Project summary

I am Gaurav Hadavale, final year undergraduate student at University Of Mumbai, India. I am seeking Travel and Conference Registration Fee Support of $1000 for presenting my sole-authored paper at the MICCAI Mechanistic Interpretability For Medical Foundation Models Workshop 2026, which is to be held in Strasbourg ,France [September 27 to October 1, 2026].

Paper: Reading Between the Lesions: Auditing Whether  Multimodal Dermatology Classifiers Actually Use Clinical Text
Code Available: [Github]

Although the experiments are in dermatology, I believe my contributions are primarily technical AI safety work, as the question investigated is whether multimodal models that seem to use visual and textual evidence are genuinely using semantic information or utilizing various shortcuts and priors.

My work interrogates this hypothesis by performing matched counterfactual tests, ablations, activation analysis, and interpreting a Top-K Sparse Autoencoder, which has been trained on fused representations, to identify whether multimodal models’ high scores are due to actual semantic grounding. I find that in my experiments, aligned clinical text often failed to increase the correct-class margin, ablation probes identify reliance on superficial text presence, and one candidate SAE feature acts similarly to a peripheral context sensitivity probe. Furthermore, I find that when using contradictory clinical histories, LLaVA-Med fails to update predictions, suggesting that it relies on class prototypes from pretraining.

A small grant would enable me to present an already-accepted, single-authored interpretability paper, by an independent undergrad researcher with no other travel funding support options, at the conference. Without institutional support, I might have to pull out of presenting this work entirely. Earlier this year, I had to skip presenting a similarly single-authored international paper at the European Congress of Radiology in Vienna because I couldn’t afford the travel costs.

Timing is critical, as I need to apply for my Schengen visa before the August 24, 2026 deadline, and cannot do so without confirming I can fund the trip

What are this project's goals? How will you achieve them?

he immediate goal is to present my paper that has been accepted to the MICCAI MI4MedFM workshop and obtain feedback from researchers who are working on similar mechanistic interpretability and multimodal foundation model questions.

The more significant goal is to contribute to technical AI safety research by publishing my findings on developing practical audits for detecting whether a model appears to utilize text-visual grounding while still relying on various shortcuts and, more generally, emphasizing the importance of interrogating the relationship between model performance and mechanism. At the workshop, I hope to:

1. Share my findings on multimodal grounding failures and discuss them with others who are interested in similar questions of foundation model shortcutting and mechanistic interpretability of vision-language models.

2. Obtain feedback on my methodology, especially the matched counterfactual design, from the perspective of workshop attendees who are experts in mechanistic interpretability and trustworthy AI.

3. Identify potential collaborators who are interested in extending these findings to other domains such as general language models and vision-language pretraining models.

4. Post a short AI-safety aligned summary of my paper findings with feedback and next directions on the workshop, along with making my code and experiment results publicly available for others to build on.

This workshop is an opportunity for me, as an independent researcher based in India, to connect with the broader mechanistic interpretability, AI safety, and vision-language model communities. In addition to presenting my work, I am hoping to establish myself within these communities by engaging with the ideas and people at the workshop.

More about me: Linkedin

How will this funding be used?

• The total estimated cost is $1,212.
  Round-trip airfare: $778
  Budget dormitory accommodation, 5 nights: $100
  Conference registration fee: $178
  Schengen visa application fee: $156
  Total estimated cost: $1,212
  Funding requested: $1,000
  Personal contribution: $212
  I have kept the budget minimal by choosing budget accommodation and only requesting support for the core costs required to attend and present the paper.

Who is on your team? What's your track record on similar projects?

I am the sole author of this paper.

This is my second single-author international paper acceptance. My first paper was accepted to the European Congress of Radiology 2026 in Vienna, but I could not attend due to financial constraints. Currently, I am interning at the A*STAR Singapore, working on trustworthy evaluation of AI systems, with a focus on mechanistic interpretability, auditing of multimodal models, shortcut discovery, and semantic grounding failures. For this project, I designed and executed the experiments, including multimodal fusion baselines, matched counterfactual audits, ablations, activation analysis, interpretation of a Top-k SAE, and testing of LLaVA-Med. I aim to contribute to technical AI safety by building auditing methodologies that highlight model failure modes and make them more transparent before deployment in high-stakes settings

How much money have you raised in the last 12 months, and from where?

I have raised US$0 in the last 12 months. I explored funding through my university, and the MICCAI Student Community , but I was not able to get financial support for international conference travel, and I have not received any external funding.