Harsh Udai

Direct-PhD, Artificial Intelligence · IIT Hyderabad

I study how vision–language models can move beyond brittle pattern-matching toward structured reasoning, building complex concepts from simpler parts.

Advised by Dr. Konda Reddy Mopuri and Dr. Vineeth N Balasubramanian. C-DOT STAR Scholar.

Email Scholar GitHub LinkedIn CV

Research

My work focuses on compositional generalization in machine learning. I treat vision–language models not only as capable systems but as imperfect approximations of multimodal cognition, and I investigate their internal representations, shortcuts and failure modes. Using mechanistic interpretability, I aim to reverse-engineer these models and turn the insights into training objectives and architectures that are more modular, transparent and reliable.

Vision–Language ModelsCompositionalityMechanistic InterpretabilityRepresentation LearningSparse Autoencoders

Publications

★ NeurIPS 2026

Beyond Hard Negatives: Grounded Positive Supervision for Compositional CLIP

Harsh Udai, Konda Reddy Mopuri, Vineeth N. Balasubramanian

Region-aligned multi-positive supervision and relation-aware triplet constraints that strengthen attribute–object binding and relational compositionality in vision–language embedding spaces.

Project Page Paper Code
★ NeurIPS 2026

LACE: Latent Alignment via Counterfactual Embeddings

Harsh Udai, Konda Reddy Mopuri

Sparse concept-space counterfactual editing that generates symmetric image-side and text-side hard negatives, with no pixel editing, external generators or retrieval mining.

Project Page Paper Code
Ongoing

Multimodal Sparse Autoencoders for Compositional Faithfulness

A shared sparse dictionary with cross-decoding that preserves compositional behavior in frozen embeddings, evaluated across CLIP, SigLIP and SigLIP2.

Background

Honors & service

📝 Reviewing

  • 🏅 Top Reviewer AwardNeurIPS 2026
  • Silver Reviewer AwardICML 2026
  • ReviewerICLR 2027, AAAI 2027, NeurIPS 2026, ICML 2026, ICVGIP 2025
  • Sub-reviewerCVPR 2026, ECCV 2026, ICVGIP 2026, NeurIPS 2025, TMLR

📢 Updates

  • 🏅 Top Reviewer Award at NeurIPSOctober 2026
  • 🎉 Two papers accepted at NeurIPSSeptember 2026
  • 🏆 Silver Reviewer Award at ICMLMay 2026
  • 🎓 Awarded C-DOT STAR ScholarshipJuly 2025

🧑‍🏫 Teaching Assistant

  • Foundations of Machine LearningAI5000 · Fall 2025
  • Deep LearningAI5100 · Spring 2025
  • Computer VisionAI3603 · Spring 2025

Contact

Happy to talk about compositionality, interpretability, or possible collaborations.

Email Scholar GitHub LinkedIn Kaggle