ICML 2026

I spent the last week at the International Conference on Machine Learning (ICML) in Korea, meeting exciting new people and learning about their research. Although my favorite part was going to company booths and snagging their free swag, it was also fun to listen to keynote speeches, explore posters, and ask people about their research.

One speaker that really stood out to me was Milind Tambe, a computer science professor at Harvard. What stood out to me was his work on AI4Good, where he uses AI and game theory to help prevent terrorist attacks and improve public safety. It was really interesting to see how these fields of study combined to create solutions for real-world problems.

Along with all of these amazing experiences, I also had the honor of presenting my own research!

📄 Read the full paper on OpenReview →

My research presents a benchmark for evaluating political bias in large language models. Existing benchmarks fail in two main ways.

They use static prompting. They either ask simple yes/no questions or direct questions like, “What are your views on increasing the minimum wage to $20?”

Why this fails? LLMs are trained to be relatively neutral, so asking a blatant political question will often yield a response such as, “As an AI, I do not have political views,” or “I can explain both perspectives.” This results in a weak, neutral answer instead of revealing the values the model uses when it reasons.

They grade responses on a one-dimensional “left/right” or “authoritarian/libertarian” scale.

Why this fails? By grading model responses on a single political axis, you miss why LLMs arrive at the positions they do. Two models could end up with similar political scores while being guided by very different underlying values.

Our research creates multi-stage role-play scenarios, where the LLM has to experience situations for itself and make difficult tradeoffs. This forces the models to reason through their decisions and provide more detailed explanations.

Next, rather than asking judge LLMs to give a single political score, we ask them to score responses across 10 personality and value dimensions that together map to the political spectrum. This lets us understand not just where a model falls politically, but why. Across eight major LLMs, we found that these interactive scenarios elicited stronger and more detailed value expression than traditional direct political questions.

Overall, presenting this work at ICML was an incredible experience. It was exciting to share our research, answer questions from other researchers, and learn about all of the amazing work happening in AI.