EACL 2026
10B and 87B models; led the bilingual Hindi–English data strategy end to end.
Research
I'm interested in what happens to reasoning when a model leaves English. Chain-of-thought and test-time compute work well in high-resource settings; whether those capabilities survive adaptation to mid- and low-resource languages — and how to tell genuine transfer apart from pattern matching — is the question I keep returning to.
Most of my work sits where that question gets decided in practice: post-training data, benchmark design, and the evaluation infrastructure that determines what a release is allowed to claim.
Language adaptationReasoning & test-time computeEvaluation designPost-training dataSafety benchmarks
Languages I've built or evaluated models for
हिन्दी/العربية/қазақ тілі/English
Selected work
arXiv 2025
70B reasoning model released with weights, data, and intermediate checkpoints.
arXiv 2025
Competitive reasoning at 32B, without frontier-scale compute.
COLM 2025
Adaptation recipe for a moderately-resourced language.
arXiv 2025
Symbolic benchmark spanning 12 domains; 30+ models evaluated.
arXiv 2025
Open instruction-tuned model for Hindi.
arXiv 2025
Entity framing and narrative role analysis over raw text.
Hugging Face
Open weights and datasets.
Experience
Research Engineer
Sep 2025 — Present
MBZUAI, Institute of Foundation Models
- Co-authored K2-V2 (70B) and K2-Think (32B); supported the K2-Think V2 release and its evaluation
- Built evaluation tooling for long-context, math, code, and safety benchmarks — prompting, deterministic scoring, reporting
- Added regression tests and automated reports that catch quality and safety regressions before release
- Gave technical talks and ran office hours for the K2-Think hackathon series
Research Assistant
Oct 2024 — Aug 2025
MBZUAI, Institute of Foundation Models
- Led development and release of the Nanda family (10B, 87B) and drove the bilingual Hindi–English data strategy
- Contributed dataset curation and evaluation to Jais-2 (Arabic) and Sherkala-Chat (Kazakh)
- Curated Suraksha Eval, a Hindi safety benchmark, and built the Hindi portion of TxT360
- Co-developed FinChain, a financial reasoning benchmark across 12 domains
Research Fellow
Mar 2024 — Sep 2024
SimPPL, with Swapneel Mehta
- Designed multi-agent experiments measuring and reducing fake-news propagation between LLM agents
- Worked with postdocs at MIT, Princeton, and Oxford on intervention design and evaluation
NLP Intern
Jun 2023 — Jan 2024
MBZUAI, with Preslav Nakov
- Built a multithreaded pipeline collecting 160K+ news articles from 5K+ sources
- Shipped an end-to-end media factuality and bias scoring system
- Trained transformer and NELA+CatBoost ensembles for article-level prediction and source profiling
Education
Thadomal Shahani Engineering College
Mumbai University
B.E. in Computer Engineering — CGPA 8.9 / 10
Contact
Happy to talk about multilingual evaluation, reasoning transfer, or open model releases — aaryamonvikram@gmail.com.