AdversaBench: Automated LLM Red-Teaming with Multi-Judge Confirmation and Cross-Model Transferability

A new automated pipeline called AdversaBench can systematically expose failures in large language models by mutating seed prompts and confirming the resulting errors through a three-judge panel, according to research published on arXiv [1]. The end-to-end red-teaming system generates hard inputs by applying five structured operators to seed prompts, then queries a target model [1]. A panel of three judges, with a meta-judge acting as tiebreaker, confirms whether each output constitutes a genuine failure [1]. In experiments across 45 seeds spanning reasoning, instruction-following, and tool-use categories, every seed produced at least one confirmed failure [1]. The study found that operator effectiveness varies sharply by category. The inject_distractor operator scored a mean reward of 0.00 on instruction-following seeds but reached 0.80 to 0.83 on reasoning and tool-use tasks [1]. The binary failure rate also masked underlying difficulty: instruction-following seeds required an average of 2.4 attacker iterations, compared with 1.1 iterations for the other categories [1]. Pairwise agreement among judges ranged from 80 to 87 percent, yet Cohen’s kappa remained near zero because of label skew [1]. The researchers note that category-level disagreement rates offer a more informative signal for evaluation reliability [1]. Adversarial prompts generated against Llama 3.1 8B transferred zero-shot to Llama 3.3 70B, indicating the mutations exploit general behavioral patterns rather than weaknesses tied to a specific model [1]. The code, dataset, and analysis scripts have been released on GitHub [1]. Large language models are machine learning systems with many parameters, trained on vast text corpora for tasks such as language generation [7]. The adversarial evaluation of such models has drawn increased attention as companies and research labs deploy them in consumer-facing products. Chinese firm DeepSeek, for instance, launched its R1 model in January 2025 with performance comparable to OpenAI’s GPT-4, while claiming training costs far below those of Western competitors [6]. The AdversaBench paper appears on arXiv, the preprint repository that since 2022 has integrated with Hugging Face Spaces to let authors and the community attach interactive demos directly to paper pages [3][4]. Researchers can link a Space to a paper by including the paper’s URL in the Space README file or by associating the Space with a model that itself cites the paper [5].