Kalman Evolve: LLM-Driven Algorithm Discovery for State Estimation
Researchers propose Kalman Evolve, a framework that uses large language models (LLMs) to discover improved filtering algorithms for state estimation. The Kalman Filter is optimal under linear dynamics, Gaussian noise, and known covariances, but fails in realistic sensing scenarios like Doppler radar and LiDAR. Kalman Evolve jointly optimizes noise parameters and the update structure, leveraging LLMs as a structured prior over program space to generate interpretable, non-affine modifications. This approach closes the performance gap that cannot be eliminated by tuning noise covariances alone.
Key facts
- Kalman Filter is optimal under linear dynamics, Gaussian noise, and known covariances.
- Realistic sensing settings like Doppler radar and LiDAR violate these assumptions.
- Kalman Evolve jointly optimizes noise parameters and update structure.
- The framework uses LLMs as a structured prior over program space.
- Kalman Evolve generates interpretable, non-affine modifications to the classical Kalman Filter.
- The approach addresses performance degradation in nonlinear estimation.
- The paper is available on arXiv with ID 2605.26830.
- The research is categorized under cross-type announcement.
Entities
Institutions
- arXiv