tsingyuai/scientify

What it solves

Scientify is an end-to-end autonomous AI research system designed to handle the entire scientific discovery process. It eliminates the need for manual iteration between literature review, hypothesis generation, and experimental validation, allowing the system to autonomously evolve its research direction based on empirical results to produce state-of-the-art (SOTA) outcomes.

How it works

The system employs a multi-agent orchestration framework that manages research hypotheses and accumulated knowledge. It schedules independent agents to handle specific tasks such as implementation, review, and experimentation. A key feature is its "continuous knowledge metabolism" mode, where the system constantly tracks frontier literature and integrates new findings with existing knowledge to refine hypotheses. Each round of experimental results is stored as experience for the next iteration, ensuring the research progresses along the most effective path.

Who it’s for

It is designed for researchers and scientists who want to automate the discovery process, from theoretical derivation and numerical analysis to code implementation and the writing of academic papers.

Highlights

• Autonomous Discovery: Capable of discovering new algorithms (e.g., the KV2 algorithm for LLM inference) and establishing theoretical frameworks in physics (e.g., black hole thermodynamics).
• Knowledge Metabolism: Uses a continuous update loop for knowledge and hypotheses, resulting in a 1.9x higher hypothesis hit rate and 92% lower token costs compared to traditional agent patterns.
• End-to-End Workflow: Manages everything from literature surveys to ablation studies and final paper drafting.
• Cloud-Based Execution: Runs in isolated cloud computers, allowing research tasks to persist and be monitored across different devices.