EvoScientist: First AI Scientist Framework with Three Co-Evolving Agents

Forum topic · ✨步子哥 · 2026-03-21

Summary

EvoScientist is an AI scientist framework that, according to this Chinese forum post, is the first to implement collaborative evolution among three specialized agents: a Researcher Agent (RA) for scientific idea generation, an Engineer Agent (EA) for executable experiment implementation, and an Evolution Manager Agent (EMA) that distills cross-task experience as a metacognitive layer. The system pairs these agents with dual persistent memory modules—an Ideation Memory (M_I) storing viable and failed research directions, and an Experimentation Memory (M_E) holding data-processing and model-training strategies. Three evolution mechanisms drive learning: IDE (extracting promising directions from successful ideas), IVE (learning to avoid infeasible paths from failures), and ESE (distilling execution strategies from code traces). A skill package system provides pre-written, verified Python modules with metadata files, implementation scripts, and utility code to support code generation and experimentation. The post reports that the framework achieved strong results at the ICAIS 2025 AI Scientist Track, with all six generated papers accepted, winning the Best Paper Award and an AI Reviewer Appreciation Award, and outperforming seven baseline systems on idea novelty, feasibility, relevance, clarity, and code execution success rates. The architecture moves AI-driven scientific discovery from static execution toward dynamic, self-improving evolution.

EvoScientist: The First AI Scientist Framework with Three Co-Evolving Agents

EvoScientist is an AI scientist framework described as the first to implement collaborative evolution of three agents, achieving a paradigm shift from static execution to dynamic, self-improving scientific discovery.

Key points

Three-agent architecture: The framework decomposes scientific discovery into three specialized, complementary, and continuously evolving roles:
Researcher Agent (RA): Generates and iteratively refines scientific ideas using idea-tree search and an Elo tournament mechanism, evaluating novelty, feasibility, and relevance.
Engineer Agent (EA): Converts ideas into executable code via a four-stage experiment-tree search covering data preprocessing, model implementation, and result analysis.
Evolution Manager Agent (EMA): Acts as a metacognitive layer that distills cross-task experience from agent interactions.
Dual persistent memory modules:
Ideation Memory (M_I): Stores summaries of viable research directions and records of failed directions.
Experimentation Memory (M_E): Stores data-processing strategies and model-training strategies (architectures, optimizers, hyperparameter configurations).
Three evolution mechanisms driven by EMA:
IDE (Idea Direction Evolution): Extracts viable research directions from successful ideas.
IVE (Idea Verification Evolution): Learns from failed cases to avoid infeasible paths.
ESE (Experiment Strategy Evolution): Distills execution strategies from code traces.
Closed-loop learning: Feedback from RA and EA execution is converted by the EMA into persistent memory, which in turn improves subsequent task performance—enabling genuine cross-task evolution.

Skill Package System

Unlike the memory modules that provide strategic guidance ("what to do"), skill packages provide concrete implementation ("how to do it"). They are pre-written, verified Python code modules consisting of:

SKILL.md: metadata file describing function, applicable scenarios, specifications, and examples
Main implementation script: complete Python code for a specific task
Auxiliary scripts: utility functions, configuration templates, and test cases
Reference materials

Reported Results

The framework reportedly achieved breakthrough results at the ICAIS 2025 AI Scientist Track:

All six generated papers accepted
Won the Best Paper Award and an AI Reviewer Appreciation Award
Outperformed seven baseline systems across four idea-quality dimensions (novelty, feasibility, relevance, clarity) and in code execution success rate

Outlook

The post argues that EvoScientist demonstrates the potential of self-evolving multi-agent systems in science, pointing toward a paradigm shift in how AI can autonomously conduct research: generating hypotheses, implementing experiments, and continuously improving from accumulated experience across tasks.

> Note: Details above reflect the claims made in the original forum post; independent verification of benchmark results is advised.