Project summary

I’m a final year CS undergrad at Delhi Technological University, India. My paper, CAMEO: A Black-Box Behavioral Diagnostic for Evaluation-Aware Concealment in Language Models got accepted as a poster at the AI4GOOD workshop at NeurIPS 2026 in Paris this december. NeurIPS gave me a financial assistance award that covers my registration and hotel. I’m raising funds for my flights and visa application.

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

The basic question behind my work is pretty simple. can we actually trust AI safety evals if a model behaves differently when it knows its being tested?

I also wanted to explore this without needing access to a model's internal weights, since thats not always possible with frontier models.

So with CAMEO, we give models the same task but change how its presented to them. Sometimes its framed as an evaluation, sometimes as private deployment, along with two other controls. The idea is to see whether these changes affect their safety behaviour.

I first tested this on fine-tuned models where we already knew what behaviour to expect. CAMEO caught the model trained to conceal its behaviour and didnt flag the controls. Then i tested six frontier and open-weight models. I didnt find evidence of evaluation-specific concealment, but the way a prompt was framed did make a difference to how models responded.

At NeurIPS, i wanna present this work, talk to people working on evaluations and interpretability, hear what they think and where they feel the approach falls short. The next thing i want to look at is whether models can pick up on evaluation cues even when we're not explicitly telling them that theyre being tested.

How will you know it worked?

I get to present the poster in Paris and talk to researchers working on evaluations.

Hopefully i come back with useful feedback, new ideas and a better sense of what to work on next with CAMEO.

How will this funding be used?

• Return flights, Delhi to Paris: $800

• Schengen visa and VFS fees: $130

• Local transport and airport transfers: $70

Total: $1,000

Registration and hotel are already covered by NeurIPS. If my visa is refused or i cant travel, i'll return any unused funds.

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

I'm co-first author on CAMEO with a team of four other researchers. Before this, i co-authored work on neurosymbolic AI, and i've been spending more time working on AI safety and evaluations. CAMEO is one of the projects through which i'm trying to contribute to this area.

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

The main issue would be not being able to travel to Paris. The poster would still exist, but i'd miss the chance to present it in person and talk to people about the work. Since i'm still figuring out the next steps for CAMEO, getting that kind of feedback would be really useful, and its much harder to get these conversations from India.

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

I havent received any cash grants in the last 12 months. NeurIPS awarded me financial assistance covering my registration and hotel, so thats in-kind support rather than a cash grant.