STOP! Benchmarking Large Language Models with Sensitivity Testing on Offensive Progressions
This work addresses the need for more comprehensive bias assessment in LLMs to create fairer models, though it is incremental as it builds on existing evaluation frameworks.
The authors tackled the problem of evaluating biases in Large Language Models by introducing the STOP dataset, which includes 450 offensive progressions across 9 demographics, and found that even top models detect bias inconsistently with success rates from 19.3% to 69.8%, while alignment with human judgments improved performance on sensitive tasks by up to 191%.
Mitigating explicit and implicit biases in Large Language Models (LLMs) has become a critical focus in the field of natural language processing. However, many current methodologies evaluate scenarios in isolation, without considering the broader context or the spectrum of potential biases within each situation. To address this, we introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, which includes 450 offensive progressions containing 2,700 unique sentences of varying severity that progressively escalate from less to more explicitly offensive. Covering a broad spectrum of 9 demographics and 46 sub-demographics, STOP ensures inclusivity and comprehensive coverage. We evaluate several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3. Our findings reveal that even the best-performing models detect bias inconsistently, with success rates ranging from 19.3% to 69.8%. We also demonstrate how aligning models with human judgments on STOP can improve model answer rates on sensitive tasks such as BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or even improving performance. STOP presents a novel framework for assessing the complex nature of biases in LLMs, which will enable more effective bias mitigation strategies and facilitates the creation of fairer language models.