[Paper] NeuroClaw Technical Report

Source: arXiv - 2604.24696v1

Overview

NeuroClaw is a multi‑agent AI assistant built specifically for neuroimaging research. By handling raw MRI, fMRI, dMRI, EEG, and related data formats out‑of‑the‑box, it lets scientists focus on the science instead of wrestling with complex pipelines, environment quirks, and reproducibility headaches.

Key Contributions

• Domain‑specialized multi‑agent framework that translates high‑level user intents into concrete neuroimaging tool calls.
• End‑to‑end environment management (pinned Python envs, Docker images, auto‑installers, GPU setup) that guarantees the same software stack across runs.
• Three‑tier skill hierarchy (user interaction → orchestration → low‑level tool skills) for modular, reusable workflow components.
• NeuroBench benchmark that quantifies executability, artifact validity, and reproducibility readiness of neuroimaging pipelines.
• Audit‑ready execution traces with checkpointing and post‑run verification, making pipelines transparent and easier to debug.

Methodology

NeuroClaw treats a neuroimaging project as a stateful graph: raw data → BIDS metadata → a sequence of tool invocations (e.g., FSL, ANTs, FreeSurfer).

1. Skill Layer – Small, atomic agents encapsulate single neuroimaging commands (e.g., “run BET skull‑stripping”).
2. Orchestration Layer – A higher‑level agent composes these skills based on the dataset’s modality and the user’s goal (e.g., “preprocess fMRI”).
3. Interaction Layer – The front‑end chat‑style interface lets developers ask natural‑language questions (“Can you generate a connectivity matrix for subject 01?”).

The system reads the BIDS side‑car JSON files to infer acquisition parameters, automatically selects the appropriate tools, and spins up a Docker container with a reproducible environment. After each step, NeuroClaw writes a structured audit log (command, inputs, outputs, checksum) and validates the produced artifact against NeuroBench’s criteria before moving on.

Results & Findings

• Across three multimodal large language models (LLMs), NeuroClaw‑augmented runs achieved 15‑30 % higher NeuroBench scores than raw LLM prompting, indicating more reliable execution and artifact quality.
• Reproducibility tests (re‑running the same pipeline on a fresh machine) showed identical outputs in 98 % of cases, thanks to pinned environments and deterministic Docker images.
• The checkpointing system reduced debugging time by ~40 %, as developers could resume from the last successful step instead of re‑executing the whole pipeline.

Practical Implications

• Accelerated prototyping – Researchers can spin up a full preprocessing pipeline with a single chat command, cutting weeks of scripting into minutes.
• Consistent CI/CD for neuroimaging – Teams can embed NeuroClaw in automated test suites, ensuring every commit produces reproducible brain maps before merging.
• Lower barrier to entry – New lab members or external collaborators can run complex analyses without deep knowledge of FSL/AFNI/FreeSurfer command‑line intricacies.
• Audit‑ready publications – The generated execution trace satisfies many journal and funding agency reproducibility requirements, simplifying data‑sharing mandates.

Limitations & Future Work

• NeuroClaw currently supports the BIDS‑standardized modalities; exotic or proprietary formats need manual conversion first.
• The benchmark focuses on executability and artifact validity, but scientific validity (e.g., statistical power) is left to the user.
• Scaling to large‑scale cloud clusters and integrating with workflow managers like Airflow or Nextflow is planned for the next release.
• Future research will explore self‑optimizing orchestration, where the system learns to pick the fastest toolchain configuration based on hardware and dataset characteristics.

Authors

• Cheng Wang
• Zhibin He
• Zhihao Peng
• Shengyuan Liu
• Yufan Hu
• Lichao Sun
• Xiang Li
• Yixuan Yuan

Paper Information

• arXiv ID: 2604.24696v1
• Categories: cs.CV
• Published: April 27, 2026
• PDF: Download PDF