OpenRaiser/NanoResearch

What it solves

NanoResearch automates the entire scientific research process, moving from an initial research idea to a complete LaTeX paper. It solves the problem of "hallucinated" data in AI-written papers by actually executing computational experiments on local GPUs or SLURM clusters, ensuring that every figure, table, and data point in the final paper is derived from real experimental results rather than LLM fabrication.

How it works

The system operates via a 9-stage deep pipeline:

1. Ideation: Searches academic literature (via OpenAlex and Semantic Scholar) to find research gaps and propose hypotheses.
2. Planning: Designs a detailed experimental blueprint including datasets, baselines, and metrics.
3. Setup: Prepares the environment, dependencies, and datasets.
4. Coding: Generates fully runnable experimental code.
5. Execution: Runs the training on local GPUs or SLURM clusters, with automatic debugging and retries if failures occur.
6. Analysis: Parses training logs and metrics to generate structured evidence.
7. Figure Generation: Creates architecture diagrams and result plots based on real data.
8. Writing: Drafts a complete LaTeX paper using the gathered evidence and citations.
9. Review: Automatically reviews and revises the chapters.

It supports multiple interfaces including a Python CLI (with a TUI mode), a Claude Code integration, and a Feishu (Lark) bot. It also features an "Evo" self-evolving pipeline that accumulates skills and project memory across multiple research cycles.

Who it’s for

• Researchers looking to quickly validate prototypes or generate high-quality first drafts of papers.
• AI Engineers wanting to automate the cycle of hypothesis, coding, and GPU execution.
• Academic Teams needing to batch-generate benchmarks or maintain fully traceable research audit trails.

Highlights

• Real-world Execution: Unlike writing tools, it submits actual jobs to GPU clusters and parses real logs.
• Self-Evolving Pipeline: The evo mode uses feedback-driven planning and RAM LoRA adapters to adapt to user preferences over time.
• Flexible Model Routing: Allows different LLMs to be assigned to different stages (e.g., DeepSeek for ideation, GPT-5.5 for coding).
• Multi-Interface Access: Can be driven via CLI, Claude Code (using Bash/WebSearch), or a chat bot.