You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Exea Labs is a free, volunteer-run AI research collective for high school and early-undergrad researchers, founded April 2026. We're about 230+ researchers across 11+ countries with volunteer regional ambassadors in the US, India, China, and Japan, no paid staff. We recently completed a rigorous shortcut-learning audit on a small (~30M-parameter) vision-language-action robotics policy we built (MicroVLA), diagnosing and fixing memorization failures at four separate layers with pre-registered predictions and statistical testing. This grant funds extending that audit methodology (a probe-plus-behavioral-randomization protocol) across more model classes in our researchers' projects, and turning it into reusable, documented tooling other small labs can run.
Goal: give our researchers (and other small labs) a cheap, ground-truth-free way to check whether a model is actually generalizing or just memorizing a fixed layout/prompt. We already built and validated the core method on MicroVLA: comparing detections across two different instructions on one frame exposes a grounding stage that ignores the instruction; swapping the instruction and rescoring exposes a stage that memorized it. Both need only a second forward pass, no annotation. We'll (1) run this protocol on 3-4 more in-progress researcher projects across different model classes, (2) package the probes and behavioral-randomization harness as a documented, reusable eval/probes.py-style tool, and (3) run a mentor-led workshop teaching ambassadors and researchers to apply it to their own work.
Almost entirely compute and people, not overhead. Our existing AMD Developer Cloud credits (~$1,800, arranged directly with an AMD engineer) are metered and mostly consumed; this grant would go to (1) additional GPU-hours to run the audit protocol across multiple researcher projects instead of just one, and (2) small stipends for mentor reviewers who currently do detailed technical review entirely unpaid. No new hires, no infrastructure spend.
I'm Avneh Bhatia, co-founder (with Joshua Selvaraj). I built and ran the MicroVLA audit myself, plus three independent preprints on parameter-efficient aerodynamic surrogate models (engrXiv, DOIs 10.31224/6870, 10.31224/6319, 10.31224/7097), and I review for Engineering Applications of Artificial Intelligence. I also work part-time as an ML researcher at c/side, a behavioral bot-detection company. Joshua co-leads Exea with me. A small volunteer team (Andre Salehi - outreach, Anthony Nguyen - systems, Devom - technical operations, Arya Addagarla - external affairs, Eknoor Singh - development, Unnat Parekh - regional expansion) runs the rest of the organization, all unpaid.
Most likely failure mode: the audit protocol doesn't generalize cleanly to other model classes (it was built and validated on one VLA policy) and needs real rework before it's reusable by others, costing more mentor time than budgeted. Lower-probability risk: we're a 4-month-old, fully volunteer org, so execution could slip if a couple of the core people get pulled away by school. If it fails, the worst case is we've still improved our own researchers' rigor and produced an honest writeup of what didn't generalize, which is useful information on its own.
No cash funding to date. In-kind only: ~$1,800 in AMD Developer Cloud compute credits (AMD, via a direct engineer contact), a Microsoft Azure nonprofit credit grant, and software/hardware from Siemens and LabJack. Everything else has been Avneh's personal money. This would be our first cash grant.
There are no bids on this project.