• Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control

  We examine safety-tuned LLMs and discover representation vectors for measuring and controlling censorship imposed through refusal and thought suppression in model outputs.

  36 min read   ·   April 24, 2025

  2025   ·   LLMs   censorship     ·   paper  

• The Mismeasure of Man and Models

  Evaluating Allocational Harms in Large Language Models

  56 min read   ·   August 10, 2024

  2024   ·   LLMs   bias   fairness     ·   paper  

• Adjectives Can Reveal Gender Biases Within NLP Models

  We extend WinoBias dataset by incorporating gender-associated adjectives and reveal underlying gender bias in GPT-3.5 model.

  10 min read   ·   August 17, 2023

  2023   ·   adversarial exmples   LLMs   generative AI   bias     ·   paper  

• Balancing Tradeoffs between Fickleness and Obstinacy in NLP Models

  We introduce Balanced Adversarial Training (BAT) to train models that are robust to two different types of adversarial examples.

  5 min read   ·   November 13, 2022

  2022   ·   adversarial exmples     ·   paper