Tatsuya Aoyama
@t_aoyam
Followers
69
Following
125
Media
0
Statuses
10
Research Scientist @AIatMeta | Ph.D. @GUCompLing
Joined May 2023
My work on cognitive modeling x emergent structures happening at X5 166, and my collaboration with @xiulin_yang on multilingual impossible language learning at X5 168!!!
I'm in Vienna for #ACL2025, presenting ( https://t.co/Vi5uU74aP6) with @weGotlieb on 7/30 11-12:30 @ Hall X4/X5! If you ever wonder why an LM's "humanlikeness" ↗️↘️ during pretraining, come check it out! DM/email/say hi in person - happy to (re)connect with old/new friends!
0
0
3
I'm in Vienna for #ACL2025, presenting ( https://t.co/Vi5uU74aP6) with @weGotlieb on 7/30 11-12:30 @ Hall X4/X5! If you ever wonder why an LM's "humanlikeness" ↗️↘️ during pretraining, come check it out! DM/email/say hi in person - happy to (re)connect with old/new friends!
arxiv.org
LMs' alignment with human reading behavior (i.e. psychometric predictive power; PPP) is known to improve during pretraining up to a tipping point, beyond which it either plateaus or degrades....
0
1
11
🥳Thrilled to present our #ACL2025 paper with @t_aoyam, @yuekun_yao, and @weGotlieb on language modeling! Can LMs’ learning dynamics distinguish typologically attested from unattested and from truly impossible languages? We test this across 12 languages and find: yes, but…👀
2
3
12
Honored to see our work with @weGotlieb mentioned by @nsaphra - we've posted an updated version on arXiv. I will be presenting it at ACL in Vienna, so please come say hi! I will post a proper thread on this paper soon, so please stay tuned :)
After induction heads form and ICL emerges, models get progressively less human-like in their predictions, leading to divergence in larger models. You can't evaluate these models like humans. They are not remotely humanlike, far better at retrieval
2
0
9
Submissions for the 2025 Workshop on Cognitive Modeling and Computational Linguistics are due Feb. 16 I humbly request your help with spreading the word
📣 We are happy to share that CMCL 2025 will be co-located with NAACL in New Mexico! 👉The call for papers is out https://t.co/KLledcVd5q SAVE THE DATES, and submit your work! ‼️ Paper submission deadline: Feb 16, 2025 🗓 May 3 or 4, 2025: Workshop (TBA) @naacl @naaclmeeting
0
10
23
Hi everyone! We are conducting a survey on the usage of Georgetown's GUM 10 corpus. If you currently use or plan on using the GUM corpus for your research, please consider filling out our survey to give us your suggestions and feedback. Thank you!!
docs.google.com
Hello all! We (Corpling@GU) are running a survey about the current usage of the GUM corpus, and we want to know user preferences for the current and potential new genres included in the GUM corpus....
0
6
6