EvoScientist: The Self-Evolving AI Scientist That Ranked #1 on DeepResearch Bench

What is EvoScientist?

EvoScientist is a self-evolving, multi-agent AI research system designed for end-to-end scientific discovery. Rather than replacing researchers, it operates as a "human-on-the-loop" research buddy that co-evolves with you — handling the tedious parts while you steer the direction.

The project ranked #1 on DeepResearch Bench II in March 2026 and won the Best Paper Award at ICAIS 2025, with all 6 AI-generated papers accepted in the AI Scientist Track.

EvoScientist DeepResearch Bench #1

The 6-Agent Team

EvoScientist employs six specialized sub-agents working in concert:

Agent	Role
Plan	Breaks down research goals into structured tasks
Research	Scans, summarizes, and synthesizes literature
Code	Implements experiments, data processing, and analysis
Debug	Diagnoses and fixes issues in code and pipelines
Analyze	Interprets results, generates visualizations, draws conclusions
Write	Produces publication-ready manuscripts in LaTeX/Markdown

Together they form a complete scientific workflow: intake → plan → execute → evaluate → write → verify.

Key Features

EvoMemory: Persistent memory across sessions — context, preferences, and findings carry over
Multi-Provider LLM: Switch between Anthropic, OpenAI, Google, and NVIDIA models via one config
200+ Built-in Skills: Predefined research skills, plus install more from GitHub on the fly
MCP Integration: Connect to any MCP server for additional tools
Multi-Channel: CLI/TUI as the hub, with Telegram, Slack, Feishu, WeChat, Discord, and QQ integrations

Quick Start

Installation

EvoScientist requires Python 3.11–3.13. The recommended way to install is via uv:

# Quick install as a CLI tool
uv tool install EvoScientist

# Or install into your current environment
uv pip install EvoScientist

Optional channel extras:

uv pip install "EvoScientist[telegram]"
uv pip install "EvoScientist[slack]"
uv pip install "EvoScientist[all-channels]"

Onboarding

Run the interactive configuration wizard:

EvoSci onboard

EvoScientist onboarding wizard

This sets up your LLM provider, API keys, and preferences. For Claude Code or Codex CLI users, it supports OAuth authentication.

Using the CLI

Once configured, start a research session:

EvoSci

EvoScientist CLI interface

The TUI provides a rich interactive interface for managing research tasks, viewing agent activity, and reviewing outputs.

MCP Integration

Add MCP servers for extended capabilities:

EvoSci mcp add arxiv-search npx -y @arxiv/mcp-server

Channel Setup

Connect to messaging platforms for remote research collaboration:

EvoSci channel setup telegram
EvoSci channel setup slack
EvoSci channel setup feishu

Example Workflow

Here is a typical end-to-end research session:

1. Define the Research Question

Research the latest advances in protein language models for drug discovery.
Focus on papers from 2025-2026. I want a comprehensive literature review.

2. The Plan Agent breaks this into tasks: identify key papers, categorize approaches, compare architectures, summarize findings.

3. The Research Agent searches arXiv, PubMed, and Semantic Scholar. It reads, summarizes, and cross-references 50+ papers.

4. The Code Agent builds comparison tables, generates citation networks, and creates benchmark visualizations.

5. The Analyze Agent identifies trends, gaps in the literature, and potential research directions.

6. The Write Agent produces a structured literature review manuscript with proper citations, figures, and a LaTeX-formatted bibliography.

Throughout the process, you can intervene at any checkpoint — approve, redirect, or ask for changes.

Awards and Recognition

#1 on DeepResearch Bench II (March 2026)
Best Paper Award at ICAIS 2025
AI Reviewer's Appraisal Award at ICAIS 2025
6/6 AI-generated papers accepted at ICAIS 2025 AI Scientist Track
Technical report: arXiv:2603.08127