Ptychographic imaging reconstructs an object from overlapping diffraction patterns — a computational inverse problem where the algorithm matters as much as the data. Standard algorithms use hand-designed regularization terms chosen by domain experts. What if the regularization itself were discovered automatically?
Note (2026-11-13; read against the paper's full text): the comparison in the paper is against reconstruction without regularization, not against manually designed regularizers. It says that for all datasets it compares against 'reconstruction without regularization' as implemented in the Pty-Chi library, the least-squares maximum likelihood algorithm without explicit priors beyond constraints inherent to it, and reports the improvements relative to this baseline; the abstract's wording is 'conventional reconstructions'. So the essay's 'outperform manually designed ones by up to +0.26 SSIM and +8.3 dB PSNR' does not match what is compared: the paper's table gives improvements of +0.12 to +0.26 SSIM across its three datasets (an X-ray integrated circuit, low-dose electron microscopy of apoferritin, and multislice imaging with crosstalk artifacts), and the 'up to' figures are the best of those, for the best discovered regularizer, against no regularizer. The evaluation pipeline in the paper supports three modes (ground-truth metrics, human expert feedback, vision-language model assessment) so the essay's 'evaluation on real data provides fitness' is one of them at most; the discovery-evaluation score is taken at a fixed iteration horizon. The paper also describes the method as using the LLM's understanding of algorithmic function to guide crossover and mutation, and records lineage; the essay's claim that the bottleneck has shifted from computation to algorithm design, and that no understanding of the physics is needed, is the essay's.
An LLM-driven evolutionary framework generates, evaluates, and refines regularization algorithms for ptychography. The LLM proposes candidate algorithms; evaluation on real data provides fitness; evolution selects and recombines the best. The discovered algorithms outperform manually designed ones by up to +0.26 SSIM and +8.3 dB PSNR across X-ray and electron microscopy datasets.
The structural insight: the LLM is not solving the imaging problem. It is solving the algorithm design problem — a meta-problem where the search space is the space of regularization strategies rather than the space of images. The evolutionary wrapper provides the selection pressure that the LLM's single-shot generation cannot: it retains algorithmic components that work and discards those that don't, accumulating design knowledge across generations that no single prompt could elicit.
The through-claim: the bottleneck in computational imaging has shifted from computation to algorithm design. Hardware is fast enough. Data is abundant. What limits reconstruction quality is the human designer's ability to anticipate which regularization assumptions match the physics of the specific imaging modality. Automating the design search doesn't require understanding the physics — it requires searching the space of assumptions and letting the data adjudicate. The LLM provides the generative capacity; evolution provides the judgment.