HCI PhD @ UMich - using technology to make education more accessible, how people's background and cognitive science affects their interaction with techonology.
Joined August 2014
- Tweets219
- Following481
- Followers284
- Likes2K
Pinned Tweet
📚How do we make unfamiliar text easy for adult readers? Excited to present my work at #EMNLP2024 Main on "Evaluating LLMs for Targeted Concept Simplification for Domain-Specific Texts" with amazing @hjrashkin @eaclark07 @fantinehuot @mlapata - arxiv.org/abs/2410.20763v1
Sumit Asthana retweeted
New paper on human-AI interaction. We hire 18 writers to edit quirks in AI writing & see if #AI can mimic this process to improve its own writing
Verdict: Writer-edited > AI-edited > AI-generated
In other words:🚨Edits enhance alignment in writing🚨
🔗arxiv.org/pdf/2409.14509
Sumit Asthana retweeted
😓 #AI co-pilots excel in information work, but what about emotional ones? Working with angry clients harms #MentalHealth
🫂 Our #CHI2025 paper w/ @MSFTResearch introduces Care-Pilot, an AI-agent to empathetically support workers. #FutureOfWork
bit.ly/carepilot2 [1/7]
Sumit Asthana retweeted
New preprint out!
GenAI is changing how we think 🧠
This work reviews its cognitive impacts.
Using Bloom’s Taxonomy & Dewey’s theory of reflective thinking, it unpacks *how* it affects human cognition.
arxiv.org/abs/2502.12447
With @krntneja @KlaraGuan @evijit 🧵
Sumit Asthana retweeted
First CfP for #EMNLP2025 is live now. Submission deadline to May ARR cycle!
Excited to be part of organizing the conference as publicity chair w/ @amuuueller @dallascard so watch out for more updates esp by following the official conf account @emnlpmeeting
The EMNLP 2025 conference website and CfP are now live! 2025.emnlp.org/calls/main-co…
Conference dates: November 5-9 in Suzhou, China
Submissions will be through ARR, and this year's theme is Interdisciplinary Recontextualization of NLP
Sumit Asthana retweeted
🤗Thanks to those asking about our CHI session on Bidirectional 👫Human-AI🤖 Alignment!
🔥Exciting news — We’ll host both the 2025 #ICLR workshop & #CHI Special Interest Group (SIG), bringing together brilliant minds from #HCI & #AI experts to collectively explore how humans & AI/LLMs should be aligned.🚀
📢 Submit via OpenReview & you may opt to present at #ICLR or #CHI! The deadline is approaching — can't wait to see your amazing work! ✨
@iclr_conf @acm_chi @uw_ischool @michigan_AI #HumanAl #Alignment
🚀 Are you passionate about #Alignment Research? Exciting news! Join us at the ICLR 2025 Workshop on 👫<>🤖Bidirectional Human-AI Alignment (April 27 or 28, Singapore). We're inviting researchers in AI, HCI, NLP, Speech, Vision, Social Science, and beyond domains to submit their work on alignment-related topics!
📄 Paper Formats: 2-page (tiny), 4-page (short), & 9-page (long)
🗓️ Key Dates: Tentative Submission Deadline: Feb 3, 2025 (AoE). Workshop Day: April 27 or 28, 2025
💠Topics:
🔸Position papers and roadmaps for alignment research
🔸Specifying human values, cognition, and societal norms
🔸RLHF and advanced algorithms, interactive/customizable alignment methods
🔸AI safety, interpretability, scalable oversight, and steerability
🔸Red teaming, benchmarks, metrics, and human evaluation for alignment
🔸Dynamic human-AI co-evolvement and societal impact
🌟Speakers: @_beenkim, @bradamyers, @fraukolos, @HungyiLee2, @RichardMCNgo, @danbohus, @Pavel_Izmailov
👫Organizers: @huashen218, @ziqiao_ma, @reshmigh, @tknearem, @lxieyang, @tongshuangwu, @andresmh, @Diyi_Yang, @ABosselut, @furongh, @tanmit, @sled_ai, Marti A. Hearst, @dawnsongtweets, @yangli169
✍️ Interested in joining our Program Committee to help review? Apply here! forms.gle/MapXyCZhbcFr9F246
Join us in advancing the Human-AI Alignment research together🌟!
