{"total":63,"rows":[{"id":"2406.11717","title":"Refusal in Language Models Is Mediated by a Single Direction","created":"2024-06-17","cited_by_count":1032,"area":"safety"},{"id":"2403.05030","title":"Defending Against Unforeseen Failure Modes with Latent Adversarial Training","created":"2024-03-08","cited_by_count":88,"area":"safety"},{"id":"2507.11878","title":"LLMs Encode Harmfulness and Refusal Separately","created":"2025-07-16","cited_by_count":88,"area":"safety"},{"id":"2406.15518","title":"Steering Without Side Effects: Improving Post-Deployment Control of Language Models","created":"2024-06-21","cited_by_count":46,"area":"safety"},{"id":"2501.01765","title":"SaLoRA: Safety-Alignment Preserved Low-Rank Adaptation","created":"2025-01-03","cited_by_count":42,"area":"safety"},{"id":"2406.12094","title":"Who's asking? User personas and the mechanics of latent misalignment","created":"2024-06-17","cited_by_count":40,"area":"safety"},{"id":"2308.08090","title":"Separate the Wheat from the Chaff: Model Deficiency Unlearning via Parameter-Efficient Module Operation","created":"2023-08-16","cited_by_count":33,"area":"safety"},{"id":"2311.09433","title":"Trojan Activation Attack: Red-Teaming Large Language Models using Activation Steering for Safety-Alignment","created":"2023-11-15","cited_by_count":30,"area":"safety"},{"id":"2411.03343","title":"What Features in Prompts Jailbreak LLMs? Investigating the Mechanisms Behind Attacks","created":"2024-11-02","cited_by_count":29,"area":"safety"},{"id":"2411.02631","title":"Extracting Unlearned Information from LLMs with Activation Steering","created":"2024-11-04","cited_by_count":21,"area":"safety"},{"id":"2503.00187","title":"Steering Dialogue Dynamics for Robustness against Multi-turn Jailbreaking Attacks","created":"2025-02-28","cited_by_count":20,"area":"safety"},{"id":"2405.04160","title":"A Causal Explainable Guardrails for Large Language Models","created":"2024-05-07","cited_by_count":18,"area":"safety"},{"id":"2507.09709","title":"Large Language Models Encode Semantics and Alignment in Linearly Separable Representations","created":"2025-07-13","cited_by_count":17,"area":"safety"},{"id":"2410.06913","title":"Utilize the Flow before Stepping into the Same River Twice: Certainty Represented Knowledge Flow for Refusal-Aware Instruction Tuning","created":"2024-10-09","cited_by_count":10,"area":"safety"},{"id":"2505.16252","title":"Does Localization Inform Unlearning? A Rigorous Examination of Local Parameter Attribution for Knowledge Unlearning in Language Models","created":"2025-05-22","cited_by_count":10,"area":"safety"},{"id":"2506.05451","title":"Interpretation Meets Safety: A Survey on Interpretation Methods and Tools for Improving LLM Safety","created":"2025-06-05","cited_by_count":10,"area":"safety"},{"id":"2603.24543","title":"Analysing the Safety Pitfalls of Steering Vectors","created":"2026-03-25","cited_by_count":10,"area":"safety"},{"id":"2505.07610","title":"Concept-Level Explainability for Auditing & Steering LLM Responses","created":"2025-05-12","cited_by_count":6,"area":"safety"},{"id":"2512.16602","title":"Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics","created":"2025-12-18","cited_by_count":6,"area":"safety"},{"id":"2602.12418","title":"Sparse Autoencoders are Capable LLM Jailbreak Mitigators","created":"2026-02-12","cited_by_count":6,"area":"safety"},{"id":"2410.10150","title":"Jailbreak Instruction-Tuned LLMs via end-of-sentence MLP Re-weighting","created":"2024-10-14","cited_by_count":5,"area":"safety"},{"id":"2604.08169","title":"Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence","created":"2026-04-09","cited_by_count":4,"area":"safety"},{"id":"2603.27412","title":"The Geometry of Harmful Intent: Training-Free Anomaly Detection via Angular Deviation in LLM Residual Streams","created":"2026-03-28","cited_by_count":3,"area":"safety"},{"id":"2510.07775","title":"The Unintended Trade-off of AI Alignment:Balancing Hallucination Mitigation and Safety in LLMs","created":"2025-10-09","cited_by_count":2,"area":"safety"},{"id":"2603.20206","title":"Enhancing Safety of Large Language Models via Embedding Space Separation","created":"2026-03-01","cited_by_count":2,"area":"safety"},{"id":"2605.12765","title":"Inference-Time Machine Unlearning via Gated Activation Redirection","created":"2026-05-12","cited_by_count":2,"area":"safety"},{"id":"2606.03291","title":"Multilingual Unlearning in LLMs: Transfer, Dynamics, and Reversibility","created":"2026-06-02","cited_by_count":2,"area":"safety"},{"id":"2607.19806","title":"OPIUM: Mitigating Steering Externalities and Over-Refusal via Dual Objective Latent Optimization","created":"2026-07-22","cited_by_count":2,"area":"safety"},{"id":"2502.05945","title":"Head-Specific Intervention Can Induce Misaligned AI Coordination in Large Language Models","created":"2025-02-09","cited_by_count":1,"area":"safety"},{"id":"2511.14195","title":"N-GLARE: An Non-Generative Latent Representation-Efficient LLM Safety Evaluator","created":"2025-11-18","cited_by_count":1,"area":"safety"},{"id":"2601.12460","title":"TrojanPraise: Jailbreak LLMs via Benign Fine-Tuning","created":"2026-01-18","cited_by_count":1,"area":"safety"},{"id":"2601.21702","title":"Beyond