Keep Calm and Avoid Harmful Content: Concept Alignment and Latent Manipulation Towards Safer Answers

2025/10/14 by Ruben Belo, Belo, Ruben, Marta Guimarães +3
Computer Science · #Advanced Text Analysis Techniques #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2510.12672

openalex publication_date 2025/10/14 · openalex created_date 2025/10/17 · openalex updated_date 2026/07/28

Abstract

Large Language Models are susceptible to jailbreak attacks that bypass built-in safety guardrails (e.g., by tricking the model with adversarial prompts). We propose Concept Alignment and Concept Manipulation CALM, an inference-time method that suppresses harmful concepts by modifying latent representations of the last layer of the model, without retraining. Leveraging concept whitening technique from Computer Vision combined with orthogonal projection, CALM removes unwanted latent directions associated with harmful content while preserving model performance. Experiments show that CALM reduces harmful outputs and outperforms baseline methods in most metrics, offering a lightweight approach to AI safety with no additional training data or model fine-tuning, while incurring only a small computational overhead at inference.

Citations

• Controlling Large Language Models Through Concept Activation Vectors
• SCAR: Sparse Conditioned Autoencoders for Concept Detection and Steering in LLMs
• The Llama 3 Herd of Models
• Gemma 2: Improving Open Language Models at a Practical Size
• Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs
• Whitening Not Recommended for Classification Tasks in LLMs
• External Invariants: A Cryptographic Trust Architecture for Institutional AI Inference
• Improving Alignment and Robustness with Circuit Breakers
• Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity
• Spectral Editing of Activations for Large Language Model Alignment
• A Language Model's Guide Through Latent Space
• Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
• Revisiting Jailbreaking for Large Language Models: A Representation Engineering Perspective
• Steering Llama 2 via Contrastive Activation Addition
• Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation
• Jailbreak and Guard Aligned Language Models with Only Few In-Context Demonstrations
• Representation Engineering: A Top-Down Approach to AI Transparency
• GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts
• Sparse Autoencoders Find Highly Interpretable Features in Language Models
• Red-Teaming Large Language Models using Chain of Utterances for Safety-Alignment
• Universal and Transferable Adversarial Attacks on Aligned Language Models
• Prompt Injection attack against LLM-integrated Applications
• Inference-Time Intervention: Eliciting Truthful Answers from a Language Model
• LEACE: Perfect linear concept erasure in closed form
• Whitening-based Contrastive Learning of Sentence Embeddings
• Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
• Pretraining Language Models with Human Preferences
• Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
• Whitening Sentence Representations for Better Semantics and Faster Retrieval
• BERT has a Moral Compass: Improvements of ethical and moral values of machines

Related

• Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models
• Attention Shift: Steering AI Away from Unsafe Content
• Reasoned Safety Alignment: Ensuring Jailbreak Defense via Answer-Then-Check
• EASE: Practical and Efficient Safety Alignment for Small Language Models
• InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance