Youze "Hargen" Zheng

Incoming CS PhD student @ucsd_cse; prev @HDSIUCSD @UCSDMathDept

[Youze "Hargen" Zheng @hgznnn___]

9/9 CT Open asks a broader AI question:

Can models forecast when the answer is not yet public -- rather than already sitting in a benchmark or on the web?

Fall Open 2026 is live. Any method is welcome.
🌐 https://t.co/38xNTVWnOS

See you at @COLM_conf in SF! 🌉

[Youze "Hargen" Zheng @hgznnn___]

Thrilled to share 𝗖𝗧 𝗢𝗽𝗲𝗻, accepted to @COLM_conf 2026! 🎉🎉🎉

𝗖𝗮𝗻 𝗔𝗜 𝗽𝗿𝗲𝗱𝗶𝗰𝘁 𝗼𝘂𝘁𝗰𝗼𝗺𝗲𝘀 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 𝗮𝗿𝗲 𝗽𝘂𝗯𝗹𝗶𝗰 -- 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗲 𝘄𝗵𝗮𝘁 𝗶𝘀 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗼𝗻𝗹𝗶𝗻𝗲?

We built a live, open-access, contamination-resistant benchmark + platform to find out.

🧵 1/9
📄 https://t.co/FHQmtxTM3L
🌐 https://t.co/W3ebZIrW3m

[Youze "Hargen" Zheng @hgznnn___]

8/9 Early results expose a real capability gap:

Simple LM and neural baselines remain competitive with -- and sometimes beat -- strong LLMs. RAG helps inconsistently. Web agents find useful evidence, but are costly and double-edged.

True forecasting remains far from solved.

[Youze "Hargen" Zheng @hgznnn___]

@yuqirose Congratulations!!

[Youze "Hargen" Zheng @hgznnn___]

@kaiwei_chang Congratulations!!!

[Youze "Hargen" Zheng @hgznnn___]

@tparekh97 @kaiwei_chang @VioletNPeng @WeiWang1973 @adityagrover_ Congrats Tanmay!!!

[Youze "Hargen" Zheng @hgznnn___]

@LiaoyaqiW @UTCompSci @EliasEskin @anqi_liu33 @ben_vandurme Congrats!🎉

[Youze "Hargen" Zheng @hgznnn___]

@AnneLe222 @berkeley_ai Congrats Heekyung!

[Youze "Hargen" Zheng @hgznnn___]

@haider1 Our paper, at COLM 2025 as Oral Spotlight, pointed directly to this direction, where we used a light-weight state-space model to retrieve important information for a downstream LLM to answer a question about extremely long documents!
Check out our paper: https://t.co/QmJtS1DO1y.