DeepScientist

DeepScientist addresses the challenge of automating frontier scientific discovery, enabling AI to progressively surpass human state-of-the-art (SOTA) performance. Targeting researchers and AI practitioners, it offers a pathway to accelerate breakthroughs across diverse scientific domains through autonomous, iterative research, aiming to make AI a true partner in scientific advancement.

How It Works

The system autonomously executes the scientific method, performing goal-oriented, continuous, and iterative discovery without human intervention. It generates novel research ideas, implements promising hypotheses, and employs structured causal reasoning, exemplified by the A2P (Abduction-Action-Prediction) method for "Agent Failure Attribution," to achieve significant advancements beyond pattern recognition. This iterative process allows for rapid exploration and refinement of scientific concepts.

Quick Start & Requirements

Access is currently phased. Phase 1 involves application-based access via a waitlist, with collaboration to refine the system. Phase 2, "Foundational Components Release," is complete, offering a website (http://deepscientist.cc) and CLI to the first 30 invited users, with plans for broader open-sourcing of foundational components. Requires Python 3.8+.

Highlighted Details

• Achieved 3 years of human research progress in 2 weeks on AI text detection, increasing AUROC by 7.9% and reducing inference latency on the RAID dataset.
• Developed the A2P method for "Agent Failure Attribution," yielding a 183.7% improvement over human SOTA on the Who&When benchmark ("algorithm-generated" setting).
• Autonomously generated 2,472 unique research ideas and implemented 600 hypotheses.
• Demonstrated a near-linear relationship between computational resources and research output, though a bottleneck in "exploration efficiency" is anticipated.

Maintenance & Community

Development follows a phased open-source plan, with foundational components released and experimental data/full source code planned. A discussion group is available via WeChat invitation.

Licensing & Compatibility

No explicit license is currently stated. The project is undergoing a phased release, with foundational components available to invited users and full source code planned for later release. Compatibility for commercial use is not yet defined.

Limitations & Caveats

Immediate adoption is restricted due to the phased release model and safety considerations. Foundational components may contain unpolished code and design. The project acknowledges the current state of AI scientists as potentially "high-throughput trial-and-error machines" rather than intuitive discoverers, and highlights the high cost and low efficiency of current RLHF training methods.