#ICLR2025 #AIAlignment #HCI #NLP #Speech #ComputerVision
Sumit Asthana retweeted
New paper on human-AI interaction. We hire 18 writers to edit quirks in AI writing & see if #AI can mimic this process to improve its own writing
Verdict: Writer-edited > AI-edited > AI-generated
In other words:🚨Edits enhance alignment in writing🚨
🔗arxiv.org/pdf/2409.14509
Sumit Asthana retweeted
New @allen_ai paper accepted to #NAACL2025! We worked with @TeachingLabHQ and @assistments to create a new dataset for evaluating VLMs. DrawEduMath includes 2,030 teacher-annotated images of students’ handwritten responses to math problems.
Website: drawedumath.org/
ALT a screenshot of the drawedumath website, showing paper title Evaluating Vision Language Models with Expert-Annotated Students’ Hand-Drawn Math Images and author list Sami Baral, Li Lucy, Ryan Knight, Alice Ng, Luca Soldaini, Neil Heffernan, Kyle Lo with affiliations at WPI, Berkeley, Insource Services, Teaching Lab, and Ai2. Below that is an image showing an example image with question-answer pairs from the dataset.
Sumit Asthana retweeted
How well can VLMs detect and explain humans' procedural mistakes, like in cooking or assembly?
🧑🍳🧑🔧
My new pre-print with Itamar Bar-Yossef, @YayuanLi, @zheyuanzhang99, @_JasonCorso_, and Joyce Chai (@SLED_AI @michigan_AI @UMichCSE) dives into this!
arxiv.org/pdf/2412.11927
ALT Dialog between a foundational VLM and itself to detect the incomplete state of the procedure "Unclip the pegs on the cloth" in an image showing a cloth pegged to a clothing line. The VLM generates the following questions and answers: 1. "Is there a cloth in the image? Yes", 2. "Are there pegs on the cloth? Yes", and 3. "Is there someone holding pegs? No". As the VLM asks these questions it becomes more confident that the procedure has not been successfully completed.
Sumit Asthana retweeted
📢I’m recruiting Ph.D students this cycle @UMichCSE , to invent and study mechanisms to support human expertise development, projects incl authoring techniques for tutoring systems, LLM-powered learning companions, simulated students, physical task tutoring, and more!
Sumit Asthana retweeted
Honored and humbled to awarded the 2024 Google PhD Fellow in #NLProc. Thankful to my advisor @kentarotoyama, my mentors & collaborators Sunayana Sitaram and Jiyin Cao and friends @umsi for their support. Thank you @GoogleAI for the honor.
Sumit Asthana retweeted
#EMNLP2024 make sure to come to Ashe Auditorium at 5pm today to learn more about 😺 CaT-Bench! Great way to end the conference day and head to the social event happy 😁
🚨 New paper alert! 🚨
📢 Announcing 😺 CaT-Bench! Our new benchmark exposes a serious deficit in LLM reasoning ability about real-world plans. CaT-Bench tests how well LLMs reason about Causal and Temporal aspects of cooking recipes. arxiv.org/abs/2406.15823 #NLProc 1/5
Sumit Asthana retweeted
Can laws be translated into software? In a new study,
@nelescher @jeffbilik @nikola_banovic and I show that the translation process is rife with misunderstandings, leading to errors and policy distortions in legal software.
dl.acm.org/doi/10.1145/36869…
Sumit Asthana retweeted
Excited to share ✨ Contextualized Evaluations ✨!
Benchmarks like Chatbot Arena contain underspecified queries, which can lead to arbitrary eval judgments. What happens if we provide evaluators with context (e.g who's the user, what's their intent) when judging LM outputs? 🧵↓
Stop by the Generation and Summarization track at 2:00 pm today to learn more about how to make it easy for adult readers to read domain-specific text containing unfamiliar concepts from our talk!
Sumit Asthana retweeted
I will be at @emnlpmeeting in Miami presenting our Main conference paper on how Human & AI perform on cognitively demanding creative thinking tasks (Nov 13th 4 PM)
I am also recruiting PhD students at @sbucompsc starting Fall 2025. Details on my webpage. Say hi #NLProc
New paper with students @BarnardCollege on testing orthogonal thinking / abstract reasoning capabilities of Large Language Models using the fascinating yet frustratingly difficult @nytimes Connections game. #NLProc #LLMs #GPT4o #Claude3opus 🧵(1/n)
LLMs struggle to balance contextually explaining difficult concepts, preserving the ideas in the definitions, and identifying the right level of explanation for readers. Automated metrics also show low correlations with human-evaluation simplification understanding.