You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
The current way we test multimodal agents for hallucinations is broken. Most setups just slap an external "checker" model on top of a "doer" model to catch errors. My tests show this approach fails catastrophically under stress. And also, everyone tests their models using single-seed evaluation runs. These single runs, most of the time, completely hide when a model is reward-hacking or quietly failing.
I am writing code to inject a Hybrid Reward Architecture (HRA) directly into LLaVA-1.5-7b's reinforcement learning loop. This anchors the model's factual behavior from within the core architecture, rather than leaning on brittle external wrappers. I am also building automated, multi-seed evaluation pipelines using PPO-Clip to expose exactly how brittle external patches are compared to internal fixes.
I want to show the alignment community why single-seed tests and external guardrails are a massive risk. To do that, I am setting up headless environments to run heavy multi-seed benchmarks on LLaVA across ScienceQA and VQA-v2. I will open-source the evaluation scripts so anyone can track variance and agent deception, document the internal HRA framework, and package the final data into a technical paper targeting ICLR 2027
If I get the $5,000 minimum, it goes entirely toward renting A100 instances on RunPod and paying for baseline API queries against closed models. If I hit the $15,000 max, I can use the rest as a living stipend to focus on this full-time without distractions and cover travel to present the paper.
Who is on your team? What's your track record on similar projects?
Just me. I am an independent researcher doing my CS Master's at Nile University in Nigeria, with a background in Electrical Engineering.
I know how to build this because I have done it. I just had a paper accepted at the ICML 2026 AgenticUQ workshop focusing on how single-seed training masks major model failures. I live in PyTorch, Python, and LangChain. Before this, I worked as a remote PM, so I actually know how to manage a technical pipeline, hit deadlines, and keep things efficient.
i.e., my project supervisor is heavily involved and a mentor.
The biggest risk is running out of money for compute before I finish the 20-seed runs, which would ruin the statistical significance of the data. There's also a chance that modifying LLaVA’s internal RL loop causes training instability early on.
If the internal HRA code completely fails to stabilize, the backup plan is still highly valuable: I will still open-source the multi-seed evaluation data to map out exactly where the external checker models broke down, giving the safety community hard data on benchmark vulnerabilities.
How much money have you raised in the last 12 months, and from where?
NO