You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$1,000 left to reach the funding goal.
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 am raising funds for my flights and visa application.
One question behind my work on CAMEO is whether we can really trust safety evaluations if a model behaves differently when it knows it is being tested.
My project mainly aims to answer whether we can trust AI safety evals if a model behaves differently when it knows it is being tested. And black-box motivation was because not everyone has access to the open weights of frontier models.
To explore this, I tested models on the same task but changed how the situation was presented. Sometimes it was framed as an evaluation and sometimes as private deployment. The idea was to see whether these changes affected safety behaviour.
I tested six frontier and open-weight models. I didn’t find evidence that they were specifically concealing their behaviour during evaluations, but I did find that the way a prompt was framed could make a real difference.
At NeurIPS, I wanna present this work, hear what other researchers think, and learn where the approach falls short. I’m esp interested in feedback from people working on evaluations and interpretability, as I work towards the next version of CAMEO.
I present the poster in Paris and talk with researchers working on evaluations.
I come back with concrete feedback that shapes the next version of my research.
Airfare (~800 USD), visa+VFS (130 USD), local transport and commute (~100 USD for 6 days)
I’m co-first author on CAMEO with a team of 4 other researchers. Before this, I co-authored projects on neurosymbolic AI, and I am actively working on the AI safety and evaluations domain.
I can’t attend in person and the poster goes without its presenter. I’d lose the chance for direct feedback and connections, which are hard to get from India.
No cash grants. NeurIPS awarded me with Financial Assistance Award, which is in-kind support.