Stella Li
@stellali.bsky.social
PhD student @uwnlp.bsky.social @uwcse.bsky.social | visiting researcher @MetaAI | previously @jhuclsp.bsky.social
https://stellalisy.com
https://stellalisy.com
Pinned
Stella Li
@stellali.bsky.social
· Jul 22
WHY do you prefer something over another?
Reward models treat preference as a black-box😶🌫️but human brains🧠decompose decisions into hidden attributes
We built the first system to mirror how people really make decisions in our recent COLM paper🎨PrefPalette✨
Why it matters👉🏻🧵
Reward models treat preference as a black-box😶🌫️but human brains🧠decompose decisions into hidden attributes
We built the first system to mirror how people really make decisions in our recent COLM paper🎨PrefPalette✨
Why it matters👉🏻🧵
🤔💭What even is reasoning? It's time to answer the hard questions!
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
November 25, 2025 at 6:26 PM
🤔💭What even is reasoning? It's time to answer the hard questions!
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
We built the first unified taxonomy of 28 cognitive elements underlying reasoning
Spoiler—LLMs commonly employ sequential reasoning, rarely self-awareness, and often fail to use correct reasoning structures🧠
Reposted by Stella Li
Because Olmo 3 is fully open, we decontaminate our evals from our pretraining and midtraining data. @stellali.bsky.social proves this with spurious rewards: RL trained on a random reward signal can't improve on the evals, unlike some previous setups
November 20, 2025 at 8:38 PM
Because Olmo 3 is fully open, we decontaminate our evals from our pretraining and midtraining data. @stellali.bsky.social proves this with spurious rewards: RL trained on a random reward signal can't improve on the evals, unlike some previous setups
Reposted by Stella Li
Day 1 (Tue Oct 7) 4:30-6:30pm, Poster Session 2
Poster #77: ALFA: Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning; led by
@stellali.bsky.social & @jiminmun.bsky.social
Poster #77: ALFA: Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning; led by
@stellali.bsky.social & @jiminmun.bsky.social
October 6, 2025 at 2:51 PM
Day 1 (Tue Oct 7) 4:30-6:30pm, Poster Session 2
Poster #77: ALFA: Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning; led by
@stellali.bsky.social & @jiminmun.bsky.social
Poster #77: ALFA: Aligning LLMs to Ask Good Questions: A Case Study in Clinical Reasoning; led by
@stellali.bsky.social & @jiminmun.bsky.social
WHY do you prefer something over another?
Reward models treat preference as a black-box😶🌫️but human brains🧠decompose decisions into hidden attributes
We built the first system to mirror how people really make decisions in our recent COLM paper🎨PrefPalette✨
Why it matters👉🏻🧵
Reward models treat preference as a black-box😶🌫️but human brains🧠decompose decisions into hidden attributes
We built the first system to mirror how people really make decisions in our recent COLM paper🎨PrefPalette✨
Why it matters👉🏻🧵
July 22, 2025 at 2:59 PM
WHY do you prefer something over another?
Reward models treat preference as a black-box😶🌫️but human brains🧠decompose decisions into hidden attributes
We built the first system to mirror how people really make decisions in our recent COLM paper🎨PrefPalette✨
Why it matters👉🏻🧵
Reward models treat preference as a black-box😶🌫️but human brains🧠decompose decisions into hidden attributes
We built the first system to mirror how people really make decisions in our recent COLM paper🎨PrefPalette✨
Why it matters👉🏻🧵
Reposted by Stella Li
Want to quickly sample high-quality images from diffusion models, but can’t afford the time or compute to distill them? Introducing S4S, or Solving for the Solver, which learns the coefficients and discretization steps for a DM solver to improve few-NFE generation.
Thread 👇 1/
Thread 👇 1/
February 27, 2025 at 6:24 PM
Want to quickly sample high-quality images from diffusion models, but can’t afford the time or compute to distill them? Introducing S4S, or Solving for the Solver, which learns the coefficients and discretization steps for a DM solver to improve few-NFE generation.
