Shristi Das Biswas

Graduate Research Assistant at Purdue University.

prof.jpg

I am a Ph.D. researcher in Electrical and Computer Engineering at Purdue University, advised by Prof. Kaushik Roy (NRL). My research develops safe, controllable, and efficient foundation models, spanning:

  • Concept Unlearning & Safety in Generative Models – Removing Harmful/Copyrighted Concepts without retraining in T2I and T2V.
  • Multimodal Coreset Learning for LVLMs - Training-free Image-Text Coreset Selection to Accelerate Finetuning of Vision-Language Models without compromising performance.
  • Generative Model Efficiency – Post-Training Inference Acceleration and Dynamic Compression of Diffusion Models.
  • Continual Adaptation of Language Models – Scalable Lifelong Learning in LLMs for Continual Customization.
  • Efficient Video Understanding – Compute-Efficient Dual-stream architectures for Efficient Video Processing.
  • Controllable and Creative Content Creation – Inference-time Exemplar-based Image Editing for Generative Models.

My work is motivated by the intersection of Trustworthy AI, Efficient Generative Modeling, Continual Capability Improvement, and Foundation Model Alignment. I focus on building training-free mathematical frameworks that enumerate and help understand knowledge inside generative image and language models.

Past Experiences

Previously, I was an Applied Science Intern at Amazon AWS in 2025 and Amazon Fashion in 2024, working on pushing SOTA in Adaptive Multi-agent LLM Routing, and Creative Diffusion-Based Imagery Generation for Producing Lifestyle Content.

I have served as a peer reviewer for NeurIPS, ICCV, CVPR, ICLR, AAAI, WACV, and AAAI, where I also serve in the Program Committee.

Research Interests

Diffusion Models Model Editing Continual Learning for Customization Concept Unlearning Trustworthy ML Multi-Modal Coreset Selection Generative Compression Efficient Foundation Models

News

Mar 31, 2026 Ecstatic to share that our paper, ICE on Theoretically Grounded Instant Concept Erasure for Safe Text-to-Image and Video Generation was accepted to CVPR Findings 2026! 🎉
Jan 05, 2026 Excited to share that my work with AWS during my internship, ELLA on efficient continual adaptation and customization of LLMs has been accepted to EACL 2026! 🎉
Nov 13, 2025 Super thrilled to share that our paper, ICE on Theoretically Grounded Instant Concept Erasure for Safe Text-to-Image and Video Generation is live on arXiv! 🎉
Nov 06, 2025 Our new work on scalable deep learning has been accepted to WACV 2026 as a Highlight! The paper, SSA, proposes a structured Local-Learning framework that operates directly on low-rank manifolds to significantly reduce standard training compute while achieving accuracy comparable to Backpropagation! 🎉
Oct 20, 2025 Absolutely delighted to share that I’ve been awarded the NeurIPS 2025 Scholar Award! 🎉
Oct 06, 2025 Excited to share that my work with AWS during my internship, ELLA on efficient continual adaptation and customization of LLMs has been submitted to the ARR Oct 2025 Cycle. Preprint to be released soon! 🎉
Sep 25, 2025 Our latest work, SlimDiff on training-free, activation-guided hands-free slimming of diffusion models, is now available on arXiv! In this paper, we propose the first closed-form, diffusion trajectory-aware structural compression of Stable Diffusion Models that is entirely training-free! 🎉
Sep 22, 2025 Excited to share that the work from my recent internship at Amazon Kumo, AWS, ELLA: Efficient Lifelong Learning for Adapters in Large Language Models, on continual learning strategies for LLMs using PEFT to support lifelong adaptation and mitigate catastrophic forgetting has been accepted at the NeurIPS 2025 CCFM Workshop! 🎉
Sep 18, 2025 Thrilled to share that our paper, CURE on efficient unlearning for responsible and safe deployment of image diffusion models has been accepted to NeurIPS 2025 as a Spotlight Paper! 🎉
Jul 30, 2025 Honored to serve as a Program Committee member for the prestigious AAAI 2026! 🎉
May 13, 2025 Super excited to announce I’ll be joining Amazon Kumo, AWS in Bellevue, WA, as an Applied Science Intern! I’ll be working with Radhika Bhargava, Anwesan Pal and Yue Zhang, focusing on a very exciting challenge: developing continual learning algorithms for large language models. We’ll be researching strategies using parameter-efficient fine-tuning to allow for scalable lifelong adaptation using regularized low-rank adapter updates via cross-task subspace decorrelation. I’m incredibly excited to get started on this research, which will form the basis of the ELLA project. 🎉
Mar 17, 2025 Our new work on efficient video understanding has been accepted to WACV 2026! The paper, Learning Unified Spatio-temporal Representations for Efficient Compressed Video Understanding, proposes a lightweight framework that learning video representations directly from compressed videos. This avoids the need for decompression and achieves state-of-the-art performance with up to 15x faster inference! 🎉
Dec 09, 2024 My latest outing with Amazon Fashion with our internship paper, PIXELS on a framework for progressive reference image-driven editing with T2I models, to enable customization by providing granular control over edits and democratizing professional-grade edits for a wider audience got accepted at AAAI 2025! 🎉
May 06, 2024 Thrilled to announce I’m joining Amazon Fashion in Sunnyvale as an Applied Science Intern! I’ll be working with Prateek Singhal and Matthew Shreve on a super exciting challenge: designing new diffusion-model-based tools for image editing. We’re hoping to find ways to give users granular, pixel-level control and allow them to draw inspiration from multiple reference images to create their perfect custom image using off-the-shelf image diffusion models. This problem is the exact inspiration for PIXELS, a project I’m passionate about, and I can’t wait to dive in and start building! 🎉
Oct 13, 2023 I’m happy to share that our paper, HALSIE: Hybrid Approach to Learning Segmentation by Simultaneously Exploiting Image and Event Modalities, was accepted to the ICCV 2023 workshop! 🎉 I had a great time presenting this work, which shows how a hybrid SNN+ANN approach can achieve highly efficient multi-modal learning.
Aug 14, 2023 My first paper as a PhD student, HALSIE, on a hybrid approach to learning segmentation by simultaneously exploiting image and event modalities has been accepted at WACV 2024 as an Oral! 🎉
Aug 17, 2021 I started a new position as a Graduate Research Assistant in the Center for Brain Inspired Computing (C-BRIC) at Purdue University! 🎉
Aug 17, 2021 I joined the PhD program in Computer Engineering at Purdue University! 🎉

