Jayin Khanna

A HEAD FULL OF DREAMS

Jayin Khanna

I hold a BSc (Research) in Mathematics from Shiv Nadar Institution of Eminence (Batch Topper, 2026), and I'm concurrently completing an online BS in Data Science from IIT Madras.

I am currently working as a Project Associate at the Deep Representation Learning Lab, IISc (Dr. Prathosh A. P.)

My current research interests are in

  • Generative models (Diffusion and Flow matching) β€” their applications and core theory.
  • Representation learning for Vision tasks (problems and applications that deal with analysing, and using the geometry of latent spaces)
  • Interpretability, alignment & safety in LLMs

I am more interested in developing a theory of Deep Learning and understanding why methods work before scaling them. My current work is on safety alignment in text-to-image/video diffusion and trustworthiness repair in fine-tuned LLMs.

I'm looking for a long-term research project aimed at publication at ICLR, ICML, NeurIPS, or TMLR.

Preprints and Publications

Repairing the Trustworthiness of Distilled Language Models

Kumar Shubham, Devendra Vyas, Vidushi Agarwal, Jayin Khanna, Prathosh AP

AI4GOOD Workshop @ NeurIPS 2026 [Under Review]

πŸ“„ arXiv preprint coming by end of month

Steering Away from Harm: Safety-Contrastive Flow Matching for Inference-Time Correction of Text-to-Image Diffusion Models

Jayin Khanna, Nasrin Imanpour, Renjith Prasad Kaippilly Mana, Suranjana Trivedy, Amitava Das, Amit Sheth

ICLR 2027 [Under Review]

πŸ“„ arXiv preprint coming by end of month

Evaluation of Classical and Deep Learning Speech Time-Scale Modification: A Comparative Study

Jayin Khanna, Jesuraj Bandekar, Prasanta Kumar Ghosh

πŸ“„ arXiv preprint coming by mid-November

Current Research

Safety-guided flow matching for T2I/T2V diffusion β€” AI Institute of South Carolina (Dr. Amit Sheth, Dr. Amitava Das). Developing a safety-potential-guided rectified flow-matching formulation in CLIP embedding space to reduce harmful generations, benchmarked on the DETONATE dataset against TRCE, CURE, SAEUron, and DoCo.

Post-hoc trustworthiness repair in LLMs β€” Deep Representation Learning Lab, IISc (Dr. Prathosh A. P.). Combining EK-FAC curvature estimation with targeted gradient ascent to mitigate bias, unethical outputs, and toxicity in fine-tuned models β€” a compute-efficient alternative to retraining or RLHF, evaluated across Qwen2 and Pythia with submodular subset selection for non-redundant repair signals.

Education

  • BSc (Research) in Mathematics β€” SNIoE, May 2026
    Minor in CSE Β· Specialisations: AI & ML Β· Mathematical Finance
  • BS in Data Science & Applications β€” IIT Madras (ongoing)

Honours & Credits

My Research Interests

  • Core & applications of Generative Models (DDPMs, Flow Matching); T2I and T2V models
  • AI Interpretability and Alignment
  • Representation learning
  • Unsupervised and self-supervised learning

Exploring and Self Reading

  • Theory of Deep Learning (slow steady reading) / Geometric Deep Learning (pleasure reading)
  • World Models
  • Mechanistic Interpretability

My undergraduate thesis on generative models for unsupervised speech time-scale modification, under the joint supervision of Prof. Prasanta Kumar Ghosh (SPIRE Lab, IISc) and Prof. Niteesh Sahni (SNIoE), won 2nd Prize for Best UG Thesis. I document this work as detailed technical notes rather than leaving it in notebooks, posted on my Research page (Flow Matching, DDPMs, VAEs, GANs, contrastive learning).

My prior research spans:

I love doing research! Check out my Academic Development & Projects.

My Experience

IISc Deep Representation Learning Lab Logo

Research Assistant

Deep Representation Learning Lab, Indian Institute of Science (IISc), Bangalore

Supervisor:Dr. Prathosh A. P.

Period: Jun 2026 – Present

Working on enhancing trustworthiness in fine-tuned LLMs by implementing the EK-FAC preconditioned gradient ascent pipeline for post-hoc bias repair in Qwen2 and Pythia LLMs using Dolly for curvature estimation and CrowS-Pairs for targeted repair gradients.

LLM TrustworthinessEK-FACBias Repair...
AIISC Logo

Research Intern

IRT Group, AI Institute of South Carolina | DETONATE Group

Supervisor:Dr. Amit Sheth & Dr. Amitava Das

Period: Feb 2026 – Present

Working on safety alignment of Text-to-Image & T2V diffusion and flow matching models using a neurosymbolic approach called scene graphs.

T2I/T2V DiffusionSafety AlignmentFlow Matching...
SPIRE Lab Logo

Machine Learning Research Fellow

SPIRE Lab, Indian Institute of Science (IISc) Bangalore

Supervisor:Prof. Prasanta Kumar Ghosh

Period: Jul 2025 – Present (Summer Research Fellowship 2025)

Research on unsupervised speech TSM using Generative Models. Working with VAEs, GANs, Diffusion, and Flow Matching based speech TSM models to preserve speaker identity, naturalness and intelligibility.

Speech ProcessingGANsDiffusion...
IIT Kharagpur Logo

Deep Learning Research Intern

Indian Institute of Technology (IIT) Kharagpur

Supervisor:Dr. Niloy Ganguly

Period: 2025

Worked on developing attribution techniques using Integrated gradients, Manifold IG, Guided IG towards neural network interpretability.

Explainable AIAttribution MethodsIntegrated Gradients...
DRDO INMAS Logo

Machine Learning Research Intern

Institute of Nuclear Medicine and Allied Sciences-DRDO, Ministry of Defence

Supervisor:Dr. Shilpi Modi

Period: 2024

Developed a Convolutional Neural Network (CNN) ResNet model for classifying sEMG stress measurements, focusing on improving accuracy and generalization.

Machine LearningCNNsEMG Classification...
UCSC Logo

Statistics Research Intern

University of California Santa Cruz, CA (ISRP)

Supervisor:Prof. Bruno SansΓ³

Period: 2024

Conducted advanced research in Analysis of Time-Varying Quantiles for Environmental Variables, under the mentorship of Professor Dr. Bruno SansΓ³.

StatisticsEnvironmental DataTime Series Analysis...
RightProfile Logo

Machine Learning Intern

RightProfile by Syntellect

Period: Dec 2024 – March 2024

Part of the Research and Development team to develop Computer Vision and Deep Learning models to automate the annotation of 10,000+ raw images.

Computer VisionObject DetectionDeep Learning...
The Habitats Trust Logo

Machine Learning Intern

The Habitats Trust

Period: Dec 2024 – Jan 2024

Conducting research and development on modern Computer Vision and object detection models such as MegaDetector, Zamba, and Timelapse to classify and analyze wildlife in camera trap images.

Wildlife AIComputer VisionConservation Tech...