EvoScientist: Huawei's Self-Evolving Multi-Agent AI Scientist That Learns From Experience

Forum topic · 小凯 · 2026-03-20

Summary

EvoScientist, developed by a Huawei research team, is a multi-agent AI scientist system designed for end-to-end scientific discovery that, unlike prior systems, accumulates and reuses experience across tasks. The architecture combines three agents—Researcher (idea generation via tree search and Elo tournament selection), Engineer (code generation and experimentation across four phases), and an Evolution Manager that distills lessons after each task into two persistent memory stores: Ideation Memory and Experimentation Memory. Three evolution mechanisms (idea direction, idea validation, and experiment strategy evolution) drive continuous improvement. Benchmarks show win-rate advantages of 29%–93% over seven baseline systems, human expert evaluations favoring EvoScientist on novelty (82.5% win rate), and average code execution success improving from 34.39% to 44.56% after evolution. Notably, six autonomously generated papers were accepted at ICAIS 2025, including a Best Paper Award. The paper is available at arXiv:2603.08127 with code on GitHub. Remaining limitations include a low 21.57% success rate for implementing novel methods and dependence on external LLM APIs.

EvoScientist is a self-evolving multi-agent AI scientist system proposed by researchers at Huawei (Yougang Lyu, Xi Zhang, Xinhao Yi, et al.), aimed at end-to-end scientific discovery. Unlike existing AI scientist systems that start from scratch on every run, EvoScientist accumulates reusable knowledge across tasks.

Paper: arXiv:2603.08127
Code: github.com/EvoScientist/EvoScientist

Why AI scientists need memory

Prior systems—Sakana's The AI Scientist, Google's AI Co-Scientist, AI-Researcher, InternAgent—are effectively stateless: they forget why past experiments failed and repeat the same mistakes, wasting compute on dead-end directions. EvoScientist's goal is to give AI something like a human researcher's accumulated intuition.

Architecture: three agents + two memories

Researcher Agent (RA): literature search (Semantic Scholar API), idea generation, tree-based propose-review-refine search, Elo tournament selection, and proposal writing.
Engineer Agent (EA): retrieves successful code strategies from memory, then runs tree search through four experiment phases (initial implementation, hyperparameter tuning, proposed method, ablation), executing code in a sandbox.
Evolution Manager Agent (EMA): after each task, analyzes the full interaction history and updates the memories.

The two persistent stores are:

Ideation Memory: promising research directions plus lessons from failed ideas.
Experimentation Memory: effective data processing, training tricks, and debugging experience.

Three evolution mechanisms

1. Idea direction evolution: top-3 high-Elo ideas per task are summarized into reusable direction insights. 2. Idea validation evolution: failed implementations and underperforming results are analyzed and recorded as lessons. 3. Experiment strategy evolution: successful and failed code attempts are distilled into implementation strategies.

Results

Idea quality: win-rate advantages over seven baselines ranging from +29.17% (Virtual Scientist) to +93.34% (InternAgent); strongest on novelty and feasibility. Human evaluation by three PhD-level experts gave win rates of 82.50% (novelty), 72.50% (feasibility), 62.50% (relevance), 75.00% (clarity).
Code execution: average success rate rose from 34.39% to 44.56% after evolution (+10.17%); Phase 4 (ablation) improved most (+16.82%), while Phase 3 (implementing the proposed method) remains hardest at 21.57%.
End-to-end: 6 autonomously generated papers submitted to ICAIS 2025 (AI Scientist Track) were all accepted (vs. a 31.71% overall acceptance rate), with one Best Paper Award and one AI Reviewer's Appraisal Award.

Comparison with The AI Scientist

The AI Scientist performs agentic tree search within a single task but discards intermediate results afterward. EvoScientist's Evolution Manager carries knowledge across tasks—the difference between a student re-researching everything each time and a scientist consulting a research log.

Limitations and future work

Low success rate (21.57%) for implementing novel methods.
Dependence on external APIs (Semantic Scholar, multiple LLM providers, Ollama embeddings).
Validated mainly in machine learning; physical-science domains would require significant adaptation.
Future directions include structured knowledge-graph memory, cross-agent co-evolution, and human-in-the-loop memory review.

Core insight: the bottleneck for AI scientists is not model capability but memory—intelligence requires the ability to learn from the past.