Harman Singh

I am a researcher at Google DeepMind, working on large scale multimodal and multilingual modelling with Dr. Partha Talukdar in the Languages team.

Previously, I was an AI Resident at Meta AI where I worked on reasoning abilities of Vision Language Models. Before this, I completed my undergrad from Indian Institute of Technology, Delhi, advised by Prof. Parag Singla. During this time, I have been a research intern at InkLab, USC, advised by Prof. Xiang Ren and a research intern at IBM Research AI.

Email  /  CV  /  Google Scholar  /  Semantic Scholar  /  Twitter  /  Github

Research

I am interested in language modeling, long-context capabilities of LLMs and multimodal LMs, reasoning, and post-training. Along with this, I am interested in building inclusive language technologies that can benefit a broader range of users.

I am looking for PhD positions starting Fall 2025

IndicGenBench is a multilingual, multi-way parallel benchmark for measuring language generation capabilities across diverse user-facing tasks in 29 Indic languages spanning 13 writing scripts and 4 language families.

Improving compositional reasoning capabilities of SOTA Vision-Language Models through a new Coarse-to-Fine contrastive learning technique as well as effective hard negative mining.

Benchmarking and improving retrieval augmented knoweledge editing of LLMs, in a cross-lingual and multi-hop setting.

New datasets and a modular method for weakly-supervised instruction guided image manipulations.

Proposed a modular method (FaiRR) for logical reasoning over natural language rule bases. Our methods ensure model faithfulness by assured causal relation from the proof step to the inference reasoning. FaiRR is more interpretable, efficient as compared to baselines, and generalizes better to OOD logical reasoning tasks.

A systematic evaluation framework designed to assess speech tokenizers comprehensively and shed light on their inherent characteristics, cheaply, without having to train large speech foundation models.

Past Work (Bioinformatics Research)