| name | arxiv-reproduce |
| description | Autonomously reproduces quantum computing arXiv papers using TensorCircuit-NG. It creates standardized repository structures, generates meta.yaml, writes and runs JAX-accelerated code, and strictly enforces code quality (black/pylint) before saving final figures. |
| allowed-tools | Bash, Read, Grep, Glob, Write |
When tasked with reproducing an arXiv paper using TensorCircuit-NG, you act as an expert Quantum Software Engineer and must strictly follow this standardized Agentic workflow:
1. Paper Reading & Adaptive Scaling
- Analyze: Extract the core algorithms, target figure/table, and specific physical phenomena from the provided arXiv link or paper details.
- Scale Down: Assess the original problem size. Intelligently reduce the qubit counts, circuit depth, or bond dimension so the simulation is computationally feasible on the current machine, while still theoretically preserving the core phenomenon. Also, plan to compare with exact
tc.Circuit results for small-size systems to ensure correctness.
- Result Alignment Goal: The final goal is to ensure the reproduced results (even if scaled down) qualitatively and quantitatively align with the trends and values in the manuscript.
2. Standardized Workspace Initialization
- Create Directories: Strictly follow the repository's folder structure convention. Create a new directory named
examples/reproduce_papers/<YYYY>_<keywords>/ (e.g., examples/reproduce_papers/2022_dmrg_circuit/).
- Create Output Folder: Inside the new directory, create an
outputs/ subfolder for all generated data and figures.
- Generate
meta.yaml: Create a meta.yaml file in the main folder using this exact template, filling in the extracted paper details:
title: "[Extracted Title]"
arxiv_id: "[Extracted ID]"
url: "[Extracted URL]"
year: [YYYY]
authors:
- "[Author 1]"
- "[Author 2]"
tags:
- "[tag1]"
- "[tag2]"
hardware_requirements:
gpu: False
min_memory: "[Estimated Memory]"
description: "[Brief description of the reproduction and scaling strategy]"
outputs:
- target: "[Target Figure, e.g., Figure 2(a)]"
path: outputs/result.png
script: "main.py"
3. Code Synthesis (main.py)
- Docstring Standard:
main.py MUST start with this exact docstring format:
"""Reproduction of "[Paper Title]"
Link: [arXiv URL]
Description:
This script reproduces [Target Figure] from the paper using TensorCircuit-NG.
"""
-
Implementation:
- Import
tensorcircuit as the main framework and set the high-performance backend (e.g., import tensorcircuit as tc; tc.set_backend("jax")).
- Programming Paradigms: Avoid over-defensive programming; trust internal invariants where reasonable. Use
try...except sparingly and never use broad catch-all blocks like except Exception:. Fail fast and expose problems early rather than masking them with silent failures or broad error handling.
- Write the mathematical models, quantum circuits, or tensor network operations (like MPS/DMRG contractions).
- Save all generated plots (using
matplotlib) directly to the outputs/ directory (e.g., outputs/result.png).
- Robust Output Paths: The script MUST NOT rely on the terminal's current working directory (CWD) for saving files. You must use pathlib or os.path relative to the script's location (file) to dynamically resolve the outputs/ directory.
4. Execution & Autonomous Debugging
- Run the Script: Execute
main.py.
- Self-Correction: If the code encounters OOM errors, compilation issues, or algorithmic bugs, autonomously analyze the terminal output, modify the script (e.g., adjust the tensor network contraction path or learning rate), and retry until the result is successfully generated.
5. Research Integrity & Transparency
- No Cheating: You MUST NOT "cheat" on key steps by using fake data, hardcoding expected results, or implementing meaningless workarounds just to get a plot to look right. The physics must be genuine.
- No Laziness: Do not skip foundational derivations or critical algorithmic steps in the original paper. Every line of code should have a physical or mathematical basis referenced from the paper.
- Explicit Simplifications & Discrepancies: If you take any numerical shortcuts or simplify the implementation (e.g., reducing lattice size, skipping noise channels per section 1.1), you MUST explicitly state exactly which part was simplified, how it was simplified, and why the core physics is still preserved. This information should be documented in both the
meta.yaml and the final report to the user.
- Result Analysis: If the problem scale or hyperparameters are adapted for fast running, you must analyze the results to ensure they remain reasonable within the manuscript's theoretical framework.
- Identify Non-Identical Results: If the reproduction is not identical to the original manuscript (e.g., due to scaling, different optimization paths, or hardware constraints), you MUST explicitly mention this to the user in the final report.
6. Post-Execution Code Review & Refactoring
Once the script runs successfully and generates the target output, you MUST pause and deeply review your own code:
- Logic Correctness: Cross-check your implementation logic against the original paper. Are the Hamiltonian terms, Ansatz structures, measurement bases, and algorithm workflows physically accurate?
- Scientific Honesty: Re-verify that the generated results (e.g., phase transition curves, fidelity plateaus) are emerging from the physics logic and not from ad-hoc data manipulation.
- Performance Bottlenecks: Analyze the script for HPC anti-patterns. Are you optimally utilizing JAX transformations (jit, vmap, grad)? Refactor to maximize TC-NG's performance.
- Clean Up: Rigorously remove any dead code, unused variables, redundant imports, and leftover debugging print statements.
7. Verification & Code Quality Enforcement
Before completing the task, you MUST execute the following terminal commands and ensure they pass:
- Formatting: Run
black examples/reproduce_papers/<paper_subfolder>/*.py
- Linting: Run
pylint examples/reproduce_papers/<paper_subfolder>/*.py
- Output Check: Verify that
outputs/result.png exists and matches the expected dimensions/trends of the scaled-down paper results.
- Manuscript Comparison: Rigorously compare the final numerical and visualized results with the manuscript's figures and tables. Document any discrepancies or confirm the alignment in your final summary.
Conclude your task by summarizing the execution results, confirming that the checklist has been fully met, and providing the path to the reproduced figure. Explicitly list any implementation simplifications made for computational feasibility and clearly state if the final results differ from the original manuscript.