Thread 👇 1/
Thread 👇 1/
Asking the right questions can make or break decisions in fields like medicine, law, and beyond✴️
Our new framework ALFA—ALignment with Fine-grained Attributes—teaches LLMs to PROACTIVE seek information through better questions through **structured rewards**🏥❓
(co-led with @jiminmun.bsky.social)
👉🏻🧵
Our new framework ALFA—ALignment with Fine-grained Attributes—teaches LLMs to PROACTIVE seek information through better questions through **structured rewards**🏥❓
(co-led with @jiminmun.bsky.social)
👉🏻🧵
February 21, 2025 at 4:00 PM
Asking the right questions can make or break decisions in fields like medicine, law, and beyond✴️
Our new framework ALFA—ALignment with Fine-grained Attributes—teaches LLMs to PROACTIVE seek information through better questions through **structured rewards**🏥❓
(co-led with @jiminmun.bsky.social)
👉🏻🧵
Our new framework ALFA—ALignment with Fine-grained Attributes—teaches LLMs to PROACTIVE seek information through better questions through **structured rewards**🏥❓
(co-led with @jiminmun.bsky.social)
👉🏻🧵
Reposted by Stella Li
New preprint!
Metaphors shape how people understand politics, but measuring them (& their real-world effects) is hard.
We develop a new method to measure metaphor & use it to study dehumanizing metaphor in 400K immigration tweets Link: bit.ly/4i3PGm3
#NLP #NLProc #polisky #polcom #compsocialsci
🐦🐦
Metaphors shape how people understand politics, but measuring them (& their real-world effects) is hard.
We develop a new method to measure metaphor & use it to study dehumanizing metaphor in 400K immigration tweets Link: bit.ly/4i3PGm3
#NLP #NLProc #polisky #polcom #compsocialsci
🐦🐦
February 20, 2025 at 7:59 PM
New preprint!
Metaphors shape how people understand politics, but measuring them (& their real-world effects) is hard.
We develop a new method to measure metaphor & use it to study dehumanizing metaphor in 400K immigration tweets Link: bit.ly/4i3PGm3
#NLP #NLProc #polisky #polcom #compsocialsci
🐦🐦
Metaphors shape how people understand politics, but measuring them (& their real-world effects) is hard.
We develop a new method to measure metaphor & use it to study dehumanizing metaphor in 400K immigration tweets Link: bit.ly/4i3PGm3
#NLP #NLProc #polisky #polcom #compsocialsci
🐦🐦
The feeling when you thought you were scooped from the title and the abstract but then unscooped after reading the actual paper is my best christmas present🤌💕🎁
December 20, 2024 at 7:19 PM
The feeling when you thought you were scooped from the title and the abstract but then unscooped after reading the actual paper is my best christmas present🤌💕🎁
Reposted by Stella Li
31% of US adults use generative AI for healthcare 🤯But most AI systems answer questions assertively—even when they don’t have the necessary context. Introducing #MediQ a framework that enables LLMs to recognize uncertainty🤔and ask the right questions❓when info is missing: 🧵
December 6, 2024 at 10:51 PM
31% of US adults use generative AI for healthcare 🤯But most AI systems answer questions assertively—even when they don’t have the necessary context. Introducing #MediQ a framework that enables LLMs to recognize uncertainty🤔and ask the right questions❓when info is missing: 🧵
study music:1, taylor swift:0 (only for September I swear😅) Spotify knows too much about me….
December 4, 2024 at 7:45 PM
study music:1, taylor swift:0 (only for September I swear😅) Spotify knows too much about me….
Reposted by Stella Li
We created an Allen School Starter Pack to help you find and connect with @uofwa.bsky.social #UWAllen labs and researchers on 🦋! We'll add to this list as we grow our community here: go.bsky.app/RyHBLJd
#AcademicBluesky #CompSci #AI #CompBio #UbiComp #Accessibility #NLP #HCI #DataViz #mHealth
#AcademicBluesky #CompSci #AI #CompBio #UbiComp #Accessibility #NLP #HCI #DataViz #mHealth
Allen School Starter Pack
Join the conversation
go.bsky.app
November 26, 2024 at 7:49 PM
We created an Allen School Starter Pack to help you find and connect with @uofwa.bsky.social #UWAllen labs and researchers on 🦋! We'll add to this list as we grow our community here: go.bsky.app/RyHBLJd
#AcademicBluesky #CompSci #AI #CompBio #UbiComp #Accessibility #NLP #HCI #DataViz #mHealth
#AcademicBluesky #CompSci #AI #CompBio #UbiComp #Accessibility #NLP #HCI #DataViz #mHealth
OpenReview turned into Reddit🤯 Can we now add upvote/downvote buttons to reviews and rebuttals plz?? Would be a very rich and interesting source of preference data🤡
Looking at ICLR submissions with the lowest score - What a work of art! 🧵
November 25, 2024 at 11:05 PM
OpenReview turned into Reddit🤯 Can we now add upvote/downvote buttons to reviews and rebuttals plz?? Would be a very rich and interesting source of preference data🤡
Reposted by Stella Li
I noticed a lot of starter packs skewed towards faculty/industry, so I made one of just NLP & ML students: go.bsky.app/vju2ux
Students do different research, go on the job market, and recruit other students. Ping me and I'll add you!
