Homepage - mukund choudhary

( ˆ𐃷ˆ) .ᐟ.ᐟ

hi! i am a 3rd year PhD student in NLP @ MBZUAI advised by Monojit Choudhury, comparing how LLMs structure language to how humans do. i call it inductive linguistic behavioral interpretability of LLMs, or LLM linguistics on a casual day.

over the last two years, i've been cornering LLMs with controlled linguistic contrasts, watching how they behave, and working backwards to which distinctions they show, ignore, or get stumped by. so far that's been (1) what makes linguistics olympiad puzzles tricky for them (COLM, 2025), (2) whether they model human linguistic variation in Hindi-English verb code-mixing (EACL, 2026), (3) how they process hierarchical sequences compared to crows, monkeys, children, and adults (CogSci abstract, 2026; journal version cooking), (4) a taxonomy of translation difficulty for the Last Translation Benchmark, and (5) whether LMs make second-language learner-like mistakes across development (WIP). next up: a survey stitching these together, and neurodivergence x AI.

before this i completed my MSc. in Computational Linguistics at IIIT Hyderabad. linguistic complexity and how tech models it, was also the theme for my masters thesis on pseudowords generation (e.g. for LLM subword probes) and native speaker evaluation design in low-resource languages. this theme started inspiring me since i was a high schooler, solving actual linguistic reasoning puzzles for fun.

everything fun i've worked on, is slotted below in the grid inspired by my thesis: rows go from linguistics-heavy to cognition-heavy work, and columns from controlled stimuli to naturalistic data. greens are done, purples are happening, blues are just starting off, join in! if a cell makes you go "huh?", that's on purpose, ask me :)

	controlledstimuli i construct	mixedcontrolled frames around natural phenomena	naturalisticlanguage as it occurs in the wild
linguistics-heavy	unveiLing (COLM '25): which typological features make linguistics olympiad puzzles hard for LLMs, across models, prompts, and scale	Hinglish verb code-mixing (EACL '26): whether LLM perplexities mirror native speakers' verb-choice variation, across model families, sizes, and language mixesFICLE (ECML PKDD '23): a dataset and classifiers for five linguistically defined types of factual inconsistency, with spans	Last Translation Benchmark (arXiv '26): a source-side taxonomy of translation difficulty across 109 languages, tested on state-of-the-art systemscode-mixed twitter morphologyTamil phonological transcription rules for speech-to-text, within dialectal variation across 9+ Indian languages (Navana Tech)Alexa: explanations for underperforming utterance clusters, from TF-IDF keyword patterns to contrastive few-shot prompting (Amazon)Alexa: multilingual synthetic data and adversarial prompting to reduce bias across demographic clusters (Amazon)
balanced	pseudowords (MS thesis '23): an LSTM + IPA generator and native-speaker evaluation for Hindi and English pseudowords, tested on password memorability and aphasia classification	Learner-Error Trajectories (ETH Zürich, WIP): whether LMs become sensitive to L2-English learner errors over training checkpoints, profiled with item response theory	Khmer MT quality and usability for immigrants
cognition-heavy	Stochastic Parrot (CogSci '26 abstract, journal version in progress): hierarchical vs. linear sequence processing in four LLM generations, compared with crows, monkeys, children, and adultsHindi representation distortion (ACCS '25): how number features and noun case distort in native speakers' memory during sentence comprehensionpassword metacognition (CogSci '21): beliefs about passphrase memorability vs. security, across demographics	nothing here yet, open to ideas	entropy-based quantification of empathy in NLP datasets and model outputsonline loneliness expressions
starting off, not placed yet	historical linguistics × LLMsthe same approach on pre-LLM language modelscreativity × LLMsneurodivergence × AI
across all of it	a survey of inductive linguistic behavioral interpretabilityconvergence tests: do the same effects recur across projects?

so, if you're into language, tech, comics, design, quizzing, board games, or just curious about what I'm up to, look around or hit me up! I keep it real, and it's always interesting :)

mukundc2k(at)gmail.com Google Scholar GitHub Twitter Bluesky LinkedIn ORCID

Action required

papers