Forgetting: Representation Misdirection Elicits Controllable Side Behaviors and Capabilities","created":"2026-01-29","cited_by_count":1,"area":"safety"},{"id":"2603.08234","title":"The Struggle Between Continuation and Refusal: A Mechanistic Analysis of the Continuation-Triggered Jailbreak in LLMs","created":"2026-03-09","cited_by_count":1,"area":"safety"},{"id":"2603.23171","title":"Adaptively Robust LLM Monitoring via Activation Watermarking","created":"2026-03-24","cited_by_count":1,"area":"safety"},{"id":"2603.24857","title":"AI Security in the Foundation Model Era: A Comprehensive Survey from a Unified Perspective","created":"2026-03-25","cited_by_count":1,"area":"safety"},{"id":"2604.15780","title":"Pruning Unsafe Tickets: A Resource-Efficient Framework for Safer and More Robust LLMs","created":"2026-04-17","cited_by_count":1,"area":"safety"},{"id":"2605.14868","title":"Fast Adversarial Attacks with Gradient Prediction","created":"2026-05-14","cited_by_count":1,"area":"safety"},{"id":"2608.03201","title":"When Refusal Looks Safe: The Refusal-Cue Shortcut in Safety Guard Models","created":"2026-08-04","cited_by_count":1,"area":"safety"},{"id":"2608.28233","title":"REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features","created":"2026-08-28","cited_by_count":1,"area":"safety"},{"id":"2510.12672","title":"Keep Calm and Avoid Harmful Content: Concept Alignment and Latent Manipulation Towards Safer Answers","created":"2025-10-14","cited_by_count":0,"area":"safety"},{"id":"2602.20102","title":"BarrierSteer: LLM Safety via Learning Barrier Steering","created":"2026-02-23","cited_by_count":0,"area":"safety"},{"id":"2604.21571","title":"Separable Expert Architecture: Toward Privacy-Preserving LLM Personalization via Composable Adapters and Deletable User Proxies","created":"2026-04-23","cited_by_count":0,"area":"safety"},{"id":"2604.28129","title":"Latent Adversarial Detection: Adaptive Probing of LLM Activations for Multi-Turn Attack Detection","created":"2026-04-30","cited_by_count":0,"area":"safety"},{"id":"2605.05687","title":"DataDignity: Training Data Attribution for Large Language Models","created":"2026-05-07","cited_by_count":0,"area":"safety"},{"id":"2606.05958","title":"Steering Vectors are an Adversarial Attack Surface","created":"2026-06-04","cited_by_count":0,"area":"safety"},{"id":"2606.25750","title":"RAS: Measuring LLM Safety Through Refusal Alignment","created":"2026-06-24","cited_by_count":0,"area":"safety"},{"id":"2607.15218","title":"When Words Are Safe But Actions Kill: Probing Physical Jailbreak Beyond Textual Jailbreak in Hidden-State Risk Space","created":"2026-07-16","cited_by_count":0,"area":"safety"},{"id":"2607.19366","title":"Geometry-Guided Constraint Learning for LLM Safety Classification","created":"2026-06-09","cited_by_count":0,"area":"safety"},{"id":"2607.22595","title":"xMIx: High-Performance Serving-Time Platform for Mechanistic Interpretability Apps","created":"2026-06-11","cited_by_count":0,"area":"safety"},{"id":"2608.23873","title":"Semantic Overlays: Mitigating Prompt Injection with Annotations Beyond Tokens and Steering Vectors","created":"2026-08-24","cited_by_count":0,"area":"safety"},{"id":"2608.30748","title":"The Fragility of Jailbreak Robustness Across Operational States","created":"2026-08-31","cited_by_count":0,"area":"safety"},{"id":"2609.16229","title":"Test-Time Unlearning via Sparse Autoencoder","created":"2026-09-14","cited_by_count":0,"area":"safety"},{"id":"2609.21996","title":"A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal","created":"2026-09-18","cited_by_count":0,"area":"safety"},{"id":"2609.32602","title":"AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion","created":"2026-09-26","cited_by_count":0,"area":"safety"},{"id":"2609.34514","title":"How to Tame a Multi-Headed Hydra? Adaptive Multi-Category Safety Steering for Large Language Models","created":"2026-09-28","cited_by_count":0,"area":"safety"},{"id":"2609.35913","title":"UNBIND: UNlearning By INference-time Directional Steering for Code LLMs","created":"2026-09-28","cited_by_count":0,"area":"safety"},{"id":"2609.36657","title":"Constitutional adapters: Inference-time interventions for misalignment and misuse","created":"2026-09-29","cited_by_count":0,"area":"safety"},{"id":"2609.39882","title":"LLM Persona Unlearning","created":"2026-09-30","cited_by_count":0,"area":"safety"},{"id":"2610.04860","title":"Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate","created":"2026-10-04","cited_by_count":0,"area":"safety"},{"id":"2610.06064","title":"TrustMI: Causally controlling how assistants trust their users","created":"2026-10-05","cited_by_count":0,"area":"safety"},{"id":"2610.10276","title":"PatchBench: Measuring Collateral Damage in Activation Patching","created":"2026-10-07","cited_by_count":0,"area":"safety"},{"id":"2610.15365","title":"TRACE: Temporal Risk Assessment and Concept Editing for Safer Language Generation","created":"2026-10-11","cited_by_count":0,"area":"safety"},{"id":"2610.17880","title":"Cross-Lingual Transfer of Evil Vectors Is Directionally Asymmetric","created":"2026-10-13","cited_by_count":0,"area":"safety"}],"next_cursor":null}