Lu Lin
Pennsylvania State University, PA·Computer Science
Large language models (LLMs), the AI systems behind chatbots, now support work in education, healthcare, media and national security. They can also be manipulated into producing harmful or misleading content.
- Fields
Computer Science
- Topics
large language models
LLM safety
AI safety
- Grants
- 1 known grant, through Jun 2027
What they work on
- Identify critical LLM safety concerns, gaps in current evaluation and infrastructure needs through structured surveys and interviews with experts across computing disciplines.
- Organize a workshop to advance strategies for evaluating large language model safety and refine the design of a shared evaluation infrastructure.
- Develop a prototype evaluation infrastructure that combines a data library, an LLM model zoo, an attack/defense repository and an evaluator hub, refined through iterative community feedback.
- Host an LLM safety challenge in which the community tests and improves the prototype evaluation infrastructure.
More about the research
Attacks and defenses keep evolving in an arms race, yet there is no standard way to measure how risky a given model truly is. Without shared benchmarks and tools, results from different groups are hard to compare.
Why it matters
AI chatbots are used in education, healthcare and security, but can be misused to produce harmful content. A shared, standardized way to evaluate LLM safety could help researchers, practitioners and policymakers judge real risks and compare attacks and defenses fairly. In turn, it could help make these widely deployed models safer.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Dongxiao Zhu
Wayne State University, MI·Computer Science
Develop optimization algorithms that make large language model unlearning, removing specific data influences without retraining, more effective, robust and efficient.
Desmond Ong
University of Texas at Austin, TX·Computer Science·Early-career
Define types of emotional reasoning, such as predicting emotions from situations and mental states and reasoning about interventions that change emotions, drawing on psychological theory.
Anthony Sicilia
West Virginia University, WV·Computer Science
Develop a Probably Approximately Correct (PAC) learning theory that measures how user disagreement, suggestions and topic shifts change the calibration of an AI system's confidence during a conversation.
The school
Pennsylvania State University: 60 labs on PI Pages, $1.3B a year on research, 743 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.