Automatic Research Kit

Offload the Labour. Steer the Science.

Automate your research and turn your idea into a paper — while you stay in control at every step.

Full Automation, Full Control

Other Tools
  • Autonomous Drift
    Runs unsupervised; no way to course-correct mid-run.
  • Broken Formatting
    Inconsistent layouts and broken LaTeX need manual cleanup.
  • Hallucinated Citations
    LLMs fabricate plausible-looking references to nonexistent papers.
  • Low-Quality Figures
    Default styles, wrong sizes, no awareness of page constraints.
  • Complex Setup
    Requires manual configuration, environment tuning, and dependency wrangling before first run.
ARK
  • Human-in-the-Loop
    Pauses at key decisions; steer via Telegram or web anytime.
  • Publication-Ready Layout
    Hard-coded LaTeX + venue templates (NeurIPS, ACL, IEEE…).
  • Verified References
    Every citation checked against DBLP — no fake references.
  • Precision-Controlled Figures
    Paper Banana + venue-aware canvas, column widths, and fonts.
  • Click to Start
    One-click launch from the web UI — zero configuration needed.

Example Papers

Real papers generated end-to-end by ARK.

MARCO Paper Template: EuroMLSys

MARCO: Budget-Constrained Multi-Modal Research Synthesis via Iterative-Deepening Agentic Search

Read Paper
MMA Paper Template: NeurIPS

A Practical Guide to CPU Matrix Multiplication: From Naive to Efficient

Read Paper

Supported Venues

NeurIPS ICML ICLR AAAI ACL IEEE ACM SIGPLAN LNCS USENIX CVPR ECCV NeurIPS ICML ICLR AAAI ACL IEEE ACM SIGPLAN LNCS USENIX CVPR ECCV NeurIPS ICML ICLR AAAI ACL IEEE ACM SIGPLAN LNCS USENIX CVPR ECCV

ARK Pipeline

Research, develop, and review — iterating until the paper meets your quality target.

1

Research Phase

Deep Research gathers background knowledge and literature survey.

2

Development Phase

Plan experiments, run on compute, analyze results, evaluate completeness, write initial draft.

3

Review Phase

Iterative paper improvement loop until the reviewer score hits the acceptance threshold.

Each iteration runs 5 steps

📄 Compile 🔍 Review 📋 Plan Execute Validate

ARK Agents

Eight specialists collaborating through shared state and structured handoffs.

Reviewer

Scores the paper against venue standards and generates improvement tasks.

Planner

Turns review feedback into a prioritized action plan with experiments and writing tasks.

Writer

Drafts and refines LaTeX sections with DBLP-verified references.

Experimenter

Designs experiments, submits Slurm jobs, and analyzes results.

Researcher

Deep literature survey, identifies gaps, and gathers evidence via academic APIs.

Visualizer

Generates figures with Paper Banana and venue-aware canvas geometry.

Meta-Debugger

Detects stalls, diagnoses failures, and triggers self-repair.

Coder

Writes and debugs experiment code and analysis scripts.

Built by SANDS Lab, KAUST

ARK is developed by SANDS Lab, KAUST, Saudi Arabia.

Visit SANDS Lab, KAUST