Swadesh Swain
Joint Secretary, DSG IIT Roorkee
—
Education
-
B.Tech, Electronics and Communication Engineering with Mathematics Minor, IIT Roorkee (2022–2026)
Research Interests
- Machine Learning & Deep Learning
- Interpretability in AI Models
- Alignment and Safety in AI
- Rienforcement Learning - Multi-Task RL, Generalization, Skill-Based Learning, World Models
- Robustness in Multi-Modal Models
-
Diffusion Conditioning for Synthetic Data Generation
Projects
-
[Re]Learning Hierarchical Image Segmentation For Recognition and by Recognition (Jan 2025-Present)
A reproducibility study under Data Science Group, IITR to reproduce results of CAST and extend it to cover SoTA benchmarks and models. -
Deep RL-enabled Fortnite Agent (Oct 2024-Present)
Developing an agent to play the complex game “Fortnite” using multi-linear ICVF via temporal difference learning and DreamerV3 as world model. -
Reproducing UPoP for Multimodal Transformer Models (Dec 2023-Feb 2024)
Reproduced results for UPop: Unified and Progressive Pruning for Compressing Vision-Language Transformers (ICML 2023). - LLM-AI-Generated-Text-Classification (Dec 2023-Jan 2024)
Research project focused on identifying AI generated text and enhancing robustness against artificial content. —Publications
- Swain, S., Singhi, S., (2024). Riemann Sum Optimization for Accurate Integrated Gradients Computation. Accepted to Interpretable AI Workshop 2024 @ NeurIPS 2024. (Co-Author)
-
Mittal, A., Pandey, A., Tiwari, A., Jindal, S., Swain, S., (2025). Revisiting CroPA: A Reproducibility Study and Enhancements for Cross-Prompt Adversarial Transferability in VLMs. Under Review in TMLR. (Co-Author)