Atharva Mehta

Interests

• Individual behaviour modelling
• User simulation
• Cognitive digital twins
• Cultural norms in LLMs
• Music AI fairness

Away from the desk

• MusicI'm a trained Hindustani classical vocalist, which is part of why I care so much about how AI treats non-Western music.
• SportBadminton is how I switch off. I played a lot through college and I'm always up for a game.
• FoodI love exploring food of every kind, and I'm always looking for the next place to try.

News

• 2026Started my PhD in NLP at MBZUAI.
• 2026Detecting LLM-Assisted Academic Dishonesty using Keystroke Dynamics published in IEEE T-BIOM.
• 2026Missing Melodies accepted at AIES 2026.
• 2025Best Paper Award Candidate at ACM Multimedia.
• 2025Honorifics paper accepted to EMNLP 2025 (Main).
• 2024Best Paper Award at IJCB 2024.

Projects

• Cognitive digital twins in mental health2026 – now

  With Prof. Monojit Choudhury

  • A systematic survey of how the mental health literature conceptualizes and operationalizes cognitive digital twins.
  • A six-branch taxonomy of how digital twins are operationalized in mental health research.
  • Screened and coded 104 papers against the taxonomy to identify gaps and open challenges.
• Digital instrument preservation2026 – now

  With Prof. Monojit Choudhury and Dr. Pravina Manoharan

  • A framework for measuring instrument fidelity in AI-generated music, using signal-processing metrics for timbre and pitch.
  • A music Turing test across listener expertise levels.
  • Models keep Western instruments more faithfully than ones like sitar and sarangi, and their output is largely distinguishable from human performance.
• Music generation for the Global South2024 – 2026

  With Shivam Chauhan and Prof. Monojit Choudhury

  • Adapter-based fine-tuning for low-resource music traditions.
  • Representational-bias audits of music datasets and generation models.
  • Fairness frameworks and metrics for automatic music evaluation.
• Honorifics and socio-cultural norms2025 – 2026

  With Prof. Monojit Choudhury and Dr. Akhil Arora

  • Extended our EMNLP 2025 study of third-person honorifics in Wikipedia from Hindi and Bengali to 8 South Asian languages.
  • Factor analysis of how age, gender, fame and cultural origin drive honorific choice; preference analysis to evaluate LLMs.
  • Honorifics favour older and positively famous people; gender bias against women is strongest in Hindi; LLMs match Wikipedia only in some settings and languages.
• Keystroke dynamics for typing behaviour modelling2024 – 2026

  With Dr. Rajesh Kumar and Dr. Rajiv Ratn Shah

  • Models typing behaviour as a signature of the writing process, to tell genuine composition from LLM-assisted writing.
  • Keystroke datasets from 124 participants across writing, transcription and paraphrasing, plus a forged-keystroke attack set that mimics user-specific typing.
  • Gradient boosting and a Siamese LSTM (TypeNet) benchmarked against DetectGPT, LLaMA-3.3-70B and 44 human evaluators.
  • Over 97% F1 where text detectors and humans were near chance; adversarial training cut attack-induced false rejections from ~95% to 0%.
• Writing style imitation2026 – now

  With Dr. Tatsuki Kuribayashi and Prof. Preslav Nakov

  • LLM simulations of individual users' writing style on PersonaConvBench (Reddit).
  • Compares in-context learning (profile-based and few-shot), LoRA fine-tuning and preference-based RL on Qwen3-8B.
  • A contrastive BERT authorship model (400 authors, 4 domains, 89% macro-F1), with lexical markers, perplexity and AI detection, measures how closely each simulation matches its user.
  • Investigating activation steering vectors as an efficient alternative to fine-tuning for per-user simulation.

* Equal contribution. Full list on Google Scholar.

Talks & writing

• Missing Genres in AI Music Generation
  Root Access Podcast, with Shivam Chauhan.
• Algorithmic Republic: Is AI Ready to Govern?
  The Bloc, essay.

Education

• MBZUAI · Ph.D., NLP2026 – now

  Computation and Mathematics Division · Advisor: Prof. Monojit Choudhury

• IIIT-Delhi · B.Tech., CSE2021 – 2024

  GPA 8.93/10

Work

• Research Engineer, MBZUAIAug 2024 – Aug 2026
  • Led research on music generation for low-resource and Global South traditions, from adapter fine-tuning to bias audits.
  • Designed fairness frameworks and metrics for automatic music evaluation and aesthetics alignment.
  • Modelled cross-cultural user behaviour through honorific usage in Wikipedia and LLMs.
• LLM Research Intern, capSwipeJan – Jul 2024
  • Built fine-tuning and RAG pipelines that cut serving cost and time-to-deploy.
  • Deployed a production financial-advice chatbot using RAG and vector search.
• Software Development Intern, Reliance JioMay – Jul 2023
  • Shipped an LLM+RAG chatbot that doubled user adoption, and a JupyterLab extension.
  • Hardened authentication with LDAP and OpenID.

Awards

• Best Paper Award Candidate, ACM Multimedia2025
• Best Paper Award, IJCB2024

Teaching

• Teaching Assistant, MBZUAIAug 2025 – Aug 2026

  Intro to AI: tutorials and mentoring for 100+ students; graded assignments, labs and exams.

• Teaching Assistant, IIIT-DelhiMay 2022 – May 2023

  Convex Optimization, Operating Systems, and Probability & Statistics: tutorials and mentoring for 600+ students.

Service

Program Committee, AIES 2025 and 2026. Reviewer, EMNLP 2025 (Main) and the NeurIPS 2025 Workshop on AI for Music.

Tools

Python, C++, TypeScript · PyTorch, HuggingFace, vLLM, TensorFlow, LangChain, FAISS, Pinecone · Docker, Git, AWS, GCP, Azure, Runpod

Details

Email

atharva [dot] mehta [at] mbzuai [dot] ac [dot] aeShow

Scholar

Google Scholar

Code

github.com/atharva20038

Professional

LinkedIn

Where

MBZUAI, Masdar City, Abu Dhabi