Project summary

I’m Gaurav Hadavale, a final-year undergraduate researcher from India. I’m looking for $1,100 to help me travel to Strasbourg and present my sole-authored paper, “Reading Between the Lesions: Auditing Whether Multimodal Dermatology Classifiers Actually Use Clinical Text,” at the MICCAI 2026 Workshop on Mechanistic Interpretability for Medical Foundation Models.

You can learn more about me:

https://gauravhadavale07.github.io/

The basic question behind the paper is pretty simple: if a multimodal model gets an answer right, how do we know it actually used the information we expected it to use?

I built the experiments myself using matched counterfactuals, ablations, activation analysis, and a Top-K Sparse Autoencoder. I used these to look at whether medical vision-language models were actually making use of clinical text or whether some of their performance could be explained by shortcuts and learned priors.

The paper has been accepted and the code is publicly available. The research is already finished. My main problem now is that I don't have enough money to cover the trip to France.

Paper: https://drive.google.com/drive/folders/1VWBZZBLynfmA5ad7iiC8N8l40reYE0XI

Code: https://github.com/gauravhadavale07/reading-between-the-lesions

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

My main goal is to present the paper at MI4MedFM and get feedback from people who work on mechanistic interpretability and multimodal foundation models.

In particular, I want to:

1. Present the results and discuss the auditing approach with researchers working on similar problems.

2. Get feedback on the counterfactual experiments and activation-level analysis, including where the methodology might be weak or could be improved.

3. Talk to researchers who are interested in applying similar ideas to other multimodal or language models.

4. Keep the code and results publicly available so that people can reproduce the experiments and build on them.

What interests me about this work from an AI safety perspective is that I don't think accuracy alone tells us enough about a model. A model can get an answer right while relying on information or patterns that we didn't actually want it to use. I'm interested in whether mechanistic interpretability can help us find these kinds of problems.

I used medical AI as the setting for this project because it gives me a concrete way to study the problem, but the underlying question isn't specific to medicine.

How will this funding be used?

Funding amount and breakdown
Budget breakdown:

$750 — Round-trip flight from Mumbai, India to Paris, France
$300 — MICCAI 2026 student registration
$100 — Schengen visa/application costs
$112 — Paris–Strasbourg and local transportation
$66 — 5% contingency buffer

Total project budget: $1,328

Personal contribution: $228 (accommodation and other expenses


Requested amount (USD): $1100

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

I'm working on this independently and I'm the sole author of the paper.

For this project, I designed and ran the experiments myself. This included the multimodal fusion baselines, matched counterfactual audits, ablations, activation analysis, Sparse Autoencoder analysis, and the LLaVA-Med experiments. I also wrote the paper and made the code publicly available.

This is my second sole-authored international research acceptance. My previous paper was accepted as a poster at the European Congress of Radiology (ECR) 2026 in Vienna, although I couldn't attend because I couldn't afford the travel costs.

I'm currently interning at A*STAR Singapore, where I'm working on trustworthy evaluation of AI systems, including work related to mechanistic interpretability and auditing multimodal models.

So far, I've mainly been doing this work independently, from coming up with the research question to running the experiments and writing the paper.

What are the most likely causes and outcomes if this project fails?

The biggest risk is simply that I don't manage to raise enough money in time to travel to Strasbourg. Since I'm travelling internationally from India, the flight cost and visa timeline are the main constraints. If I can't secure the money in time, I may have to give up the trip and not present the paper in person.

There is also a chance that attending the workshop doesn't lead to as much useful feedback or follow-up as I hope. I can't really predict how much interest the work will generate beforehand.

In either case, the research itself won't disappear. The paper has already been accepted, and the code and results will remain publicly available. The main thing I'd lose would be the opportunity to present the work in person, get feedback from researchers in the area, and have conversations about where I could take the work next.

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

$0.

I asked my university and the MICCAI Student Community about possible travel support, but I wasn't able to get funding for the trip.

I'm also applying to other funders, including BlueDot Impact and EA Funds, because I can't cover the full cost of the trip myself. If I receive funding from another source, I won't use multiple grants to pay for the same expense.