Twitxy Saved

Juho Kim

@imjuhokim

Interaction-Centric AI, HCI, HAI researcher. Running @kixlab_kaist & member of @hcikaist. Associate Professor at @kaistcsdept. @mit, @Stanford, @SNUnow alum.
Daejeon, south Korea juhokim.com Joined March 2007
2.1K Following 6.2K Followers 2.9K Posts

Latest posts 5

Juho Kim @imjuhokim
Our paper on LLM user simulation at #EMNLP2026 Findings.

When evaluating agents w/ simulated users, they have to learn, get confused, and give up the way real people do. We investigate this bias.

w/ Daeheon Jeong & @yoonjoo_le2

arxiv.org/abs/2609.00608
assistantbias.kixlab.org
arxiv.org Investigating Assistant Bias in LLM User Simulators Using a Role Vector LLM-based user simulators are increasingly used to evaluate autonomous agents at scale, in place of costly human evaluations. Despite this promise, these simulators exhibit "assistant bias," a...
Yoonjoo Lee Verified account @yoonjoo_le2 Oct 25
Heading to #EMNLP2026! Excited to share two papers on LLM user simulation:

馃 KnowSim: Simulating how users learn through AI interactions, validated against real human鈥揂I interaction data across different knowledge levels.

馃幁 User Role Vector: Investigating assistant bias in LLM user simulators through internal representations and activation steering.

I鈥檓 also on the job market! I鈥檇 love to connect with folks interested in user simulation, interactive AI evaluation, and oversight for agentic systems. Feel free to DM me!
View quoted post on X
5 replies 2 reposts 26 likes 1.8K views View on X
Juho Kim @imjuhokim
Thanks @Steve_Yegge! This means a lot.

Token efficiency is one concrete place where the problem shows up, but the bigger question is: how do we understand, develop, and support the human skills emerging around AI-native work?

That鈥檚 the product we鈥檙e building. Stay tuned!
Steve Yegge Verified account @Steve_Yegge Jul 2
SkillBench is one of the most crazily important startups I know about, and it's been tough not to talk about them. Congrats to @mattbeane on this huge move!

SkillBench is poised to solve a tremendous number of problems in the industry, not least of which could be token efficiency. SkillBench is really one of the most useful things I've ever seen come out of the AI era.

In short, and this is butchering it, they scan your coding agent session traces and build a skills profile from it. And you get a character sheet, of sorts. Life as RPG. Are you inefficient with tokens? They'll help you see how, and even connect you to mentors.

They're actually much more significant than I've hinted at here, but I figured I'd start with token efficiency since it's arguably the industry's biggest problem today. Everyone is hacking on token efficiency metrics with heuristics that try to tie outcomes to dev inputs. Only SkillBench treats the problem as a fine-grained skills discovery issue, and I think that means they're the only ones who are doing it right.

I'll be talking about them a lot more. I think they are huge piece of the puzzle for "How do organizations work in the future, in a way that helps humans flourish?"
View quoted post on X
0 replies 0 reposts 6 likes 1.4K views View on X
Juho Kim @imjuhokim
Honored to be appointed to the @UN Independent International Scientific Panel on AI鈥攖he first global scientific body dedicated to assessing AI's real-world impact. 40 experts selected from 2,600+ applicants. Humbled to be the only member from Korea.
5 replies 7 reposts 57 likes 3.4K views View on X
Juho Kim @imjuhokim
Check out how we generate LLM personas with an intermediate layer of dimensions/values for better interpretability and trust #chi2025
Exciting news 馃
I'll present our upcoming paper Proxona: Supporting Creators' Sensemaking and Ideation with LLM-Powered Audience Personas at #CHI2025 in Yokohama! 馃嚡馃嚨

Can we simulate virtual yet plausible and meaningful audience personas for content creators?
View quoted post on X