Ph.D. student @nyuniversity | B.S. @PKU1898
Joined April 2023
- Tweets25
- Following539
- Followers124
- Likes88
New preprint! Why do LLMs abandon a correct answer as soon as a user disagrees? We trace the mechanisms behind sycophantic agreement.
arxiv.org/abs/2609.35822
Zooming out: we identify a mechanism for sycophantic agreement, where a user’s opinion is registered early and then biases the model’s later answer retrieval. Pinpointing these components could help guide more targeted interventions to make models more truthful.
This work was done as part of the @cbai_ai fellowship. Joint work with @zfjoshying, @loganriggssmith, @JeremyWertheime, @NatalieShapira. Check out our preprint for more information.
Now accepted to #NeurIPS2026!
New preprint! When reasoning LLMs deliberate over possible futures, are they actually planning?
arxiv.org/abs/2605.06840
Sixing Chen retweeted
Applications are open for the CBAI Fall Research Fellowship in AI Safety & AIxBiosecurity. Apply by September 6th at 11:59 PM EDT!
📅October 13 - December 18
💵 $15,000 stipend for 10 weeks
💰 ~$1,500+/week compute support
🏠 Housing close to our offices, arranged by CBAI
New preprint! When reasoning LLMs deliberate over possible futures, are they actually planning?
arxiv.org/abs/2605.06840
Zooming out: We found that LLMs write out deep lookahead, but they don't act on it.
More broadly, our work provides a generalizable approach for interpreting LLM planning across strategic domains.
Joint work with @Ji_An_Li , @sanerp110, Sinan Akcali, @kaylalee278, and @marcelomattar. Check out our preprint for more information.
Sixing Chen retweeted
Excited to share our work now published in Nature Communications - shared functional specialization in tansformer language models and the human brain! With co-first author @tedsumers, senior author @samnastase and others (@hawkrobe, @TakaYamakoshi). [1/n]
nature.com/articles/s41467-0…
