Stress Testing Concept Erasure with Large Language Model Agents
2026-07-20 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors address the problem of checking if a trained AI model has truly forgotten certain concepts, which is important for safe AI use. They propose a new method called STACE that uses multiple language models to automatically create and test challenging questions about the removed concepts, improving the way evaluations cover potential weaknesses. Their experiments show STACE works better than previous methods across different models and erased concepts, and it can even be used for other tasks like testing AI security. This approach helps make sure AI models really forget targeted information as intended.
concept erasuregenerative modellarge language model (LLM)stress testingadaptive hypothesis searchevaluation metricstext-to-image (T2I) modelsmodel robustnessjailbreakingsemantic concepts
Authors
Yuyang Xue, Feng Chen, Zhihua Liu, Edward Moroshko, Jingyu Sun, Steven McDonagh, Sotirios A. Tsaftaris
Abstract
Concept erasure aims to remove semantic concepts from a trained generative model and is increasingly important for responsible AI deployment. However, verifying whether a model has robustly removed targeted concepts remains a critical challenge. Existing evaluation methods are typically pre-defined and static, failing to expose vulnerabilities under diverse natural-language probes and challenging conditions. Moreover, manually designed evaluation strategies can be biased and difficult to scale. We posit that concept erasure evaluation is best formulated as an adaptive hypothesis search, operationalised by agents that iteratively propose, critique, and verify tests to systematically expand coverage of failure modes. To this end, we propose Stress Testing Agents for Concept Erasure (STACE), a framework that autonomously stress-tests concept-erased models using multiple Large Language Model (LLM) agents, by iteratively generating and verifying stress-testing hypotheses grounded by external knowledge. We also introduce a suite of metrics for assessing the performance and efficiency of LLM-agent-powered stress-testing frameworks. Our extensive experiments show that STACE outperforms five LLM-based evaluation baselines on four concept categories. Further analysis across two T2I models, six concept erasure approaches, and various erasure strengths show that STACE is robust for different settings. We also show that STACE can be adapted beyond concept erasure evaluation to other problem domains, such as LLM jailbreaking. Our code is available anonymously.