Incredible work from Hita on a timeless challenging NLP+HCI problem: how do we help readers effectively navigate a complex answer with citation to effectively retrieve evidence. Check it out!
Commencing the annual defrost of my academic social media to share delightful projects done with delightful collaborators that I’m excited to present at #UIST2026, the first of which is: Attribution Gradients, which also got a 🏆 Best Paper Honorable Mention!
I'm at COLM this week, presenting:
- LLMs Corrupt Your Documents When You Delegate. At today (Wednesday's) afternoon poster 4.30-6.30. Come check out our very (!) colorful poster.
- Giving a fun talk on mostly ongoing work at the Workshop on Lifelong Agents early Friday morning.
Check out Yoonjoo's fantastic work on simulating users of different expertise levels?
It makes progress on this concrete question: for a given task, how do we simulate how a novice vs. an expert would approach it?
The validation study is a great example of careful research
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
Check out Jihoon's very interesting work on simulating long "situated" conversations where the user can change their mind, switch tasks, underspecify, etc.
We need more work evaluating model performance outside of the single-turn, single-task, lab-like setting.
Excited to share my first work after joining @MSFTResearch!
LLMs have entered the agentic era, and we now collaborate with agents on complex tasks over many turns of interaction.
But does your agent actually follow what you intended?
We show where agents get lost: LLMs Get Lost
New paper! LLMs Corrupt Your Documents When You Delegate
LLMs are enabling a new way of working: delegated work, where users supervise an LLM as it edits documents on their behalf.
Delegation requires trust: does the LLM complete tasks without introducing errors?
We simulate