Selected Publications

  1. CVPR Findings
    Now You See It, Now You Don’t - Instant Concept Erasure for Safe Text-to-Image and Video Generation
    Shristi Das Biswas , Arani Roy , and Kaushik Roy
    2026
  2. arXiv
    SlimDiff: Training-Free, Activation-Guided Hands-free Slimming of Diffusion Models
    Arani Roy , Shristi Das Biswas , and Kaushik Roy
    arXiv preprint arXiv:2509.21498, 2025
  3. EACL
    ELLA: Efficient Lifelong Learning for Adapters in Large Language Models
    Shristi Das Biswas , Yue Zhang , Anwesan Pal , Radhika Bhargava , and Kaushik Roy
    In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics , 2026
  4. NeurIPS
    ELLA: Efficient lifelong learning for adapters in large language models
    Shristi Das Biswas , Yue Zhang , Anwesan Pal , Radhika Bhargava , and Kaushik Roy
    In AI That Keeps Up: NeurIPS 2025 Workshop on Continual and Compatible Foundation Model Updates , 2025
  5. NeurIPS
    CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models
    Shristi Das Biswas , Arani Roy , and Kaushik Roy
    In The Thirty-ninth Annual Conference on Neural Information Processing Systems , 2025
  6. AAAI
    PIXELS: Progressive image xemplar-based editing with latent surgery
    Shristi Das Biswas , Matthew Shreve , Xuelu Li , Prateek Singhal , and Kaushik Roy
    In Proceedings of the AAAI Conference on Artificial Intelligence , 2025
  7. WACV
    Feedback Alignment Meets Low-Rank Manifolds: A Structured Recipe for Local Learning
    Arani Roy , Marco Paul E. Apolinario , Shristi Das Biswas , and Kaushik Roy
    2025
  8. WACV
    Towards Scalable Modeling of Compressed Videos for Efficient Action Recognition
    Shristi Das Biswas , Efstathia Soufleri , Arani Roy , and Kaushik Roy
    arXiv preprint arXiv:2503.13724, 2025
  9. ICCV, WACV
    HALSIE: Hybrid approach to learning segmentation by simultaneously exploiting image and event modalities
    Shristi Das Biswas, Adarsh Kosta , Chamika Liyanagedera , Marco Apolinario , and Kaushik Roy
    In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , 2024