~/research

Calvin and Hobbes panel: Calvin looks up at Hobbes and says, she says we don't really understand it, but there are many things we don't understand, and we just have to do the best we can with the knowledge we have.

As a relatively naive undergrad, my research interests are broad and somewhat unfettered by trivial constraints like logic, reason, and capability.

At a high level, my work focuses on Human-Centered AI (HCAI) and agent behavior. More specifically, how humans use AI, what effects AI has on humans, and developing better AI systems for human interaction (safety, intent alignment, ease of use/human-ai interaction). I maintain some over-flowing notes of personal research ideas and topics that can be found below. If you have any interest in talking anything related to research, have a low barrier for reaching out at: mprince [at] uw [dot] edu

research interests

My overarching research interests might seem a bit unrelated and broad; that's because they are. Humans are historically terrible at predicting the future, and I am no different, so I want to learn the intersection of the most impactful research done today, then apply that knowledge to maximize potential for a good future, however that world looks.

  • Human-Centered AI
  • Human-AI Interaction
  • AI Safety / Alignment
  • Agentic Systems / Multi-Agent Systems
  • Agent Reasoning / Decision-Making
  • LLM/ML Architectures (Transformer, SSM, LNN, etc.)
  • RL / Post-Training
  • Continual Learning
  • Applied Social Sciences
  • Economics (Behavioral + Strategic) in AI

All in all, a main goal and research direction is defining human value, culture, productivity, and labor in a post-AGI world: what it means to be human, how humans play a role where they are technically outmatched in many disciplines, and what type of labor is important in these systems.

As an extension to this, my research interests come together with the intent to improve the human-condition amongst increasingly capable AI. I wish for a future where technology empowers and supports humanity, not destroys it.

## past research

My past research has largely focused on agent reasoning, particularly in social situations:

  1. 2024–25

    In 2024, I earned a merit scholarship and funded compute to work under a Berkeley grad student and a mentor at META FAIR. Our research was on agent theory-of-mind capabilities across different model architectures. At the time, state-space models, like mamba showed promising results, especially with compute efficiency; I was particularly interested in whether the way models internally manage context and information would change how well they reason about other minds, especially on more complex tasks. We had some interesting results in that the state-space model's performance on complex theory-of-mind tasks degraded less on long-context tasks than its transformer-based counterpart. This work was subsequently presented at the NAACL 2025 Student Research Workshop.

  2. 2025–26

    After that, I wanted to develop a more rigorous technical base, and also enjoy my senior year of high school. In this time, I started getting more into AI Safety, reading a multitude of technical papers and blogs about the more transformative aspects of AI. I also took a data science course through UConn, served as a TA for 2 UConn courses on economics, completed the ARENA curriculum, and worked on some other projects (prince string lists).

  3. summer 2026

    Somewhere in northern Nevada, around 2 a.m. on a cross-country road trip, I was scrolling in the passenger seat when I found a paper that changed how I think about measuring persuasion. Over the months before this, I had become much more invested in how people use AI, and the extent to which AI use affects users. As such, I looked into it more and more, and as the summer progressed, I turned my interests into reality by testing some of my ideas regarding agent-to-agent persuasion. Particularly, I looked at (1) whether the social framing of situations (who the agent is, who the agent is talking to, under what circumstances) affected persuasion, and (2) whether agents could reliably target their persuasion towards an entity that isn't the primary adversary (akin to how candidates in a presidential debate might technically be debating each other, but are more accurately trying to persuade the audience). This was an especially salient topic as ideas regarding emergent norms and agent behaviors in multi-agent systems grew with the OpenAI-HuggingFace breach. While this research did not lead to any significant results, I still found it very enriching experience wise, and hopefully inspiring for future research not just in multi-agent persuasion, but also in human-AI interactions as AI systems' capacities to persuade humans are incredibly important in both issues of alignment and safety. This paper would have been presented at NeurIPS '26 UserSim, but due to undergraduate funding constraints through UW, it will unfortunately not be presented. It is still a working paper, and in the coming months I will be working with a mentor through the NAACL 27 SRW to polish and submit this work archivally.

  4. summer 2026–

    Over the summer and to the present, I have also worked on a few projects at MatrAIx, an MIT CSAIL affiliate lab with multiple Nobel laureate advisors. This work has been on agent persona faithfulness (how well an agent follows a given persona [ex: if you tell an agent to act as a vegan, does it order steak or a salad at a restaurant]), and more recently on evaluating personal and finance agent failures.

  5. now

    My most recent project is working under Dante Lok of votee.ai and PolyU on identifying and preventing internal LLM reasoning failures. I am also in the Mercial Research Fellowship, working on systems for handling AI-generated research.

  6. community

    I have been a contributor to Terminal-Bench, a benchmark designed to measure agent work with diverse, difficult, and evolving tasks, and was an invited reviewer at the NeurIPS AI and the Self and AI4Good workshops. I am also currently helping organize the AI Safety movement at UW.

  7. coursework

    I have taken some coursework related to my research including two 30 hour courses on AGI Strategy and Technical AI Safety, year-long courses on Quantum Computing and AI, a course on Computer Ethics, and an online course on Frontier Systems; this year I am taking Human-AI Interaction, AI Ethics, Game Theory for CS, and a course on Social RL.

## ideas

Through my personal readings and research, I have also developed some ideas for future research directions that I am interested in. Below is a (non-exhaustive) list of ideas and topics that I think are interesting:

  • Agent societies: emergent norms, capabilities, and communication; collusion, scheming, and altruism; multi-agent coordination, learning, and games
  • Agent economics: alignment through trade; Coasean bargaining at scale; multi-agent marketplaces; principal-agent conflicts; information asymmetries; the economic value of AI work
  • Reasoning and decision-making: agent decision theory; game theory for agents; dynamic theory of mind; confidence and internal vs. external beliefs; hypersensitivity to prompts
  • Persuasion: alignment via debate; persuasion under pressure; situational awareness and sandbagging in evals; usersim and value/applications of usersim
  • Human-AI use: how and why people use AI; long-horizon autonomy vs. human-coupled systems; effects on culture, critical thinking, and decision-making; homogenization of language; ethics of AI detection software
  • User-facing LLMs: different approaches to building them, like humans&'s work on models for coordination and collaboration between people rather than automation
  • Training and alignment: effects of post-training on behavior; moving past RL; value drift under continual learning; pluralistic and bi-directional alignment; continual learning; personal AI models; test-time training; continual alignment
  • Architectures and personalization: state-space and other non-transformer approaches; what compaction and compression lose; personal user models
  • Other: post-AGI human value, labor, and rights; RSI; AI self-replication; AI sentience