An LLM agent that writes your ML paper and code

What it does AI-Researcher is a NeurIPS 2025 Spotlight project that packages an LLM agent inside a Docker container to automate machine learning research. You give it either a detailed research idea or a set of reference papers, and it attempts to chain together literature review, algorithm design, code implementation, validation, and manuscript writing into a single pipeline.

The interesting bit Rather than treating LLMs as chatbots, the system treats them as assembly-line workers in a containerized research factory, with distinct stages for ideation, coding, and LaTeX generation. It even includes a benchmark suite and leaderboard to judge whether the resulting papers and code actually pass muster.

Key highlights

• NeurIPS 2025 Spotlight project with a hosted demo at novix.science/chat
• Two input modes: detailed idea prompts or reference-based “come up with something novel”
• End-to-end pipeline covers lit review, code implementation, experiment validation, and paper drafting
• Web GUI built on Gradio for configuring and monitoring runs
• Benchmark suite evaluates outputs across specific ML domains like GNNs, VQ-VAE, and flow matching

Caveats

• Requires external LLM API keys (OpenRouter, GitHub AI token) and GPU Docker setup; not a local, offline tool
• README leans heavily on marketing language and emoji, while remaining vague on actual paper quality or reproducibility rates
• Domain coverage appears narrow, focused on a handful of ML subfields rather than general science

Verdict Useful if you are experimenting with agentic research automation or need a baseline for ML paper-generation systems. Avoid if you want a general-purpose, offline research assistant that works out of the box.