You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
My project tries to understand what audio foundation models learn at different layers, and how those representations change depending on the model architecture and pre-training objective. In my solo-author paper, “What Do Audio Models Really Hear? Layer Selection and Mechanistic Structure in Sound Representations,” I compare representations across several audio foundational models using probing, similarity analysis, representation geometry, and mechanistic interpretability tools.
One of the main findings is that the final layer is often not the best one for downstream tasks in a lot of cases. In many models, useful information appears much earlier, and the pattern varies a lot depending on how the model was trained and what data it was trained on. I also look at how these layer-wise differences relate to robustness to noise and changes in audio. This gives a more practical way to identify which parts of a model are actually useful, without always having to retrain or evaluate the full model end to end.
The paper has been accepted to the NeurIPS 2026 Main Track as a poster with scores 4, 5, and 5, and I am the sole author and an undergraduate researcher.
NeurIPS is one of the best AI conferences in the world where the best researchers present their work.
The main research work have already been accepted and completed, and the final paper has been accepted to NeurIPS 2026 as well as multiple ICML workshops. The next stage of the project is presenting it for which i need funds. I would like to present the work in person, discuss the results with researchers working on representation learning, interpretability, and audio foundation models, and use their feedback to progress the field.
As a solo undergraduate author, attending the conference would also give me the opportunity to directly explain and defend my work, meet researchers around the world working on AI SAfety and Mechanistic Interpretabilty
The funding would be used exclusively to support my attendance at NeurIPS 2026 and presentation of the accepted paper.
Round-trip economy airfare from New Delhi, India to Sydney, Australia (~$900–$1,100 USD), accommodation in Sydney for the duration necessary to attend and present at the conference (~$1,000–$1,300 USD for 6–7 nights), Australian Visitor (subclass 600) Business Visitor visa expenses (~$174 USD / AUD 250), and NeurIPS 2026 student conference registration ($250 USD).
Estimated total: ~$2500 USD.
I’m Arnesh Batra, an undergraduate student at Indraprastha Institute of Information Technology Delhi (IIIT Delhi), my research interests are mainly representation learning, mechanistic interpretability, and AI safety.
This NeurIPS paper is a solo-authored project that i worked alone for months. I have also worked on several research projects across these areas, including an ICML 2026 Spotlight paper on intermediate representations in Foundational Models, two papers accepted to TMLR, and work accepted to EMNLP 2026. All of my work is as first auuthor and this is the first time I attempted solo work.
I have also been involved in research through academic labs and external research programs like SPAR and MARS, particularly on foundation models and mechanistic interpretability. NeurIPS 2026 would be my first time attending the conference, and it would also be my first opportunity to present a main-track NeurIPS paper in person. I also have 3 workshop papers accepted at NeurIPS.
My research page - https://arnesh2212.github.io/
Without external travel support, I would likely be unable to attend the conference and present my solo-authored paper in person.
I currently have $0 in confirmed external funding.