Students do different research, go on the job market, and recruit other students. Ping me and I'll add you!
November 23, 2024 at 7:54 PM
I noticed a lot of starter packs skewed towards faculty/industry, so I made one of just NLP & ML students: go.bsky.app/vju2ux
Students do different research, go on the job market, and recruit other students. Ping me and I'll add you!
Students do different research, go on the job market, and recruit other students. Ping me and I'll add you!
Came across this paper, very cute of @hamishivi.bsky.social to say "popular BERT model"🤡
November 21, 2024 at 1:26 AM
Came across this paper, very cute of @hamishivi.bsky.social to say "popular BERT model"🤡
Reposted by Stella Li
I'm biased ofc but IMO the ~right theoretical framework for thinking and talking about instructable LLM-based agents is assistance games (a.k.a. Cooperative Inverse Reinforcement Learning (CIRL) games).
Setting aside the fact that people are literally re-inventing a bunch of terminology around agents when we have several decades of framing from RL (POMDPs, anyone?) that we can rely on to describe what we're doing, the main challenges are as follows., working backwards from the end point. [2/11]
November 19, 2024 at 2:07 PM
I'm biased ofc but IMO the ~right theoretical framework for thinking and talking about instructable LLM-based agents is assistance games (a.k.a. Cooperative Inverse Reinforcement Learning (CIRL) games).
Reposted by Stella Li
Ready for another Computational Social Science Starter Pack?
Here is number 2! More amazing folks to follow! Many students and the next gen represented!
go.bsky.app/GoEyD7d
Here is number 2! More amazing folks to follow! Many students and the next gen represented!
go.bsky.app/GoEyD7d
November 14, 2024 at 11:42 PM
Ready for another Computational Social Science Starter Pack?
Here is number 2! More amazing folks to follow! Many students and the next gen represented!
go.bsky.app/GoEyD7d
Here is number 2! More amazing folks to follow! Many students and the next gen represented!
go.bsky.app/GoEyD7d
Reposted by Stella Li
1/ Introducing ᴏᴘᴇɴꜱᴄʜᴏʟᴀʀ: a retrieval-augmented LM to help scientists synthesize knowledge 📚
@uwnlp.bsky.social & Ai2
With open models & 45M-paper datastores, it outperforms proprietary systems & match human experts.
Try out our demo!
openscholar.allen.ai
@uwnlp.bsky.social & Ai2
With open models & 45M-paper datastores, it outperforms proprietary systems & match human experts.
Try out our demo!
openscholar.allen.ai
November 19, 2024 at 4:30 PM
1/ Introducing ᴏᴘᴇɴꜱᴄʜᴏʟᴀʀ: a retrieval-augmented LM to help scientists synthesize knowledge 📚
@uwnlp.bsky.social & Ai2
With open models & 45M-paper datastores, it outperforms proprietary systems & match human experts.
Try out our demo!
openscholar.allen.ai
@uwnlp.bsky.social & Ai2
With open models & 45M-paper datastores, it outperforms proprietary systems & match human experts.
Try out our demo!
openscholar.allen.ai
I noticed that all the X posts about migrating to 🦋 (with the app name, the emoji, or even “the other platform”) get extremely down ranked😅 very sad/disturbed to see the level of content control over there🤯
November 19, 2024 at 3:25 AM
I noticed that all the X posts about migrating to 🦋 (with the app name, the emoji, or even “the other platform”) get extremely down ranked😅 very sad/disturbed to see the level of content control over there🤯
Had so much fun presenting our work ValueScope (arxiv.org/abs/2407.024...) at #EMNLP2024 and (mostly) enjoying Miami with friends last week🫶🏻🏖️!! Also learned about so many amazing works and now I’m inspired to start working again🫡💕
November 18, 2024 at 5:29 PM
Had so much fun presenting our work ValueScope (arxiv.org/abs/2407.024...) at #EMNLP2024 and (mostly) enjoying Miami with friends last week🫶🏻🏖️!! Also learned about so many amazing works and now I’m inspired to start working again🫡💕