LLM Safety & Alignment
Auditing and repairing the safety of fine-tuned language models: measuring how alignment drifts during fine-tuning, and restoring it without giving up task performance.
Anshul Kaushal
AI Researcher · M.Tech. Artificial Intelligence, IIT Patna
Hello! 👋 I study how large models work on the inside and how to keep them safe. My research covers LLM safety, mechanistic interpretability, multimodal learning and computer vision.
Across five research internships at Lexsi Labs, IIT Jodhpur, the National Astronomical Observatory of Japan, IISER Bhopal and IIIT Allahabad, I have co-authored six papers at venues including EMNLP, BMVC, The Astrophysical Journal and a NeurIPS workshop. Most recently I built SAFETUNE, a framework for auditing and repairing safety drift in fine-tuned LLMs.
If you share these interests, I'd love to talk and collaborate. 😉
Research Interests
Mechanistic Interpretability
Reverse-engineering model internals through safety-circuit discovery, steering vectors and representation analysis, with a focus on multilingual LLM safety.
Multimodal Learning
Data-efficient image–text representation learning, such as pairing self-supervised vision encoders like DINOv2 with CLIP-style objectives for multimodal LLMs.
Computer Vision
Efficient medical image segmentation, state-space vision models (Mamba), uncertainty-aware architectures and explainable AI for visual recognition.
Experience
Education
Indian Institute of Technology, Patna
M.Tech., Artificial Intelligence
Panjab University
B.E., Computer Science & Engineering · CGPA 9.07 / 10
Research Experience
-
Lexsi Labs (Aurionpro)
Research Intern: Mechanistic Interpretability and Safety in LLMs
- Built SAFETUNE, a unified AI-safety framework that integrates 118+ methods for auditing, repairing and preserving safety in fine-tuned language models.
- Applied mechanistic interpretability, safety-circuit discovery and steering vectors to analyse multilingual LLM safety and support model alignment.
-
Indian Institute of Technology Jodhpur
ACM IKDD Scholar: Self-Supervised Vision and Multimodal Learning
- Built a CLIP-like image–text representation model by integrating the DINOv2 self-supervised vision encoder into MetaCLIP, for downstream multimodal LLM applications.
- Aligned the vision backbone using 99K CC and LAION image–caption pairs, reducing the data and compute needed for multimodal representation learning.
- Improved zero-shot classification by ~2% on ImageNet-1K, ObjectNet and iNat2021.
-
National Astronomical Observatory of Japan
Research Intern: Sparse Signal Reconstruction
- Reconstructed sparse Gamma-Ray Burst (GRB) light curves across four morphologies: Good, Break, Flares, and Break with Flares.
- Benchmarked Conditional GANs, Mamba state-space models and polynomial fitting for sparse signal reconstruction.
- Reduced uncertainty relative to prior studies to 37% for log Ta, 38% for log Fa and 41% for α.
-
IISER Bhopal
Research Intern: Medical Image Segmentation · Guide: Prof. Vinod K. Kurmi
- Studied Vision-Mamba state-space models for multi-domain medical imaging and semantic segmentation.
- Extended CNN–Transformer segmentation models with an adaptive, learnable fuzzy-logic module across four public datasets.
- Improved accuracy by 2% while cutting FLOPs, memory use and parameter count by 30%.
-
IIIT Allahabad
Research Intern: Emotion Classification using EEG · Guides: Prof. Uma Shankar Tiwary & Prof. Mohammad Asif
- Classified 24 emotions from high-density EEG data for affective computing.
- Tuned hyperparameters with a genetic algorithm, improving accuracy and F1-score by 4%.
- Used explainable AI (XAI) to identify the EEG frequency bands linked to specific emotional states.
Publications
-
NeurIPS 2026 Workshop
Drift, Then Repair: A Safety Audit of Fine-Tuned Language Models
NeurIPS 2026 FLLMPT Workshop
-
EMNLP 2026 Demo
SAFETUNE: A Unified, Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMs
EMNLP 2026, System Demonstrations Track
-
BMVC 2025 · Core A
CLFSeg: A Fuzzy-Logic based Solution for Boundary Clarity and Uncertainty Reduction in Medical Image Segmentation
British Machine Vision Conference (BMVC) 2025
-
JHEAp
Multi-Model Framework for Reconstructing Gamma-Ray Burst Light Curves
Journal of High Energy Astrophysics
-
ApJ
Gamma-Ray Burst Light Curve Reconstruction: A Comparative Machine and Deep Learning Analysis
The Astrophysical Journal
-
ICCCNT 2024
Optimizing Emotion Recognition in EEG Data: A Genetic Algorithm Approach with XAI Insights
ICCCNT 2024
Full list on Google Scholar.
Projects
COVID-19 Lung Infection Segmentation
- Implemented Trans, Residual, Squared and Recurrent U-Net variants, reaching a Jaccard index above 0.87.
- Served model inference through FastAPI and Streamlit for real-time predictions.
Academic Research RAG-bot
- Built a retrieval-augmented generation pipeline with GPT-3.5 Turbo and text embeddings for answering questions about research papers.
- Used Pinecone with dot-product similarity for semantic search over embedded papers.
Rock-Paper-Scissors Classifier
- Fine-tuned a pre-trained ResNet-18 with data augmentation, reaching 85% training accuracy.
- Interpreted learned features using Grad-CAM and t-SNE visualisations with OpenCV.
Skills
PyTorch
Hugging Face
TensorFlow
Keras
Scikit-learn
OpenCV
NumPy
Pandas
Python
C / C++
SQL
LaTeX
- LLM & Multimodal AI
- LLM safety, mechanistic interpretability, model alignment & auditing, safety-circuit discovery, steering vectors, multimodal LLMs, RAG, text embeddings, vector databases
- Machine Learning
- Deep learning, representation & self-supervised learning, transfer learning, generative AI, computer vision, semantic segmentation, hyperparameter optimisation, explainable AI
- Models & Methods
- DINOv2, MetaCLIP, CLIP, Mamba SSMs, Conditional GANs, CNN–Transformers, U-Net, ResNet, genetic algorithms, fuzzy logic, Grad-CAM, t-SNE
- Languages, Libraries & Tools
- Python, C, C++, SQL, MATLAB, LaTeX, JavaScript · PyTorch, TensorFlow, Hugging Face, Matplotlib, Seaborn, FastAPI, Streamlit · Linux, Git, Docker, CUDA, AWS, GCP, Jupyter
Honours
Achievements
- MEXT 2027 Scholarship: top 186 of 7,300 applicants (top 2.5%)
- ACM IKDD Uplink Research Internship 2025: 2% acceptance rate
- Oxford Machine Learning Summer School (OxML) 2025: shortlisted
- IELTS Academic: 8.0 overall band
Certifications & Service
- MLx Fundamentals (OxML): ML, DL, Representation Learning, Optimisation, Generative AI
- DeepLearning.AI: Unsupervised Learning, Recommender Systems, Reinforcement Learning
- Imperial College London: Mathematics for ML (Linear Algebra)
- AWS Elements of Data Science · Duke University: Data Science Math Skills
- Core Team, Training & Placement Cell (Oct 2023 – present): co-organised 10+ ML/AI workshops reaching 200+ students
Connect
Open to research collaborations and conversations about AI safety and interpretability.
