Chain of Operators: An Inference-Time Harness for In-Context Operator Learning

Paper: arXiv:2606.12318

This repository contains the evaluation code, data-generation utilities, checkpoints, and saved results used for ICON-CHOP experiments.

• conservation laws;
• mean-field control (MFC); and
• graph-in-context air-quality forecasting (GICON).

Repository layout

Directory	Contents
ckpt/	Checkpoints used by the evaluation entry points
data/	Data-generation utilities and dataset instructions
eval/	Reproducible evaluation entry points
conservation_law/	Conservation-law ICON-CHOP code and saved results
mean_field_control/	MFC ICON-CHOP code, prompt examples, and saved results
graph_in_context/	GICON-CHOP code and saved results
cross_transfer/	Cross-task MFC-to-WENO evaluation results

The READMEs inside these directories document task-specific protocols, file formats, and result layouts.

Environment setup

The project requires Python 3.12 and uses uv to manage the environment and install the dependencies recorded in uv.lock. From the repository root, run:

uv sync

The evaluation entry points should be run from the repository root. They use the shared environment created by uv.

Data

Conservation-law and MFC generation scripts are included under data/:

• data/conservation_law/README.md
• data/mean_filed/README.md

The mean_filed directory name is retained for compatibility. Air-quality data are distributed separately; see data/air/README.md for the download location.

Evaluation

Evaluation entry points are organized by task under eval/:

• eval/conservation_law/ for autoregressive conservation-law evaluation and cross-task transfer;
• eval/mean_field_control/ for MFC out-of-distribution evaluation and ablations; and
• eval/graph_in_context/ for cross-region air-quality evaluation.

The Bash wrappers expose task-specific data, checkpoint, and evaluation options. See eval/README.md and the README in each task directory for commands and protocol details.