Muzhe Wu retweeted
Excited to share that our paper “Personalizing LLMs Through User Value Profiles” has been accepted to @EMNLP 2026 Main Track! #emnlp2026
Muzhe Wu retweeted
Heading to #EMNLP2026! Excited to share two papers on LLM user simulation:
🧠 KnowSim: Simulating how users learn through AI interactions, validated against real human–AI interaction data across different knowledge levels.
🎭 User Role Vector: Investigating assistant bias in LLM user simulators through internal representations and activation steering.
I’m also on the job market! I’d love to connect with folks interested in user simulation, interactive AI evaluation, and oversight for agentic systems. Feel free to DM me!
AI assistants can provide correct, comprehensive information and still be poorly calibrated to the user: too difficult for novices, redundant for experts.
KnowSim is a user simulator that tracks what a user knows and updates that knowledge as the conversation unfolds, so we can evaluate information calibration over multi-turn interactions.
Across 9 LLMs, model rankings changed with the user’s knowledge level.
📄 arxiv.org/abs/2608.17150
Muzhe Wu retweeted
AI assistants can provide correct, comprehensive information and still be poorly calibrated to the user: too difficult for novices, redundant for experts.
KnowSim is a user simulator that tracks what a user knows and updates that knowledge as the conversation unfolds, so we can evaluate information calibration over multi-turn interactions.
Across 9 LLMs, model rankings changed with the user’s knowledge level.
📄 arxiv.org/abs/2608.17150
Muzhe Wu retweeted
Announcing our new work: Toward Honest Language Models for Deductive Reasoning 🎉
1. We study honesty in settings where a model must derive a conclusion only when it logically follows and abstain when the query is unanswerable.
2. To support this, we curate two multi step deductive reasoning datasets built from graph structures, one for linear algebra and one for logical inference.
3. We find that prompting and existing training methods, including GRPO with or without supervised fine tuning, struggle to reason honestly on these tasks.
4. We introduce ANCHOR ⚓️, a reinforcement learning method that injects ground truth trajectories into rollouts to stabilize learning and prevent early collapse.
📄 Check out the paper: arxiv.org/abs/2511.09222
Muzhe Wu retweeted
New lab alert ‼️🚨
I’m looking for PhD students who are passionate about creating assistive technologies and human-centered AI tools to join my research group at Tufts CS! I'm starting the Tufts AI, Design, and Accessibility Lab (TADA!) in January 2026 🥳
(1/4)
ALT Graphic showing research areas for the TADA lab. In the center, a three circle Venn Diagram with the research areas Accessibility, AI, and Creation Tools. Surrounding the Venn Diagram are projects from Jaylin's research, pointing to the research areas. (1) Making Mobile AR Accessible, ASSETS 2020 (Accessibility). (2) CollabAlly: Collaborative Writing Accessibility, CHI 2022 (Accessibility and Creation Tools). (3) XSpace: AR Collaboration Toolkit, ISS 2022 (Creation Tools). (4) Hacking, Switching, Combining, CHI 2023 (Accessibility and AI). (5) ImageExplorer: Verifying Image Captions, CHI 2022 (Accessibility and AI). (6) ProgramAlly: Programming for DIY Assistive AI, UIST 2024 (Accessibility, AI, and Creation Tools). The side of the image lists possible future research directions: "(1) DIY Assistive Software: How to help people create personal assistive tools? (2) Customizing AI: How to customize AI effectively and on-the-fly? (3) Understanding AI Errors: How to use imperfect AI?"
Muzhe Wu retweeted
Ever wondered if more AI help actually makes us learn or think less?
I’ll be talking about this at #CSCW2025 — AI & Writing Session🗓️ Tue, Oct 21 | 11:00 AM
Paper: dl.acm.org/doi/pdf/10.1145/3…
Muzhe Wu retweeted
🚨 New on arXiv! 🚨
"Not There Yet: Evaluating Vision Language Models in Simulating the Visual Perception of People with Low Vision"
Can VLMs really simulate what people with low vision perceive?
Our findings: Not quite yet.
📄 arxiv.org/abs/2508.10972v1
Muzhe Wu retweeted
Hello! We are researchers from Carnegie Mellon University. We are looking for mental health professionals or end users who have sought mental health support from LLM-based products to participate in a survey and workshop study. (Details in flyer)
Sign-up: forms.gle/AHniGREb1PuXngoVA
Muzhe Wu retweeted
Introducing HandProxy, a speech-controlled virtual proxy hand that interprets the user’s natural speech input and performs hand interactions in XR on the user's behalf. Accepted in #IMWUT 2025 (1/n)
Paper: bit.ly/4fjdEJN
Demo Video: bit.ly/40KXtPy
Rubikon shows the promises of integrating cognitive tutoring principles with mixed reality interfaces. We aim to extend these concepts to broader physical task domains, enabling enhanced personalized learning.
#DIS2025 Paper: arxiv.org/abs/2503.12619
Video: youtu.be/4Jes_0nd3kY
Muzhe Wu retweeted
Meetings often drift off-track. Can AI help us reflect in the moment and realign?
Our #CHI2025 paper explores how AI-assisted interventions influence goal reflection during workplace meetings.
w/@lev_tank @rishivanukuru @PayodPanda @seanrintel @ava___scott
Muzhe Wu retweeted
A week before heading to Japan for #chi2025! I will be presenting a work focusing on immersive story viewpoint and navigation design (arxiv.org/abs/2502.04542). Happy to chat about any experiences on storytelling or VR/AR design!
Muzhe Wu retweeted
🌸New #CHI2025 paper alert!!
🤖👩🏫 How can teachers ensure that AI tutors work for every student?
Our team envisions a future where teachers build chatbots (e.g., AI tutors) tailored to their curriculum and students.
📄 Paper: arxiv.org/abs/2410.04078
