22719","title":"Interpreting and Steering State-Space Models via Activation Subspace Bottlenecks","created":"2026-02-26","cited_by_count":3,"area":"ssm"},{"id":"2605.06342","title":"Don't Lose Focus: Activation Steering via Key-Orthogonal Projections","created":"2026-05-07","cited_by_count":3,"area":"mechinterp"},{"id":"2507.12950","title":"Insights into a radiology-specialised multimodal large language model with sparse autoencoders","created":"2025-07-17","cited_by_count":2,"area":"mllm"},{"id":"2508.16325","title":"ConceptGuard: Neuro-Symbolic Safety Guardrails via Sparse Interpretable Jailbreak Concepts","created":"2025-08-22","cited_by_count":2,"area":"safety"},{"id":"2509.19839","title":"LatentGuard: Controllable Latent Steering for Robust Refusal of Attacks and Reliable Response Generation","created":"2025-09-24","cited_by_count":2,"area":"alignment"},{"id":"2601.22010","title":"Exploring Diverse Generation Paths via Inference-time Stiefel Activation Steering","created":"2026-01-29","cited_by_count":2,"area":"mechinterp"},{"id":"2603.05773","title":"Knowing without Acting: The Disentangled Geometry of Safety Mechanisms in Large Language Models","created":"2026-03-06","cited_by_count":2,"area":"alignment"},{"id":"2603.18329","title":"FaithSteer-BENCH: A Deployment-Aligned Stress-Testing Benchmark for Inference-Time Steering","created":"2026-03-18","cited_by_count":2,"area":null},{"id":"2605.08326","title":"LLM Advertisement based on Neuron Auctions","created":"2026-05-08","cited_by_count":2,"area":null},{"id":"2605.31245","title":"Toward Identifiable Sparse Autoencoders","created":"2026-05-29","cited_by_count":2,"area":"mechinterp"},{"id":"2508.10404","title":"Layer-Wise Perturbations via Sparse Autoencoders for Adversarial Text Generation","created":"2025-08-14","cited_by_count":1,"area":"safety"},{"id":"2509.25843","title":"ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack","created":"2025-09-30","cited_by_count":1,"area":"safety"},{"id":"2512.06655","title":"Graph-Regularized Sparse Autoencoders for LLM Safety Steering","created":"2025-12-07","cited_by_count":1,"area":"safety"},{"id":"2603.05772","title":"Depth Charge: Jailbreak Large Language Models from Deep Safety Attention Heads","created":"2026-03-06","cited_by_count":1,"area":"safety"},{"id":"2603.13359","title":"From Refusal Tokens to Refusal Control: Discovering and Steering Category-Specific Refusal Directions","created":"2026-03-09","cited_by_count":1,"area":"alignment"},{"id":"2604.04946","title":"Sparse Autoencoders as a Steering Basis for Phase Synchronization in Graph-Based CFD Surrogates","created":"2026-03-28","cited_by_count":1,"area":"mechinterp"},{"id":"2604.06263","title":"Incentive-Aware Multi-Fidelity Optimization for Generative Advertising in Large Language Models","created":"2026-04-07","cited_by_count":1,"area":null},{"id":"2605.16362","title":"When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search","created":"2026-05-09","cited_by_count":1,"area":"mechinterp"},{"id":"2605.17413","title":"Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications","created":"2026-05-17","cited_by_count":1,"area":"alignment"},{"id":"2605.24154","title":"Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMs","created":"2026-05-22","cited_by_count":1,"area":"alignment"},{"id":"2606.15092","title":"High-Dimensional Random Projection for Activation Steering in Language Models","created":"2026-06-13","cited_by_count":1,"area":"mechinterp"},{"id":"2606.18322","title":"SAE Interventions are Unreliable: Post-Intervention Recovery of Suppressed Behavior","created":"2026-06-16","cited_by_count":1,"area":"mechinterp"},{"id":"2606.31699","title":"Look But Don't Touch with Sparse Autoencoders for Unlearning in Diffusion Models","created":"2026-06-30","cited_by_count":1,"area":"diffusion"},{"id":"2608.28233","title":"REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features","created":"2026-08-28","cited_by_count":1,"area":"safety"},{"id":"2609.16204","title":"Decoy Direction Optimization: A Post-Hoc Defense Against LLM Abliteration","created":"2026-09-14","cited_by_count":1,"area":"safety"},{"id":"2604.08846","title":"Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs","created":"2026-04-10","cited_by_count":0,"area":"mllm"},{"id":"2604.18756","title":"Towards