Vidushee Vats

Multimodal Learning and Generative Modelling

Contact: {X @ Y}, X=vatsvidushee, Y=gmail.com

Biography

Hi, I am Vidushee. I am a Research Fellow at X-NLP Lab, INSAIT, advised by Yuxia Wang. My main interests are multimodal learning, representation learning, any-to-any models, and diffusion. I also work on reinforcement learning and multi-agent systems.

Previously, I was a Research Intern at Halmstad University, working on any-to-any multimodal generation and alignment, advised by Shengzhi Li and Dr. Prayag Tiwari. I worked as a Research Engineer at Lexsi Labs and an AI Engineer at TUM & SAP. I was also a Computer Vision Research Affiliate at Georgia Institute of Technology, advised by Prof. Vijay Madisetti, where I worked on visual grounding in vision-language models. I graduated with a B.Tech in Computer Science (AI) from Bennett University.

Experience

X-NLP Lab, INSAIT

Research Fellow

Feb 2026 - Present

Working with Yuxia Wang on continual learning for personalized models that update after each user interaction. Built a multimodal system for personalized slide generation, with about 10% higher F1 than Claude Code and generation twice as fast and cheaper than Manus, Skywork, and Kimi.

Lexsi Labs

Research Engineer

Jun 2026 - Aug 2026

Built an agentic training framework for autonomous retraining of LLMs and VLMs. Led a three-stage move to an isolated execution environment to scale production agents.

Halmstad University

Research Intern

From Sep 2024

Advised by Shengzhi Li and Dr. Prayag Tiwari. Trained Qwen3.5-27B with SkyRL and Claude Code in a Harbor environment for autonomous scientific research. Built an asynchronous off-policy RL sampler for the Tinker API, cutting training cost from $1M to $300K. Worked on reward modeling for scientific question generation (Findings of ACL 2026), reduced visual information loss in any-to-any models, and replaced the NExT-GPT image decoder with FLUX.

Technical University of Munich & SAP

AI Engineer

Mar 2025 - May 2025

Built AI agents for SAP UI navigation and testing. Converted traditional test scripts to VLM-based navigation with 100% test coverage. Built a pipeline to fix failed scripts, reducing maintenance from weeks to hours.

Georgia Institute of Technology

Computer Vision Research Affiliate

Aug 2024 - Dec 2024

Designed a plug-and-play Visual Grounding framework. Outperformed SOTA models and GPT-4V by mIoU of 0.11 & 0.48 on RefCOCOg. Extended the grounding framework to accept multiple modalities (image, video).

Zocket AI

Computer Vision Intern

Apr 2024 - Aug 2024

Built a pipeline to fetch visually semantic videos from 100K+ videos with sub-0.5s latency. Designed an AI module for cross-platform ad resizing. Fine-tuned inpainting models for text-based background generation.

Research

IntelliAsk: Learning to Ask High-Quality Research Questions via RLVR

Karun Sharma*, Vidushee Vats*, Shengzhi Li*, Yuxiang Wang, Zhongtian Sun, and Prayag Tiwari

Findings of ACL 2026 / Project / Paper / arXiv / bibtex

Developed IntelliAsk to teach LLMs to ask critical, evidence-based research questions beyond imitating reviewer style. Designed IntelliReward with separate Transformer heads over a frozen LLM to assess effort, evidence, and grounding, achieving 72% agreement with expert labels versus 53% for GPT-5. RL training raises question quality to 0.55/3 versus 0.10 for OpenReviewer and DeepReviewer and 0.45 for GPT-5. Experts also rate IntelliAsk above Gemini 2.5 Pro (0.66 vs 0.60/3).

* Equal contribution

@inproceedings{sharma-etal-2026-preference,
  title={Preference Optimization for Review Question Generation Improves Writing Quality},
  author={Sharma, Karun and Vats, Vidushee and Li, Shengzhi and Wang, Yuxiang and Sun, Zhongtian and Tiwari, Prayag},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2026},
  year={2026},
  pages={25081--25104},
  doi={10.18653/v1/2026.findings-acl.1256},
  url={https://aclanthology.org/2026.findings-acl.1256/}
}

Think to Ground: Improving Spatial Reasoning in LLMs for better Visual Grounding

Karun Sharma, Vidushee Vats

ICLR 2025: Workshop on Reasoning and Planning for LLMs / Paper / bibtex

Improves spatial reasoning in LLMs to locate objects in images more accurately.

@inproceedings{sharma2025think,
  title={Think to Ground: Improving Spatial Reasoning in LLMs for better Visual Grounding},
  author={Sharma, Karun and Vats, Vidushee},
  booktitle={ICLR 2025 Workshop on Reasoning and Planning for Large Language Models},
  year={2025}
}

LLaVA-PlantDiag: Integrating Vision-Language Abilities for Conversational Plant Pathology Diagnosis

Karun Sharma, Vidushee Vats, Abhinendra Singh, Rahul Sahani, Deepak Rai, and Ashok Sharma

IJCNN 2024 / Paper / bibtex

@INPROCEEDINGS{10651096,
  author={Sharma, Karun and Vats, Vidushee and Singh, Abhinendra and Sahani, Rahul and Rai, Deepak and Sharma, Ashok},
  booktitle={2024 International Joint Conference on Neural Networks (IJCNN)},
  title={LLaVA-PlantDiag: Integrating Large-scale Vision-Language Abilities for Conversational Plant Pathology Diagnosis},
  year={2024},
  pages={1-7},
  doi={10.1109/IJCNN60899.2024.10651096}
}

An Improved Hybrid Model for Target Detection

Ujjwal Gupta, Roshan Golash, Vidushee Vats, and Karun Sharma

ICETCI 2023 / Paper / bibtex

@INPROCEEDINGS{10330945,
  author={Gupta, Ujjwal and Golash, Roshan and Vats, Vidushee and Sharma, Karun},
  booktitle={2023 International Conference on Emerging Techniques in Computational Intelligence (ICETCI)},
  title={An Improved Hybrid Model for Target Detection},
  year={2023},
  pages={265-270},
  doi={10.1109/ICETCI58599.2023.10330945}
}

Projects

AppliedScientist

Vidushee Vats, Karun Sharma, Shengzhi Li, and Shichao Pei

In progress / Project

An AI scientist and reviewer work in a loop to revise research papers. The scientist addresses feedback, runs experiments, and updates the manuscript; the reviewer checks each revision.

SlideLab

Vidushee Vats, Karun Sharma, and Yuxia Wang

In progress / Project

Generates scientific presentations from research papers, with agents for planning, visuals, layout, and grounding. ConfArena evaluates the slides through a simulated talk and audience questions.

LLMs for Scientific Discovery

Trained Qwen3.5 and Nemotron models with GRPO for scientific research. Built custom skills to help models read research papers and cited literature.

Any-to-Any Model

Trained encoder, LLM, and decoder alignment for multimodal generation. Integrated FLUX into NExT-GPT for image output and worked on interleaved outputs.

LLaVA-PlantDiag

Built a plant pathology conversation dataset with over 6,000 responses. Trained and evaluated vision-language models with PyTorch and DeepSpeed.