Run records for "Prior or Feedback? What an LLM Uses When Adapting Neural Operators"

Recorded runs behind the reported results of the NeurIPS 2026 AI for Science workshop paper Prior or Feedback? What an LLM Uses When Adapting Neural Operators. An LLM (DeepSeek, model deepseek-v4-pro) selects fine-tuning configurations for pretrained Fourier Neural Operators on PDEBench transfer tasks under a fixed trial budget, compared with random search and TPE.

Contents: per-run search histories, LLM proposals and reasoning, validation and held-out test metrics, and resolved run configurations for the main panels and the TPE startup control; and decision records of the cold-start description intervention and the validation-score reassignment replays. Reasoning text and proposals are AI-generated model outputs. Absolute paths were replaced by <repo>; no other content was changed. MANIFEST.json lists the SHA-256 checksum of every copied data file and maps each run panel and campaign to the paper claims it supports.

The paper's adaptation and intervention results can be recomputed from these records with paper/verify_claims.py in the accompanying code repository. The FNO training code is not included. The pretrained FNO checkpoints and datasets are from PDEBench (Takamoto et al., 2022; DaRUS doi:10.18419/darus-2986 and doi:10.18419/darus-2987, CC BY 4.0) and are not included.