Sangmitra Madhusudan

Robert Morabito, Sangmitra Madhusudan, Tyler McDonald et al.

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.

Nikta Gohari Sadr, Sangmitra Madhusudan, Hassan Sajjad et al.

While zero-shot instructional prompts like "Let's think step-by-step" have revolutionized Large Language Model performance, a fundamental question remains unanswered: which specific words drive their remarkable effectiveness? We introduce the ZIP score (Zero-shot Importance of Perturbation), the first systematic method to quantify individual word importance in instructional prompts through controlled perturbations including synonym replacement, co-hyponym substitution, and strategic removal. Our analysis across four flagship models, seven widely-adopted prompts, and multiple task domains reveals four key findings: (1) Task-specific word hierarchies exist where mathematical problems prioritize "step-by-step" while reasoning tasks favor "think"; (2) Proprietary models show superior alignment with human intuitions compared to open-source alternatives; (3) Nouns dominate importance rankings, consistently representing the majority of significant words; and (4) Word importance inversely correlates with model performance, indicating prompts have greatest impact where models struggle most. Beyond revealing these patterns, we establish the first ground-truth benchmark for prompt interpretability through 20 validation prompts with predetermined key words, where ZIP achieves 90% accuracy versus LIME's 60%. Our findings advance prompt science, the study of how language shapes model behavior, providing both practical insights for prompt engineering and theoretical understanding of word-level effects in LLMs.

Sangmitra Madhusudan, Robert Morabito, Skye Reid et al.

Books, while often rich in cultural insights, can also mirror societal biases of their eras - biases that Large Language Models (LLMs) may learn and perpetuate during training. We introduce a novel method to trace and quantify these biases using fine-tuned LLMs. We develop BookPAGE, a corpus comprising 593 fictional books across seven decades (1950-2019), to track bias evolution. By fine-tuning LLMs on books from each decade and using targeted prompts, we examine shifts in biases related to gender, sexual orientation, race, and religion. Our findings indicate that LLMs trained on decade-specific books manifest biases reflective of their times, with both gradual trends and notable shifts. For example, model responses showed a progressive increase in the portrayal of women in leadership roles (from 8% to 22%) from the 1950s to 2010s, with a significant uptick in the 1990s (from 4% to 12%), possibly aligning with third-wave feminism. Same-sex relationship references increased markedly from the 1980s to 2000s (from 0% to 10%), mirroring growing LGBTQ+ visibility. Concerningly, negative portrayals of Islam rose sharply in the 2000s (26% to 38%), likely reflecting post-9/11 sentiments. Importantly, we demonstrate that these biases stem mainly from the books' content and not the models' architecture or initial training. Our study offers a new perspective on societal bias trends by bridging AI, literary studies, and social science research.