Just Change the URL: alphaXiv’s AutoArxiv Lets You Reproduce Papers on a Single GPU

AI‑assisted paper reproduction is entering a new stage with alphaXiv’s AutoArxiv, an autoresearch service for arXiv papers. Users simply replace the word “arxiv” in a paper’s URL with “autoarxiv”, and the system automatically handles dependency resolution, environment configuration, and minimal‑reproduction execution while also estimating the resources required for a full replication.

The workflow, demonstrated in an official video, begins by extracting the paper’s arXiv ID, searching GitHub for the corresponding open‑source code repository, and importing it into an AI‑driven development environment. When the user asks the agent to produce a runnable minimal version, the agent clones the repo, reads the README, inspects the project structure, and analyzes the original experiment’s requirements.

The agent discovers that the original experiment for the selected paper (e.g., the 2017 "Attention Is All You Need" model) needs four H100 GPUs, runs for about 15 minutes, and performs 100 training steps, with a hard‑coded dataset path. To make the experiment fit a single GPU, the agent devises a "minimal reproduction" plan that includes:

Model replacement: swapping the large base model for a lightweight alternative.

Parameter compression: limiting training to 40 steps and checkpointing every 20 steps.

Resource optimization: setting num_processes to 1, disabling multi‑GPU accelerators such as DeepSpeed, and enabling LoRA training to save memory.

Script generation: automatically creating or modifying run.sh and a log‑analysis script summarize_eval.py.

The video shows a side‑by‑side diff of the original and modified scripts, highlighting the concrete code changes made by the agent. This demonstrates that AutoArxiv’s AI agent does more than merely read a paper; it acts as a practical reproduction assistant that adapts complex training code to the user’s available compute.

For researchers and engineering teams, this capability is valuable because it allows a quick sanity check—determining whether the code can run at all and estimating the compute budget before committing to a full replication. In the hands‑‑on test, the system successfully linked the "Attention Is All You Need" paper to its TensorFlow repository, prompted the user to confirm the import, and proceeded with the minimal‑reproduction steps.

AlphaXiv clarifies that users may run the agent on their own compute resources, preserving flexibility for labs and enterprises that wish to integrate private hardware or codebases. Overall, AutoArxiv lowers the entry barrier for paper reproduction, enabling both casual readers and professional teams to experiment with state‑of‑the‑art research without needing to provision multiple high‑end GPUs.