LLM Agent Autonomously Conducts CT Reconstruction Research, Matching SOTA with 969 Parameters

maier_ak · x · 2026-07-31

An LLM-driven agent successfully ran a full autonomous research loop for CT image reconstruction. Operating within a fixed compute budget, the agent wrote code, tuned hyperparameters, and benchmarked 26 state-of-the-art methods on a dedicated cluster, iterating based on a calibrated headroom score.

The experiment yielded a highly compact 969-parameter CT solver that tied for top performance on the low-dose Mayo dataset, matching much larger SOTA models. Furthermore, the research exposed the fragility of ideal-data leaderboards: when realistic noise was added to the clean DL-Sparse-View breast benchmark, top-performing clean-data models dropped to near-zero improvement, while a learned primal-dual method proved robust, achieving a headroom score of 0.93.

Original post →

More from coding & agent

coding & agent channel →