Project Summary

The thing is, the typical use case with a large language model on a regular computer is to reduce the size of the language model so it fits in video memory. This can be achieved using a method known as 4-bit NF4 compression. Now when I was testing this, what I found out is that some of the safety features that you can build into the model can be broken by compressing the model in this way. For instance, asking the model a question in a language such as Hindi or Bengali could cause a harmful answer to be given by the compressed model, even if it would not have been given by the original model. This is an issue as it implies that the compressed model is not as secure to use as we had speculated.

I was just recently accepted into the NeurIPS2026 in Sydney for the GlobalSouthAI workshop for my research paper. The problem is, that I am from Mumbai and require some assistance in getting here. I'm searching for funding for a travel grant of $2,000 to pay for my travel. It would be incredible to show my work in person and get to meet some of the other people in the red-teaming area, and just have a conversation in person about ideas. If it can happen, I hope to have a lot of great conversations and learn from those in the field.

I've uploaded my paper and acceptance email here.

[Google Drive folder](https://drive.google.com/drive/u/2/folders/11aUMaut7oxrEQkzu_5dnM7xsfhWSHQtk)

What is the purpose of this project? How will you do it?

I want to present my work from NeurIPS, and have some experts in the area discuss how they find themselves simplifying complex models. I'm especially interested in talking to folks who work on the control of activations and see if they have any suggestions on how to keep our models safe when we decrease their precision, without resorting to really high precision numbers.

I have been cleaning up my evaluation scripts following the conference and am making everything public on GitHub. The aim is to develop a simple test that can be used by any user to assess the performance of their quantized model, whether it be AWQ, GPTQ or a 70B checkpoint, on our multilingual safety tests and find where it's failing. In this manner, individuals will be able to easily conduct the tests and determine what the weaknesses are in the safety measures.

How will you know if it is successful?

I think this will be successful if I manage to present my poster at NeurIPS and finally complete the paper for publication, and put our code and test data sets in the public domain at GitHub, so that other people can easily reproduce our results. This way, everyone can see what we've accomplished and build upon our work.

What impact will this funding have?

The trip from Mumbai to Sydney is just $2000.

So I want to go to Sydney, I've worked out the cost. Mumbai to Sydney is approximately $1200 for an economy ticket on both flights, reasonably priced for a round trip economy ticket. There is the lodging - I'm considering a no-frills hostel or sharing a room in the area around the convention center - that should be about $550 for 5 nights. About $250, will go towards the Australian visa fee, airport transfers and event registration. Well, if it feels a tad bit pricier, I am okay with that as well.

Who will you need in your team? What have been your successes with projects like this?

Only one person is working on this project. Doing independent research on edge cases of failure modes in ML models.

A few weeks ago I had the opportunity to present my research in a large conference named MICCAI in Strasbourg. My paper was about the learning of medical models from the use of visual shortcuts. Fortunately, I received financial assistance with my travels from Emergent Ventures India. And just before that, I learned that another one of my papers was accepted at NeurIPS - this one is on quantization.

The paper for NeurIPS is in this Google Drive folder along with the acceptance email.

[Google Drive folder](https://drive.google.com/drive/u/2/folders/11aUMaut7oxrEQkzu_5dnM7xsfhWSHQtk)

What are the consequences and most likely causes of failure for this project?

The biggest worry is how long it takes to get an Australian visa. So, I shall apply as soon as the grant is approved so that I do not face any problem. If I don't receive my visa in time, I will present my work online and immediately send all $2,000 back to Manifund.

We don't have to worry that the study isn't finished as it is already finished and peer-reviewed.

How much money have you raised in the past year and where did it come from?
No funding received