Anshul Singh

I am a Research Associate at the IACV Lab, IISc Bangalore, advised by Prof. Soma Biswas. My research centers on Large Language Models, with a particular focus on multimodal systems, visual reasoning, and document understanding. My work explores how to enhance the reasoning capabilities of language and vision models. Specifically, I have worked on:

• Adverserial Robustness and Safety of Multimodal LLMs and Diffusion Models
• Multimodal Analysis of open source information using LLM-guided Active Learning
• Multimodal reasoning for multi-tabular data using Vision-Language Models
• Multilingual alignment for low-resource Indic languages
• Vision Transformers and Signal analysis for fault detection and vocal pattern recognition

Previously, I was a research intern at the LT Research Group, University of Hamburg, with Prof. Chris Biemann and Jan Strich, and a MITACS Globalink Research Intern at Dalhousie University. I completed my Bachelor of Engineering in Information Technology at Panjab University. During my undergraduate studies, I also worked as a machine learning research intern at IIT Roorkee and served as the ML/AI Lead at Google Developer Student Clubs.

Research Interests: Multimodal and Multilingual Reasoning | Vision-Language Models | Interpretability of LLMs | Information Retrieval

Hobbies: Outside of research, I enjoy traveling, reading, writing blogs, minimalist living, and tinkering with small experiments. I'm always happy to connect for a discussion or collaboration!

[Email / CV / Github / Twitter / Google Scholar / Linkedin / Blogs]

In this work, we address a critical gap in Vision-Language Model (VLM) evaluation by introducing MTabVQA, a novel benchmark for multi-tabular visual question answering. Our benchmark comprises 3,745 complex question-answer pairs that require multi-hop reasoning across several visually rendered table images, simulating real-world documents. We benchmark state-of-the-art VLMs, revealing significant limitations in their ability to reason over complex visual data. To address this, we release MTabVQA-Instruct, a large-scale instruction-tuning dataset. Our experiments demonstrate that fine-tuning with our dataset substantially improves VLM performance, bridging the gap between existing benchmarks that rely on single or non-visual tables.

In this work, we address the growing security challenges in IoT networks by conducting a comprehensive comparative analysis of machine learning classifiers for intrusion detection. We evaluated five distinct models on two real-world IoT network traffic datasets to identify the most effective algorithms for detecting malicious activity. Our findings show that tree-based models, specifically Random Forest and Decision Trees, deliver outstanding performance, achieving accuracies exceeding 99%. This research provides a clear benchmark and practical guidance for developing robust and high-performance security systems to protect vulnerable IoT environments.

We propose HybridNet, a data-efficient framework that leverages hybrid active learning to select the most informative samples, drastically reducing labeling cost. We also propose a lightweight Reasoning-Aware Classifier (RAC) for challenging cases, which combines Vision–Language Model (VLM) features with reasoning from a Multimodal Large Language Model (MLLM) to further improve detection performance and provide human- interpretable explanations.