iiinova — where your digital self keeps evolving

Reposted by Lechen Zhang

Indu Panigrahi @indupanigrahi.bsky.social · 07/08/2026

📌 Do you have experience or professional interest engaging with an unfamiliar discipline or a discipline beyond your primary discipline? Complete this 10-15 minute survey! 👉 illinois.qualtrics.com/jfe/form/SV_... Eligibility criteria and compensation details in 🧵 cc: @talaugust.bsky.social

302

Lechen Zhang @leczhang.bsky.social · 15/06/2026

10/ This is the first paper at UIUC, and I'm deeply grateful to my advisor @talaugust.bsky.social for his thoughtful guidance, constant support, and sharp insights at every step. Huge thanks also to my collaborator Jiarui Liu and everyone in the lab for their tremendous help! 🙏

000

Lechen Zhang @leczhang.bsky.social · 15/06/2026

9/ Takeaway: Synthetic users and LLM judges aren't enough to capture the complex nature of human preferences. We highlight the importance of human data and call for re-centering humans in LLM personalization Paper: arxiv.org/abs/2606.06614 Code & data: github.com/orange0629/recenter-personalization

github.com

orange0629/recenter-personalization

Contribute to orange0629/recenter-personalization development by creating an account on GitHub.

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

8/ We also show our data serves as a useful training source. But personalization is far more complex than any one-size-fits-all model: even humans vary meaningfully in what they consider good personalization, and each individual deserves to be modeled seriously.

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

7/ 👀One especially surprising thing: LLMs differ sharply on surface-level personalization (explicitly mentioning user attributes), and their preferences as generators and judges are highly correlated. E.g., Claude strongly avoids AND penalizes mentions. Humans? They barely care.

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

6/ When rating effectiveness of personalization (Stage 3): humans and LLMs disagree again. As generators, 54.6% of personalized responses are judged by humans as no better than generic ones. As judges, LLMs assign inflated scores and are sensitive to surface-level personalization

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

5/ When deciding which attributes to use (Stage 2): LLMs and humans disagree, but each agree well within their own group. LLMs mark 40–60% of attributes as relevant vs. ~20% for humans, revealing a systematic misalignment that drives over-personalization.

120

Lechen Zhang @leczhang.bsky.social · 15/06/2026

4/ When inferring user attributes from real conversations (Stage 1): LLMs struggle to extract reliably. Compared to synthetic benchmarks, +22% attributes extracted from real chats are judged problematic by humans. Synthetic data underestimate how hard reliable extraction is.

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

3/ We study personalization in 3 stages 🧠Attribute Extraction: what to infer from conversations? 🎯Relevance Matching: which attributes matter now? ✍️Response Generation: does personalization improve user experience? With human data, we find clear human-LLM gaps at every stage

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

2/ Our main research question: 𝗵𝗼𝘄 𝘄𝗲𝗹𝗹 𝗱𝗼𝗲𝘀 𝗟𝗟𝗠 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘄𝗼𝗿𝗸 𝗳𝗼𝗿 𝗿𝗲𝗮𝗹 𝘂𝘀𝗲𝗿𝘀? We collected 550 real conversations + ~19K human judgments to find out. 📄 Paper: arxiv.org/abs/2606.06614

100

Lechen Zhang @leczhang.bsky.social · 15/06/2026

1/ When chatbots remember something about you in “𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹𝗶𝘇𝗲𝗱” answers, does it ever feel uncomfortable, offensive, or just unnecessary? 🚨 Our paper argues these misalignments reflects a deeper issue in LLM personalization: 𝗼𝘃𝗲𝗿-𝗿𝗲𝗹𝗶𝗮𝗻𝗰𝗲 𝗼𝗻 𝘀𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝘂𝘀𝗲𝗿𝘀 𝗮𝗻𝗱 𝗟𝗟𝗠 𝗷𝘂𝗱𝗴𝗲𝘀. 🧵👇

110

Lechen Zhang @leczhang.bsky.social · 24/04/2026

🚨 𝙋𝙤𝙨𝙩𝙚𝙧 @ #ICLR2026 🇧🇷 🚨 Is your 𝘀𝘆𝘀𝘁𝗲𝗺 𝗽𝗿𝗼𝗺𝗽𝘁 actually optimal? 🚀 Catch me this Saturday (Apr 25) 3:15–5:45pm at Pavilion 3, Poster #1314 presenting 𝙎𝙋𝙍𝙄𝙂, an automatic genetic optimizer for system prompts! Come check it out and let's chat!

020

Lechen Zhang @leczhang.bsky.social · 07/04/2026

I am so sorry to see that PCs are going through this situation. I also felt deeply upset by the toxicity yesterday. It's really unfair to those who are still trying to be responsible and take their work seriously. Thank you so much for your time and dedication 🙏🏻

000

Reposted by Lechen Zhang

David Jurgens @davidjurgens.bsky.social · 29/01/2026

This major release of my lab's annotation tool, Potato 2.0, was a long time in the making but I'm really excited by all the new features (and quality of life improvements). If you're doing data annotation of any kind, check it out. Need any other features? We'd love to hear from you too!

1294

Reposted by Lechen Zhang

David Jurgens @davidjurgens.bsky.social · 25/11/2025

I've been really impressed with the presentation modalities from NotebookLM. My former student @leczhang.bsky.social used it to make this really cool illustration of our SPRIG paper arxiv.org/abs/2410.14826 and I was blown away by the layout and design. Here's the illustration of the genetic alg...

141

Reposted by Lechen Zhang

David Jurgens @davidjurgens.bsky.social · 01/03/2025

I'm grateful for the chance to write this and so impressed by all the amazing things scientists are doing in Central Illinois and the impact the NSF/NIH are having in my hometown. Many thanks to the folks at @sciencehomecoming.bsky.social for this initiave and all the resources to get started

23512