PhD researcher at IIT Bombay Monash

I'm Anirban Pal, a PhD researcher at IIT Bombay working across machine learning, biology, and chemistry. I study how models learn useful representations, retain knowledge, and generalize under distribution shift. My research spans continual learning, self-supervised learning, and JEPA-based world models, alongside biological and molecular representation learning for cancer research, drug discovery, and catalysis.

OpenReview profile

• Representation learning and generalization under distribution shift.
• Continual learning, self-supervised learning, and JEPA-based world models.
• Biological and molecular representation learning for cancer research and drug discovery.
• Machine learning for catalysis.

• Reconstruction-Based Selection Can Hurt Retrieval Compatibility
  Atharva Airen, Anirban Pal. Submitted to TMLR, 2026.

• Test-Time Feature Alignment Under Uncertain Correspondence
  Atharva Airen, Anirban Pal. ICLR 2027 conference submission.

• The Interface Contract: Understanding Representation Drift in Continual Representation Learning
  Atharva Airen, Anirban Pal. Submitted to NewInML workshop @ NeurIPS 2026.

• Beyond Marginals: What Test-Time Feature Alignment Can and Cannot Recover
  Atharva Airen, Anirban Pal. NewInML workshop @ NeurIPS 2026, poster.

• Deconstructing Gradient Reuse: A Prospective Target-Permutation Diagnostic for Online Learning
  Atharva Airen, Anirban Pal. NewInML workshop @ NeurIPS 2026, poster.

• The Interface Contract: Repairable Geometry and Partially Predictable Drift in Continual Representation Learning
  Atharva Airen, Anirban Pal. NeurIPS 2026 Workshop: NeurIPS Extended Abstracts.

• Pixel Reconstruction Keeps an Absolute-Coordinate Channel that Latent Prediction Lets Go
  Aarav Gupta, Anirban Pal. CAISc 2026, accepted paper.

• Action-Conditioned Sparsity for Latent-Space Planning
  Aarav Gupta, Anirban Pal. CAISc 2026, accepted paper.

Publication records and submission updates are available on my OpenReview profile.

VECTOR+ combines property-guided contrastive representation learning with latent-space sampling and a molecular decoder. My associated evaluation work examines the contributions of representation choice, activity supervision, latent sampling, and decoding on the released PD-L1 dataset.

Amartya Banerjee, Somnath Kar, Anirban Pal, and Debabrata Maiti. Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design. arXiv preprint, 2025.

Read the paper · Author code

Cross-electrophile coupling: molecular descriptor regression and fragment-based phosphine-ligand design. Workflow and data

For project questions, please use the issue tracker in the relevant public repository.