Mukund Choudhary

📜 PhD in NLP @ MBZUAI, advised by @monojitchou

🔬 figuring out what LM linguistics looks like

💬 cover image from Rare Flavors

[Mukund Choudhary @mukundc2k]

@NeurIPSConf @monojitchou DM / mail us for any further questions.

honored to be organising this with Abdoul Jalil Djiberou Mahamadou, @SandrineHansen, @monojitchou, @aroraakhilcs, Dr. Haneesha Pinnamaraju, and Jacki O'Neill 😇

[Mukund Choudhary @mukundc2k]

a cool new @NeurIPSConf workshop just dropped! check out the details on the poster below and submit your thoughts/research on the concept of Self in this AI pervasive time. submission deadline on 24th of this month, workshop in Paris, December 12th/13th!🧘🤖

[NeurIPS Conference @NeurIPSConf]

We are happy to announce the list of accepted NeurIPS 2026 workshops: https://t.co/qiBOZgP6Ao

The workshops will take place on:
- Fri Dec 11 and Sat Dec 12, 2026 – Sydney
- Sat Dec 12 and Sun Dec 13, 2026 – Paris and Atlanta

We want to thank everyone who submitted a proposal and look forward to hosting the accepted workshops!

[Mukund Choudhary @mukundc2k]

DM / mail us for any further questions.

honored to be organising this with Abdoul Jalil Djiberou Mahamadou, @SandrineHansen, @monojitchou, @aroraakhilcs, Dr. Haneesha Pinnamaraju, and Jacki O'Neill 😇

[Mukund Choudhary @mukundc2k]

@AbhishekDedhe @rbanda86 @monojitchou @cogsci_soc #CogSci2026

[Mukund Choudhary @mukundc2k]

hello CogSci 2026 people!

We (@AbhishekDedhe @rbanda86 @monojitchou ) are excited to virtually discuss and present a sneak peek at our ongoing work comparing 🐦‍⬛, 🖥️, 🐒, 🧒, and 🧓 (very casual about the sneak there). If this piques your curiosity, stop by!

[Abhishek Dedhe | अभिषेक देढे @AbhishekDedhe]

Had a great time presenting at #Analogy2024 and #CogSci2024 about the algorithms underlying hierarchical patterning!

Thanks to @ClaireS012345, @margipavlova, and @IchienNick for organizing!

[Mukund Choudhary @mukundc2k]

@AbhishekDedhe @rbanda86 @monojitchou (imported from @rbanda86 's bluseky thread)

[Mukund Choudhary @mukundc2k]

@zouharvi cannot get over this one ever XD

[Mukund Choudhary @mukundc2k]

@zouharvi 🧑‍🦳🍰

[Vilém Zouhar @zouharvi]

There are many things machine translation still can't do. Help us steer the next direction by contributing hard-to-translate inputs (and be on a cool paper).

[Cohere Labs @Cohere_Labs]

We're excited to be sponsoring and co-organizing IOL-AI 2026, a new open challenge with the International Linguistics Olympiad @IOLing_official.🌎💬

🌐 Open-science
🗓️ One-month competition
🎯 Targeting AI reasoning with challenging, unseen test problems from the 2026 olympiad.

[Tom McCoy @RTomMcCoy]

🤖🧠 New commentary 🧠🤖

What role should large language models (LLMs) play in linguistics?

I reflect on this question in a piece now on arXiv: https://t.co/S4W7MASiqV

To appear in BBS as a commentary on @rljfutrell and @kmahowald's excellent piece on LLMs & Linguistics!

[Mukund Choudhary @mukundc2k]

@kanishkamisra they exaggerated counting back then?

[Mukund Choudhary @mukundc2k]

but I will stop here, for a more detailed discussion and implications of this kind of study, or any other feedback/comments, etc., drop by the poster or DM me! (apart from https://t.co/3EQMXn09NX you can find the paper and the poster from the QR code above.)

[Mukund Choudhary @mukundc2k]

On the other extreme, it turns out that LLMs don't 'naturally pick up' norms in linguistic variation, like native speakers in a high-resource-language, either. Drop by my poster for more, at EACL'26 poster sessions, where I continue presenting my findings about LLM linguistics 🧵

[Mukund Choudhary @mukundc2k]

LLMs fail at linguistic puzzles that high schoolers find easy. I am excited to share our #COLM2025 work that unveils the why, via linguistics. https://t.co/5GNL22MXqn🧩

Using uncontaminated (endangered languages) + self-contained, metalinguistic reasoning puzzles, we find: (1/8)

[Mukund Choudhary @mukundc2k]

4. we find that just classifying verbs (through LLM perplexity ratios) into the language preferences native speakers agreed on and where they didn't, was in itself a very hard task (lesser than baseline performance, across model families, sizes, and pre-training mixes)