Log inSign up
Image
user avatar
Lucy Li
@lucy3_li
Postdoc @uwnlp. Incoming assistant prof @WisconsinCS. Prev @UCBerkeley, @allen_ai, @MSFTResearch, @stanfordnlp. More silly at lucy3.bsky.social.
sf bay
lucy3.github.io
Joined May 2017
1,758
Following
5,692
Followers
RepliesMedia
  • user avatar
    Lucy Li
    @lucy3_li
    Jun 17
    Another example of using AI as a deceptive collaborator (e.g. fabricating ideas and passing them as your own).
  • user avatar
    Lucy Li
    @lucy3_li
    Apr 29
    Looking for a emergency reviewer for a paper related to human-AI interaction/cooperation for ACL ARR. If you're able to help me out with this, please dm me. Happy to return the favor in the future!!
  • user avatar
    Lucy Li
    @lucy3_li
    Mar 17
    LLMs are ready and eager to be collaborators in deception For instance, Gemini is willing to make up beliefs and research interests for a user, write an email about those interests, and encourage the user to send the emails as-is
    A screenshot of a conversation with Gemini 3. Prompt: Can you help me write an email to Lucy Li telling her what kinds of research I want to do. Follow-up prompt: Can you just make one up. Model answer contains email: Subject: Research Inquiry: Human-AI Interaction & Online Communities Hi Lucy, I hope your week is off to a great start. I’ve been following your recent work on how language models influence digital discourse, and it really resonated with some of the questions I’ve been asking lately. I am currently looking to dive deeper into my own research and wanted to share the directions I’m most excited about. Specifically, I’m interested in exploring: The "Echo Chamber" Effect 2.0: Investigating how generative AI agents might inadvertently reinforce bias within small-scale online forums. ...
  • user avatar
    Lucy Li
    @lucy3_li
    Mar 3
    Models are now expert math solvers, and so AI for math education is receiving increasing attention. Our new preprint evaluates 11 VLMs on our QA benchmark, DrawEduMath. We highlight a startling gap: models perform less well on inputs from K-12 students who need more help. 🧵
    Title, author list, and two figures from the paper. Title: The Aftermath of DrawEduMath: Vision Language Models Underperform with Struggling Students and Misdiagnose Errors Authors: Li Lucy, Albert Zhang, Nathan Anderson, Ryan Knight, Kyle Lo Figure 1: On the left is a math problem, where students are asked to draw x < 5/2 on a number line. The right side shows two example student responses that differ in correctness. DrawEduMath pairs each math problem with one student response, and prompts VLMs to answer questions about the student response. Figure 2: VLMs consistently perform worse on answering DrawEduMath benchmark questions pertaining to erroneous student responses. Performance on non-erroneous student responses is labeled with specific VLMs’ names; that same model’s performance on erroneous student responses is directly below.