ContextBench: Modifying Contexts for Targeted Latent Activation

2025/06/15 by Graham, Robert, Stevinson, Edward, Richter, Leo +3
Computer Science · #Artificial Intelligence (cs.AI) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Data Classification #Topic Modeling

paper · pdf · doi:10.48550/arxiv.2506.15735

openalex publication_date 2025/06/15 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Identifying inputs that trigger specific behaviours or latent features in language models could have a wide range of safety use cases. We investigate a class of methods capable of generating targeted, linguistically fluent inputs that activate specific latent features or elicit model behaviours. We formalise this approach as context modification and present ContextBench -- a benchmark with tasks assessing core method capabilities and potential safety applications. Our evaluation framework measures both elicitation strength (activation of latent features or behaviours) and linguistic fluency, highlighting how current state-of-the-art methods struggle to balance these objectives. We enhance Evolutionary Prompt Optimisation (EPO) with LLM-assistance and diffusion model inpainting, and demonstrate that these variants achieve state-of-the-art performance in balancing elicitation effectiveness and fluency.

Citations

• Adversarial Attacks Leverage Interference Between Features in Superposition
• Large Language Diffusion Models
• Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoders
• StablePrompt: Automatic Prompt Tuning using Reinforcement Learning for Large Language Models
• Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2
• Future Events as Backdoor Triggers: Investigating Temporal Vulnerabilities in LLMs
• External Invariants: A Cryptographic Trust Architecture for Institutional AI Inference
• AI Sandbagging: Language Models can Strategically Underperform on Evaluations
• Competition Report: Finding Universal Jailbreak Backdoors in Aligned LLMs
• Fast Adversarial Attacks on Language Models In One GPU Minute
• MORL-Prompt: An Empirical Analysis of Multi-Objective Reinforcement Learning for Discrete Prompt Optimization
• Black-Box Access is Insufficient for Rigorous AI Audits
• Fluent dreaming for language models
• Unsupervised Approach to Evaluate Sentence-Level Fluency: Do We Really Need Reference?
• AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models
• Code of "Sirens' Whisper: Inaudible Near-Ultrasonic Jailbreaks of Speech-Driven LLMs"
• Automatically Auditing Large Language Models via Discrete Optimization
• Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery
• Governance Architecture for Neural Network Superposition: A Structural Solution to Hallucination via Routing and Interference Filtering
• The Power of Scale for Parameter-Efficient Prompt Tuning
• Disentangling Adversarial Robustness and Generalization
• Sentence-Level Fluency Evaluation: References Help, But Can Be Spared!
• Interpretable Textual Neuron Representations for NLP
• Crowdsourcing Multiple Choice Science Questions

Related

• Unleashing the potential of prompt engineering for large language models
• Context-Tuning: Learning Contextualized Prompts for Natural Language Generation
• A Brief History of Prompt: Leveraging Language Models. (Through Advanced Prompting)
• Task-driven Layerwise Additive Activation Intervention
• Invasive Context Engineering to Control Large Language Models