Feolu Kolawole
Microsoft · Stanford University · Stanford AI Lab (SAIL)
I'm an undergraduate at Stanford University studying computer science, and a machine learning researcher at the Stanford AI Lab (SAIL) and the Stanford Human Perception Lab. My work spans computer vision, world models, and spatial computing.
I also serve as VP of External Affairs at Stanford XR. My research interests span computer vision, world models, reinforcement learning, and augmented reality / spatial computing.
Reach me at flukol@stanford.edu.
Highlighted work
- 1.Mixture-of-Steering Vectors (MoSV): Sparse Gating for Compositional Hallucination Mitigation
Vedant Srinivas, Daniel Lee, Feolu Kolawole · ICML 2026 Workshop on Mechanistic Interpretability · 2026
A framework that dynamically selects from multiple learned correction vectors per prompt, improving factual accuracy by +2.4pp (vs. +0.3pp for prior methods) and discovering distinct hallucination types without labels across 10,615 items in 8 domains.
- 2.Cross-Game Semantic Alignment of Latent Action Representations
Feolu Kolawole, Khizer Khaderi · IJCAI 2026 Workshop on Generalizing from Limited Resources (GLOW) · 2026
A single inverse dynamics model trained across racing, Atari, and first-person games aligns action embeddings by action class rather than game identity; a 50-label centroid calibration recovers 91.1% accuracy on Pole Position, within 0.3pp of a fully-supervised single-game classifier.
- 3.Eous: Embodied AR Robot Assistant
Winner, Stanford Hackathon · 2026
A hands-free AR system pairing AR glasses, a smartphone, and a Raspberry Pi robot for gesture and voice control with live camera feedback — all on-device with no cloud dependencies.
- 4.DYNAMO: Reinforcement Learning Portfolio Manager
Reinforcement Learning · 2026
An RL agent managing a portfolio across 10 asset classes, achieving 15.76% annualized returns at a 1.44 Sharpe ratio — outperforming traditional strategies by 38–54%.
- 5.Power Lever: GPU-Efficient LLM Inference Gateway
Winner, Stanford Hackathon · 2026
An inference gateway that dynamically routes LLM prompts to right-sized GPU hardware across 4 tiers, cutting energy use by 75% on simple queries while reserving high-end GPUs for complex tasks.