SciAgent: A Unified Multi-Agent System for Generalistic Scientific Reasoning

Xuchen Li, Ruitao Wu, Xuanbo Liu, Xukai Wang, Jinbo Hu, Zhixin Bai, Bohan Zeng, Hao Liang, Leheng Chen, Mingrui Chen, Haitian Zhong, Xuanlin Yang, Xu-Yao Zhang, Jia Li, Kaiqi Huang, Jiahao Xu, Haitao Mi, Wentao Zhang, Bin DongarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces SciAgent, a novel multi-agent AI system designed for "generalistic scientific reasoning" – the ability to tackle complex problems across various scientific domains and difficulty levels without domain-specific redesign. SciAgent achieves gold-medalist performance in math and physics Olympiads, demonstrating a significant step towards AI systems that can reason coherently across disciplines.

What Problem Does This Paper Address?

Current AI systems excel at specific scientific tasks but struggle to generalize. They are often "narrow" and require extensive handcrafted pipelines for each new domain or problem type. This limits their ability to truly mimic human scientific intelligence, which involves flexible adaptation of reasoning strategies across diverse fields. The authors aim to bridge this gap by developing an AI that can autonomously select, compose, and adapt reasoning procedures for any given scientific problem, moving beyond domain-bounded reasoning.

Methods

SciAgent employs a hierarchical multi-agent architecture:

  • Coordinator Agent: Acts as a meta-reasoner, analyzing the problem's domain, modality, and difficulty to route it to the appropriate Worker System.
  • Worker Systems: Domain-specific multi-agent systems (e.g., for Math, Physics, Chemistry) that encapsulate specialized reasoning paradigms.
  • Sub-agents: Within each Worker, these agents perform specific tasks like symbolic deduction, conceptual modeling, numerical computation, and verification.

These agents collaborate through adaptive, self-assembling reasoning pipelines. This means the system dynamically constructs and refines solution strategies based on the problem's demands, rather than relying on fixed templates. The architecture emphasizes hierarchical meta-reasoning, where the Coordinator plans and delegates, Workers specialize and adapt, and Sub-agents execute and verify.

Datasets

The paper evaluates SciAgent on several challenging benchmarks:

  • Mathematics Olympiads: International Mathematical Olympiad (IMO) 2025, International Mathematics Competition (IMC) 2025.
  • Physics Olympiads: International Physics Olympiad (IPhO) 2024, 2025, Chinese Physics Olympiad (CPhO) 2025.
  • Chemistry Olympiads: International Chemistry Olympiad (IChO) 2025 (though full results are limited due to data availability).
  • General Scientific Reasoning: Humanity's Last Exam (HLE) benchmark (selected math, physics, chemistry, and biology problems).

Key Insights and Ideas

  • Generalistic Scientific Reasoning: The paper proposes this as a new paradigm for AI in science, shifting focus from narrow specialization to adaptive, cross-disciplinary problem-solving.
  • Hierarchical Meta-Reasoning: The Coordinator-Worker-Sub-agent structure is crucial. It allows for explicit reasoning about reasoning, enabling intelligent task delegation and strategy selection.
  • Adaptive Pipeline Assembly: The ability to dynamically construct reasoning pipelines, rather than using static ones, is key to SciAgent's flexibility and generalization.
  • Beyond LLM Scaling: The authors argue that achieving true generalistic scientific intelligence requires more than just scaling up large language models; it necessitates sophisticated architectures that orchestrate reasoning as a dynamic social process among specialized agents.

Key Results

SciAgent consistently achieves gold-medalist performance across multiple mathematics and physics Olympiad benchmarks:

  • IMO 2025: Achieved a total score of 36/42, surpassing the average gold medalist score.
  • IMC 2025: Achieved a perfect score of 100, matching the highest human performance.
  • IPhO 2025: Achieved a theoretical score of 25.0/30.0, exceeding the average gold medalist score.
  • IPhO 2024: Attained 27.6/30.0, outperforming the average gold medalist score.
  • CPhO 2025: Achieved a total of 264/320, significantly surpassing the human gold medalist record.

The system also demonstrated strong generalization on the HLE benchmark, tackling problems that require adapting reasoning strategies across different scientific fields.

Limitations and Future Work

  • Data Availability for Chemistry/Biology Olympiads: Full quantitative comparisons for IChO and IBO were not possible due to data limitations (e.g., lack of publicly released individual scores for IChO, restricted access to IBO problems).
  • Chemistry Agent Development: While showing potential, the Chemistry Agent System is still under development and has not yet reached gold-medalist performance levels.
  • Future Work: The authors propose expanding SciAgent to more scientific domains (chemistry, biology, interdisciplinary sciences), incorporating multimodal reasoning (text, images, data), developing collaborative and self-evolving agents, and integrating SciAgent into real-world scientific workflows for hypothesis generation, simulation verification, and analytical modeling. The ultimate goal is to move towards AI systems that can contribute original scientific insight, not just reproduce expert reasoning.

Summary updated

Abstract

Recent advances in large language models have enabled AI systems to achieve expert-level performance on domain-specific scientific tasks, yet these systems remain narrow and handcrafted. We introduce SciAgent, a unified multi-agent system designed for generalistic scientific reasoning-the ability to adapt reasoning strategies across disciplines and difficulty levels. SciAgent organizes problem solving as a hierarchical process: a Coordinator Agent interprets each problem's domain and complexity, dynamically orchestrating specialized Worker Systems, each composed of interacting reasoning Sub-agents for symbolic deduction, conceptual modeling, numerical computation, and verification. These agents collaboratively assemble and refine reasoning pipelines tailored to each task. Across mathematics and physics Olympiads (IMO, IMC, IPhO, CPhO), SciAgent consistently attains or surpasses human gold-medalist performance, demonstrating both domain generality and reasoning adaptability. Additionally, SciAgent has been tested on the International Chemistry Olympiad (IChO) and selected problems from the Humanity's Last Exam (HLE) benchmark, further confirming the system's ability to generalize across diverse scientific domains. This work establishes SciAgent as a concrete step toward generalistic scientific intelligence-AI systems capable of coherent, cross-disciplinary reasoning at expert levels.
BibTeX
@misc{li2025sciagent,
  author = {Xuchen Li and Ruitao Wu and Xuanbo Liu and Xukai Wang and Jinbo Hu and Zhixin Bai and Bohan Zeng and Hao Liang and Leheng Chen and Mingrui Chen and Haitian Zhong and Xuanlin Yang and Xu-Yao Zhang and Liu Liu and Jia Li and Kaiqi Huang and Jiahao Xu and Haitao Mi and Wentao Zhang and Bin Dong},
  title = {{SciAgent: A Unified Multi-Agent System for Generalistic Scientific Reasoning}},
  year = {2025},
  eprint = {2511.08151},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.08151},
}
Figure