idhant

over the past years, i've spent countless hours understanding, thinking about, and working on beautiful problems in machine learning. i'm deeply interested in understanding these systems in a mechanistic manner—how their elegant architectures give rise to complex yet beautiful phenomena. i'm also passionate about exploring how these highly capable systems can be applied to real-world problems and how other fields can benefit from this technology.

currently, researching vision-language models, specifically investigating how misalignment emerges broadly across domains when it's only induced in a narrow domain during post-training.

over the summer, as part of Tzafon, i researched reasoning models. i spent my time developing a framework around sampling trajectories from language models to improve their reasoning capabilities. also, contributed to the pre-training regime by helping curate clean, high-quality data for model training.

in the past, independently researched on interpreting the inner working of mixture of experts (MoEs) for expert specialization and it got accepted at ICLR '25.

worked with my friends on densetex during buildspace, a fast and accurate image-to-latex model using transformers and CNNs. i also led my university's computer vision team for robomaster, a competitive combat robotics tournament, for over a year—where i learned a lot about optimizing systems for resource-constrained environments. additionally, i was part of a research lab at my university where i worked on machine learning algorithms to predict and optimize power consumption in 3d printers.

i like interacting with new people. if anything from the above interests you, feel free to reach out via twitter or email.

- idhant