ContextBench: Modifying Contexts for Targeted Latent Activation

TL;DR

This paper introduces ContextBench, a benchmark for evaluating methods that modify input contexts to activate specific latent features in language models, highlighting challenges and improvements in balancing activation strength and linguistic fluency.

Contribution

The paper formalizes context modification as a method for targeted activation and presents ContextBench for evaluating such techniques, along with enhanced EPO methods achieving state-of-the-art results.

Findings

Current methods struggle to balance activation strength and fluency.

Enhanced EPO with LLM assistance improves targeted activation.

ContextBench provides a standardized evaluation framework.

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.