Ashish Ramayee Asokan

Ashish Ramayee Asokan

Founding MTS · Stealth

I am a Founding Member at a stealth startup, working on post-training and distillation for language models and agents. I completed my MS in Machine Learning at Carnegie Mellon University (CMU), advised by Fernando De la Torre and closely collaborating with Raviteja Vemulapalli and Oncel Tuzel from Apple MLR. Before that, I was a Predoctoral Fellow at the Indian Institute of Science, advised by R. Venkatesh Babu.

My work examines how much of a frontier model's capability can be recovered through distillation and post-training, and how to rigorously characterize what changes between model versions — with the goal of making powerful models more accessible and their development more transparent.

CV Scholar Email GitHub LinkedIn Twitter
Research Interests

Distillation. Frontier models are powerful but locked behind APIs — you can query them but not inspect or modify their weights. I'm interested in how much of their capability can be recovered through careful distillation, and what data regimes and query strategies make this transfer most effective.

Model diffing. As models evolve rapidly, it's hard to know what actually changed between versions beyond a few benchmark numbers. I want to develop better tools for comparing models — frameworks that surface meaningful behavioral differences and track compatibility across a model family over time.

Distribution shift. I've also worked on getting models to transfer reliably when the test distribution differs from training. This spans transformer-based approaches for source-free adaptation, vision-language supervision for cross-domain generalization, long-tail strategies for visual recognition, and federated methods for fine-tuning large models efficiently.

Publications

* Equal contribution

Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification
CVPR 2024
Leveraging Vision-Language Models for Improving Domain Generalization in Image Classification
Sravanti Addepalli*, Ashish Asokan*, Lakshay Sharma, R. Venkatesh Babu
Distills OOD generalization from black-box VLMs such as CLIP into standard vision classifiers.
ArXiv Project Code BibTeX
DeiT-LT: Distillation Strikes Back for Vision Transformer Training on Long-Tailed Datasets
CVPR 2024
DeiT-LT: Distillation Strikes Back for Vision Transformer Training on Long-Tailed Datasets
Harsh Rangwani, Pradipto Mondal, Mayank Mishra, Ashish Asokan, R. Venkatesh Babu
Trains Vision Transformers from scratch on long-tailed datasets (for the first time in literature!) via distillation from CNNs.
Paper Project Code BibTeX
Domain-Specificity-inducing Transformers for Source-Free Domain Adaptation
ICCV 2023
Domain-Specificity-inducing Transformers for Source-Free Domain Adaptation
Sunandini Sanyal*, Ashish Asokan*, Suvaansh Bhambri*, Akshay Kulkarni, Jogendra Nath Kundu, R. Venkatesh Babu
A source-free domain adaptation framework for Vision Transformers (ViTs) that augments goal-task representations in the ViTs with domain-specific features learned via query weights for improved target adaptation.
ArXiv Project Code BibTeX
Aligning Non-Causal Factors for Transformer-based Source-Free Domain Adaptation
WACV 2024
Aligning Non-Causal Factors for Transformer-based Source-Free Domain Adaptation
Sunandini Sanyal*, Ashish Asokan*, Suvaansh Bhambri, Pradyumna YM, Akshay Kulkarni, Jogendra Nath Kundu, R. Venkatesh Babu
Aligning non-causal (spurious) factors before causal ones bridges the domain gap more effectively. Proposes a two-stage source-free adaptation framework using ViTs, whose shape bias makes them well-suited for this disentanglement.
ArXiv Project Code BibTeX
Distilling from Vision-Language Models for Improved OOD Generalization in Vision Tasks
ODRUM Workshop, CVPR 2023
Distilling from Vision-Language Models for Improved OOD Generalization in Vision Tasks
Sravanti Addepalli*, Ashish Asokan*, Lakshay Sharma, R. Venkatesh Babu
Shows that naive distillation from VLMs fails to transfer OOD generalization, and proposes alignment as the fix.
ArXiv Code BibTeX
Interpretability for Multimodal Emotion Recognition using Concept Activation Vectors
WCCI IJCNN 2022
Interpretability for Multimodal Emotion Recognition using Concept Activation Vectors
Ashish Asokan, K. Nidarshan, Anirudh V Ragam, Shylaja S S
Defines human-understandable concepts for Multimodal Emotion Recognition and uses Concept Activation Vectors to show that a BC-LSTM's reasoning can be expressed in terms of these concepts.
ArXiv BibTeX
Experience
Jun '26–Present
Founding Member of Technical Staff
Stealth
Working on post-training and distillation for language models and agents.
Post-Training Distillation LLMs
Feb '26–Mar '26
Founding Research Intern
VLM Run
Research on agentic visual reasoning with open-source vision-language models.
VLMs Agents
Jan–May '25
Research Scholar
Carnegie Mellon University
Advisor: Fernando De la Torre · Collaborators: Raviteja Vemulapalli, Oncel Tuzel (Apple MLR)
  • Designed a framework for model diffing — discovering divergent behaviors between foundation models.
  • Used model diffing to investigate gains and regressions in LLM-as-a-judge evaluation results across model versions.
LLMs VLMs Model Diffing
May '22–Jul '24
Predoctoral Fellow
Indian Institute of Science — Vision and AI Lab
Advisor: Prof. Venkatesh Babu
  • Research on domain generalization and vision-language models (CVPRW '23, CVPR '24), long-tail learning (CVPR '24), domain adaptation (ICCV '23, WACV '24), and federated learning.
  • Led a collaboration with Boeing, Wipro, & HCL to build an airport analytics system for vehicle collision detection, aircraft classification, and activity recognition from surveillance feeds.
  • Supervised and mentored 3 undergraduate interns across multiple research projects.
Distribution Shift Foundation Models Long-Tail Learning
Aug–Dec '21
Research Intern
Intel Corporation
Research on continual learning, federated learning, and neural network pruning.
Continual Learning Federated Learning Pruning
Education
Aug '24–May '25
M.S. in Machine Learning
Carnegie Mellon University
Coursework: Advanced Intro to ML (PhD), Convex Optimization (PhD), Probabilistic Graphical Models (PhD), Deep Reinforcement Learning (PhD), Probability & Statistics.
Aug '18–May '22
B.Tech (Honors) in Computer Science & Engineering
PES University, Bangalore
  • Ranked in the top 3% (30/975) of the CS Department.
  • Thesis: Interpretability for Multimodal Emotion Recognition.
  • Coursework: Linear Algebra, Intro to ML, Topics in Deep Learning, Information Retrieval.
Recognition
Outstanding Reviewer — Top 2%
Recognised among the top 2% of reviewers
CVPR & ECCV 2024
Kotak IISc AI-ML Pre-Doctoral Fellowship
Competitive research fellowship to work at the Vision and AI Lab
Indian Institute of Science, 2023
Prof. CNR Rao Merit Scholarship — Top 2%
Top 2% of the CS Department
PES University
Prof. MRD Merit Scholarship — Top 20%
Top 20% of the CS Department
PES University
Academic Service
NeurIPS'23 · '24 · '25 ICML'24 · '25 CVPR'24 · '25 · '26 ICLR'24 · '25 ICCV'23 ECCV'24 · '26 AAAI'24 AI-ML SystemsPC · '23 · '24