Prasanna Mayilvahanan

I'm a fourth-year Ph.D. student at the Max Planck Institute for Intelligent Systems, Germany. Previously, I was a Research Intern at Cohere working on exploration methods and reward modeling for LLM reasoning, and at Apple MLR working on LLM pre-training and data curation. Prior to this, I completed my master's studies at USI Lugano / ETH Zurich, and my bachelor's at IIT Guwahati.

Email  /  CV  /  Research Statement  /  Blog  /  Scholar  /  Twitter  /  Github

Research

My research spans benchmarking LLM reasoning capabilities, robustness of LLMs and VLMs, and foundations of representation learning. I am particularly interested in developing methods that enable models to 'extrapolate beyond their training distribution.' Currently, I focus on (1) exploration-driven RL approaches that push models beyond their base capabilities, and (2) novel architectures, training objectives, and data curation strategies that promote creative generalization. Selected papers are highlighted.

Can models reason better with intermediate visualizations, akin to human mental imagery? MentisOculi evaluates this across frontier models and finds that visual thoughts do not yet improve reasoning.

Can RL post-training discover genuinely new solution paths, or does it merely sharpen what the base model already knows? MATH-B is a diagnostic benchmark designed to tell the two apart.

We find that two substantially different training setups—differing in architectures, tokenizers, optimizers, etc.—when trained on the same data and achieving identical training losses, consistently yield matching downstream performance across diverse tasks.

CLIP's high performance on style-centric domain shifts is significantly influenced by the presence of such images in its training set.

CLIP's ability to generalize to standard OOD benchmarks does not mainly stem from exact duplicates and near-duplicates in its training dataset.

We introduce a theoretical framework to analyze compositional generalization of neural networks within the regression setting.

We provide a neural network-based representation learning and clustering method for multi-omics data integration.

* denotes equal contribution
Template from Jon Barron's website.