The Ohio State University

Vishnu Kabir Chhabra

PhD Candidate · AI Safety & Mechanistic Interpretability

I'm a PhD candidate at The Ohio State University, advised by Prof. Xueru Zhang and Prof. Mohammad Mahdi Khalili, working on AI safety through mechanistic interpretability. I study how safety mechanisms in language models, such as refusal behavior and steering directions, change under compression, fine-tuning and corruption, and I build lightweight methods that make models safer.

Portrait of Vishnu Kabir Chhabra

01 research interests

AI Safety & Interpretability

  • Mechanistic interpretability
  • Refusal & safety mechanisms
  • Steering vectors
  • Sparse autoencoders
  • Model compression
  • Fine-tuning robustness

Machine Learning

  • Diffusion models
  • Reinforcement learning
  • Statistical learning
  • ML for wireless networks

02 news

  1. Advanced to PhD candidacy.

  2. Workshop paper on refusal in compressed models accepted at ICML 2025 (AIW).

  3. Neuroplasticity and Corruption in Model Mechanisms accepted at NAACL 2025.

  4. Workshop paper on ECG signal denoising accepted at ICML 2024 (NGSM).

  5. Our work HORCRUX on accurate cross-band channel prediction was accepted at MobiCom 2024.

  6. Passed my PhD qualifying exams.

  7. Paper accepted at CSCW.

  8. Graduated from the University of Minnesota and started my PhD at Ohio State.

03 publications

AI Safety & Interpretability

  1. UR

    Speaking the Steering Vector: Steering Vectors as Universally Transferable System Prompts

    V.K. Chhabra, X. Zhang, M.M. Khalili

    Under review

  2. UR

    Scoping the Emergence of Steering Vectors in Large Language Models

    V.K. Chhabra, Z. Zhu, X. Zhang, M.M. Khalili

    Under review

  3. UR

    Alignment Gradients are Pseudo-Refusal Directions

    V.K. Chhabra, X. Zhang, M.M. Khalili

    Under review

  4. UR

    Fixing What Fine-Tuning Breaks: A Simple and Efficient Approach

    V.K. Chhabra, D. Zhu, Z. Zhu, X. Zhang, M.M. Khalili

    Under review

  5. UR

    On the Transferability of Sparse Autoencoders for Interpreting Compressed Models

    S. Gupte, V.K. Chhabra, M.M. Khalili

    Under review

  6. NAACL2025

    Neuroplasticity and Corruption in Model Mechanisms: A Case Study of Indirect Object Identification

    V.K. Chhabra, D. Zhu, M.M. Khalili

    NAACL 2025

  7. ICML-W2025

    Towards Understanding and Improving Refusal in Compressed Models via Mechanistic Interpretability

    V.K. Chhabra, D. Zhu, Z. Zhu, X. Zhang, M.M. Khalili

    ICML 2025 Workshop (AIW)

Machine Learning, Systems & HCI

  1. MobiCom2024

    HORCRUX: Accurate Cross Band Channel Prediction

    A. Banerjee, X. Zhao, V.K. Chhabra, K. Srinivasan, S. Parthasarathy

    MobiCom 2024

  2. ICML-W2024

    ECG Signal Denoising Using Multi-scale Patch Embedding and Transformers

    D. Zhu, V.K. Chhabra, M.M. Khalili

    ICML 2024 Workshop (NGSM)

  3. CSCW2022

    Multi-Objective Personalization in Multi-Stakeholder Organizational Bulk E-mail: A Field Experiment

    R. Kong, C. Zhang, R. Sun, V.K. Chhabra, T. Nadimpalli, J. Konstan

    Proceedings of the ACM on Human-Computer Interaction (CSCW 2022)

    doi

04 selected projects

Cross-Band Channel Prediction

2022

Python · PyTorch · Diffusion models

End-to-end framework built on a conditional diffusion model that beats the prior state of the art by 2 dB SNR for downlink channel prediction; collaborated on the physics-based HORCRUX approach (MobiCom 2024).

Face Recognition for Dental-Assisting Robots

2023

Python · PyTorch

Generated synthetic benchmark data with Stable Diffusion and evaluated AdaFace and MogFace, reaching 90% accuracy for dual-factor authentication.

Food Calorie Estimation with WiFi

2023

MATLAB · Python · Spectral analysis

Estimated calories and nutrients of liquid and solid foods from WiFi signals with up to 10% error, robust to interference and food containers.

Deep RL for Automated Trading

2022

Undergraduate thesis

Benchmarked DQN, DDPG, PPO, SAC, A2C, TD3 and multi-agent DDPG for cryptocurrency trading, modeled as a POMDP. Awarded a Distinguished Poster.

05 experience

  1. Graduate Associate, AI Safety & Mechanistic Interpretability

    2022 – now

    The Ohio State University

    • Safety of compressed models: showed how pruning and quantization change the mechanisms behind refusal, and introduced AIRD, a lightweight algorithm that improves the safety of compressed models by up to 41% (ICML 2025 Workshop)
    • Fine-tuning and corruption: discovered neuroplasticity in language model mechanisms, a mechanistic account of how task-specific and harmful fine-tuning change a model (NAACL 2025)
    • Steering vectors and refusal directions: how steering vectors emerge in LLMs, their use as transferable system prompts, and alignment gradients as pseudo-refusal directions
    • Mentored undergraduates in the AIEdge REU (2023) on resource allocation and deepfake detection

  2. Software Engineering Intern

    2021

    Scriptulate

    • Built databases and anomaly detection on healthcare data (k-NN, SMOTE, spectral clustering, k-means, SVM)
    • Built knowledge graphs to analyze doctor–patient relationships

  3. Teaching Assistant

    2020 – 2022

    University of Minnesota – Twin Cities

    • Taught Java and OCaml to undergraduates remotely during the pandemic
    • Designed grading schemes, project assignments and debugging solutions

06 education

  1. PhD, Computer Science and Engineering

    2022 – 2027 (expected)

    The Ohio State University

    • Major: Artificial Intelligence. Minors: Statistical Learning, High Performance Computing
    • Advised by Prof. Xueru Zhang and Prof. Mohammad Mahdi Khalili
    • GPA: 3.9

  2. B.S., Computer Science and Mathematics

    2018 – 2022

    University of Minnesota – Twin Cities

    • Advised by Prof. Joseph Konstan and Prof. Ju Sun

07 skills

Research
Mechanistic interpretability, sparse autoencoders, steering vectors, circuit analysis, LLM safety & refusal evaluation, pruning & quantization, fine-tuning
Languages
Python, C/C++, Java, OCaml, SQL, CUDA
ML & interpretability
PyTorch, Hugging Face Transformers, TransformerLens, SAELens, TensorFlow, scikit-learn
Training & infra
PyTorch Distributed, DeepSpeed/FSDP, vLLM, Weights & Biases, Slurm, MPI, Docker, Kubernetes, AWS SageMaker

08 contact

Email

vishnu.kabirc@gmail.com

Languages: English, Hindi, French

Office

418 Baker Systems
1971 Neil Ave
Columbus, OH 43210