Understanding the Robustness of Sparse Autoencoders","created":"2026-04-20","cited_by_count":0,"area":"safety"},{"id":"2605.01827","title":"Referring Multiple Regions with Large Multimodal Models via Contextual Latent Steering","created":"2026-05-03","cited_by_count":0,"area":"mllm"},{"id":"2605.20262","title":"Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing","created":"2026-05-18","cited_by_count":0,"area":null},{"id":"2605.21609","title":"CR4T: Rewrite-Based Guardrails for Adolescent LLM Safety","created":"2026-05-20","cited_by_count":0,"area":"safety"},{"id":"2605.23040","title":"Steered Generation via Gradient-Based Optimization on Sparse Query Features","created":"2026-05-21","cited_by_count":0,"area":"mechinterp"},{"id":"2605.24977","title":"Universal Boosts, Specific Suppressors: Sparse Autoencoder Steering of Medical Vision-Language Models","created":"2026-05-24","cited_by_count":0,"area":"mllm"},{"id":"2606.03601","title":"DDOR: Delta Debugging for Explainable Overrefusal Testing and Repair","created":"2026-06-02","cited_by_count":0,"area":"alignment"},{"id":"2606.08365","title":"Pre-Intervention Prediction of Sparse Autoencoder Steering Side Effects","created":"2026-06-06","cited_by_count":0,"area":"mechinterp"},{"id":"2606.26396","title":"At the Edge of Understanding: Sparse Autoencoders Trace The Limits of Transformer Generalization","created":"2026-06-24","cited_by_count":0,"area":"mechinterp"},{"id":"2606.26620","title":"Discovering Millions of Interpretable Features with Sparse Autoencoders","created":"2026-06-25","cited_by_count":0,"area":"mechinterp"},{"id":"2606.27199","title":"Forecasting With LLMs: Improved Generalization Through Feature Steering","created":"2026-06-25","cited_by_count":0,"area":null},{"id":"2608.05409","title":"Mood Matters: How Syntactic Sensitivity Undermines Safety Alignment","created":"2026-08-05","cited_by_count":0,"area":"alignment"},{"id":"2608.09928","title":"Multimodal Model Diffing for Feature Discovery and Control","created":"2026-08-10","cited_by_count":0,"area":"mllm"},{"id":"2608.18093","title":"Abliteration Mitigation via Refusal Aliases","created":"2026-06-07","cited_by_count":0,"area":null},{"id":"2608.25390","title":"Refusal geometry reflects refusal training: diverse refusal prefixes can raise stable rank and weaken refusal vector ablation attacks","created":"2026-08-26","cited_by_count":0,"area":"safety"},{"id":"2609.18471","title":"First Token Matters: Understanding Safety Collapse in Large Reasoning Models","created":"2026-09-16","cited_by_count":0,"area":"lr"},{"id":"2609.22153","title":"SafeTune: A Unified Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMs","created":"2026-08-26","cited_by_count":0,"area":null},{"id":"2609.22224","title":"From Trait Vectors to Circuits: Tracing Refusal and Sycophancy Through Language Models","created":"2026-09-02","cited_by_count":0,"area":null},{"id":"2609.32457","title":"Write Back the $\\Delta$: Revisiting the Same Tokens with Fresh Representations","created":"2026-09-26","cited_by_count":0,"area":null},{"id":"2609.39701","title":"Values as Style: Disentangling Values from Semantics with One-Way Mixing for Low-Damage LLM Steering","created":"2026-09-30","cited_by_count":0,"area":null},{"id":"2610.01062","title":"Kernelized Activation Steering","created":"2026-10-01","cited_by_count":0,"area":"mechinterp"},{"id":"2610.03498","title":"Detect and Suppress: A Mechanistic Defense against Adversarial Patches in VLA Models","created":"2026-10-02","cited_by_count":0,"area":"vla"},{"id":"2610.04905","title":"ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection","created":"2026-10-04","cited_by_count":0,"area":"mechinterp"},{"id":"2610.06049","title":"Backdooring Sparse Autoencoders","created":"2026-10-05","cited_by_count":0,"area":"safety"},{"id":"2610.19114","title":"Genesis Steering: Better Activation Steering Through Training Data","created":"2026-10-13","cited_by_count":0,"area":"mechinterp"},{"id":"2610.20139","title":"Sparse Autoencoders for Mechanistic Interpretability in NLP: A Critical Survey of Recoverability, Interpretability, and Utility","created":"2026-10-14","cited_by_count":0,"area":"mechinterp"}],"next_cursor":null}