UW Tech Policy LAB: 2026 Spring Internships

Dear Students:

The 2026 Spring Internships are open for research projects in AI Safety, Personalized AI Systems, and Large Language Models (LLMs) at the UW Tech Policy LAB.

Our group focuses on building next-generation AI systems that are safe, personalized, and aligned with real-world human needs, especially in high-risk scenarios such as health, finance, and decision-making.

This is a great opportunity to get involved in cutting-edge AI research and contribute to projects that are actively being prepared for top conferences such as NeurIPS, ICML, and ACL. We provide structured guidance and training to help students participate in research and system building.

Strong students are given opportunities to publish their work, contribute to open-source projects, collaborate on research papers, and receive letters of recommendation for graduate school or career advancement. Many students may continue into long-term or paid research roles based on performance.

Students are expected to commit 10 to 15 hours per week. Hours are flexible and can include telecommuting, evenings, or weekends. Project availability changes depending on student interests and research needs.

Students will gain hands-on experience using AI and large language models (LLMs), from data analysis to building AI-powered systems. They will also develop AI fluency and build a personal portfolio showcasing the tools, systems, and research contributions they have developed. No prior AI research experience is required.

Highlighted projects are listed below:

Personalized AI Safety and Risk Modeling Study how large language models behave under user-specific contexts, especially in high-risk scenarios where background information is incomplete, missing, or misleading. This project focuses on identifying failure modes of one-size-fits-all responses, designing benchmarks for personalized safety, and developing evaluation protocols that capture risk sensitivity, emotional alignment, and user-specific correctness. Great for students interested in AI safety, evaluation, and social impact.

Memory and Information Acquisition in LLM Agents Develop agent frameworks that actively acquire and utilize user background information to improve response quality and safety. This includes studying when to ask for additional information, how to store and retrieve relevant context, and how memory mechanisms influence long-horizon reasoning. Current LLM agents rely on ad-hoc memory or multi-agent pipelines; this project explores principled and learnable memory strategies. Great for students interested in AI agents, reasoning, and decision-making.

Reasoning Trace Optimization and Control Investigate whether current LLM reasoning traces (e.g., chain-of-thought) are optimal for decision-making and safety. This project explores how reasoning processes can be modified, structured, or guided to improve consistency, correctness, and robustness—especially under uncertainty or personalized contexts. It also studies the relationship between reasoning traces and final model behavior. Great for students interested in reasoning, interpretability, and model control.

Multimodal Personalized Safety (Vision-Language Models) Analyze how visual and textual signals interact in multimodal models and how this affects personalized safety. This includes studying failure modes such as cross-modal interference, evaluating safety under conflicting signals, and designing multimodal benchmarks that incorporate user context. The goal is to understand and mitigate risks unique to vision-language systems. Great for students interested in multimodal AI, computer vision, and safety.

Training Personalized LLMs via Risk-Aware Interaction (RL-based) Develop training methods for personalized LLMs that proactively seek missing user information in high-risk scenarios. Instead of directly generating responses, the model learns to recognize uncertainty and ask clarifying questions when necessary. This project explores reinforcement learning or RLHF-based approaches to train models that balance helpfulness and safety by deciding when to answer and when to ask. It also studies how such behaviors affect downstream safety, user trust, and decision quality. Great for students interested in RL, alignment, and interactive AI systems.