This paper builds an LLM agent that edits reconstruction solvers, runs short cluster jobs, reads a frozen metric, and iterates. The agent implemented, tuned, and benchmarked 26 CT reconstruction methods. On a noiseless sparse-view breast task, the supervised image denoiser ranked first with a headroom score of 0.89, but collapsed to 0.00 when evaluated with noise. A learned primal-dual method rose from 0.72 to 0.93. A compact 969-parameter solver tied the top Mayo low-dose CT tier at the 1% level, using 0.4% of the champion's parameters.
No heat snapshots are available in the last 24 hours.