Overview
EvoScientist is described as the first AI scientist framework to implement three-agent co-evolution. It decomposes scientific discovery into three specialized, complementary, and continuously evolving roles, paired with a persistent dual-memory design, achieving a paradigm shift from static execution to dynamic evolution.
Key Breakthrough
At the ICAIS 2025 AI Scientist Track, all six papers generated by the framework were accepted, winning the Best Paper Award and an AI Reviewer Appreciation Award. The system reportedly outperformed 7 baseline systems on four idea-generation metrics (novelty, feasibility, relevance, clarity) and on code execution success rate.
1. Multi-Agent Architecture
The framework's core innovation is EMA-driven evolution, enabling continuous distillation of strategy knowledge from interaction history for cross-task self-improvement.
Researcher Agent (RA)
- Generates and iteratively refines scientific ideas via idea-tree search and an Elo tournament mechanism
- Evaluates novelty, feasibility, and relevance
- Translates ideas into executable code through a four-stage experiment-tree search
- Handles data preprocessing, model implementation, and result analysis
- Meta-cognitive layer distilling cross-task experience via three mechanisms:
- IDE (Ideation Direction Evolution)
- IVE (Ideation Validation Evolution)
- ESE (Experiment Strategy Evolution)
- Ideation Memory (M_I): supports idea generation; stores summaries of viable directions (from IDE) and records of failed directions (from IVE)
- Experimentation Memory (M_E): supports experiment execution; stores data-processing strategies (cleaning, transformation, feature engineering) and model-training strategies (architecture, optimization, hyperparameters)
SKILL.md— metadata (function, use cases, conventions, examples)- Main implementation script — the core Python implementation
- Auxiliary scripts — utility functions, configuration templates, test cases
- Reference materials
Engineer Agent (EA)
Evolution Manager Agent (EMA)
Core insight: RA and EA execution feedback is converted by EMA into persistent memory, which in turn improves future task quality — forming a closed-loop learning system with true cross-task evolution capability.
2. Persistent Memory Modules
The design reflects the highly structured, multi-dimensional, evolving nature of scientific knowledge, which prompt engineering or in-context learning alone cannot capture effectively.
3. Skill Packages
Skill Packages provide the concrete "how-to" counterpart to memory's "what-to-do" guidance — pre-written, verified Python code modules for code generation and experiment execution. Each package contains:
Significance
The post positions EvoScientist as evidence of AI's growing potential in scientific research, illustrating a shift toward self-improving, evolving multi-agent systems for autonomous scientific discovery.
*Note: the source is a Chinese forum post; details are reported as claimed by the original author and have not been independently verified.*