I attended COLM 2025 for a very specific reason: I study how people write specifications for AI systems, and these days the way we write specifications is increasingly by learning specifications from human preferences (to my chagrin, as I quite enjoy an explicitly-defined objective function). Intuitively, collecting and learning from human preferences feels like an intensely democratic process: we are listening to the perspectives of the people and learning from their perspectives. Alas, I wish it were so. 

It is not uncommon to screen the people who provide preferences for scaling AI systems to assess whether their preferences are agreeable to the companies that are hiring them to label data. It is also not uncommon to then provide instructions to these chosen people about what their preferences should mean—words like helpful, harmless, and honest abound, though I would caution anyone against asking a machine learning researcher what these words mean. After choosing people with agreeable preferences, telling them what their preferences should mean, and asking these people to label data in accordance with those definitions, we then learn an objective function from their preferences—but, to do so, we impose inductive biases of our choosing. While choosing an inductive bias is necessary to make the problem of learning an objective function from preferences well-specified, this choice is impactful and may or may not preserve either their original or their skewed preference.

I find this process endlessly frustrating: to train our LLMs, we’ve obscured the meaning of human preference to the point of being unrecognizable. To reconsider how preferences can be converted into a more pure currency of democracy, I am eager to establish collaborations with public opinion researchers, survey statisticians, and demographers; so, I traveled to COLM to join a panel at the NLPOR Workshop, a community effort to bridge NLP and Public Opinion Research. Building these bridges will require developing shared languages—both conceptual and methodological—so that insights about representation, sampling, and measurement from the social sciences can directly inform how we model human preferences in AI.

Attending COLM 2025 also made me reflect on how the nature of machine learning research itself is changing. Throughout my graduate education, I felt more and more compelled to share my research with the giant AI conferences: AAAI, NeurIPS, ICML, ICLR. But, in some ways, it feels like these bubbles have finally burst: it now feels far more enjoyable and meaningful to share my work with smaller, niche communities that are capable of engaging more deeply. Be it COLM, RLC (the reinforcement learning conference), or HRI (the human-robot interaction conference), these smaller communities are where I find my intellectual fulfillment and joy. 

After COLM, I touched the grass in Montreal by taking a joyous Autumn tandem bike ride with my partner on the Le Petit Train du Nord trail and eating copious amounts of cheese.