| Oct 2026 | Released ZAP (arXiv), accepted to CoRL 2026 (Late-Breaking Results). A small open VLM trained only in simulation matches GPT-6 Astra on real-world robot tasks. |
| Mar 2026 | Luma released Uni-1, our unified model for image understanding and generation. I owned temporal reasoning and character consistency in pretraining, and RL for legible text rendering. |
| Dec 2025 | GRPO learns mostly from hard examples: training on the hardest 10% gives up to 47% gains, and only hard examples improve on AIME 2025. Hard Examples Are All You Need at the NeurIPS 2025 Workshop on Efficient Reasoning. |
| Sep 2025 | Joined Luma AI as a research scientist. |
| Jul 2025 | New paper on conditioning diffusion models for synthetic data: Understanding Trade-offs When Conditioning Synthetic Data. |
| Dec 2024 | AbeT, our out-of-distribution detection method, accepted to the NeurIPS 2024 Workshop on Bayesian Decision-making and Uncertainty. |
| Feb 2024 | Presenting our analysis of reward-model robustness under distribution shift at an AAAI 2024 workshop. |
| May 2023 | Presenting calibration for zero-shot vision-language models at the ICLR 2023 Workshop on Trustworthy ML. |