TLDR: SciAgent is a new multi-agent AI system designed for “generalistic scientific reasoning,” meaning it can adapt its problem-solving strategies across different scientific disciplines and difficulty levels. Unlike specialized AI, SciAgent uses a hierarchical structure with a Coordinator Agent, specialized Worker Systems (e.g., Math, Physics), and Sub-agents that collaborate to dynamically solve complex problems. It has achieved gold-medal performance in major mathematics and physics Olympiads and shows strong generalization across other scientific tasks, marking a significant step towards AI capable of coherent, cross-disciplinary scientific intelligence.
Recent advancements in large language models have empowered AI systems to tackle complex scientific problems, even reaching expert-level performance in specific domains. However, these systems often remain narrowly focused and require significant manual crafting for each new task. Addressing this limitation, researchers have introduced SciAgent, a groundbreaking unified multi-agent system designed for what they call ‘generalistic scientific reasoning’.
Generalistic scientific reasoning refers to an AI system’s ability to flexibly adapt its reasoning strategies across different scientific disciplines and varying levels of difficulty. SciAgent aims to move beyond specialized AI, which often relies on pre-engineered pipelines, towards a more adaptable and intelligent approach that mirrors human scientific cognition.
SciAgent’s Innovative Architecture
SciAgent’s core innovation lies in its hierarchical multi-agent architecture, which organizes problem-solving into three interconnected layers:
- Coordinator Agent (Meta Level): This top-level agent acts as a meta-reasoner. It interprets the problem’s domain, modality (e.g., symbolic, numerical, conceptual), and difficulty. Based on this analysis, it dynamically routes the problem to the most suitable specialized Worker System.
- Worker Systems (Domain Level): Each Worker System is a self-contained multi-agent ensemble specializing in a particular scientific domain or reasoning mode. For instance, there are dedicated Worker Systems for Math Olympiads, Physics Olympiads, Chemistry Olympiads, and General Exams. These systems translate the Coordinator’s high-level instructions into executable reasoning plans and manage the collaboration among their internal Sub-agents.
- Sub-agents (Execution Level): Within each Worker System, specialized Sub-agents perform concrete operations. Examples include Generator Agents for producing initial solutions, Reviewer Agents for validation and correction, Image Analyser Agents for interpreting visual data, and Molecule Agents for chemical structure analysis. These Sub-agents interact through structured message passing and critique-revision loops, refining intermediate results until a solution converges.
This hierarchical structure allows SciAgent to decompose complex problems, delegate tasks to specialized modules, and maintain coherence across diverse reasoning processes. It enables the system to dynamically assemble and refine reasoning pipelines tailored to each specific task, fostering adaptability and extensibility.
Achieving Gold-Medal Performance
SciAgent has been rigorously evaluated across various scientific competitions and benchmarks, demonstrating remarkable performance:
- Mathematics Olympiads: In the International Mathematical Olympiad (IMO 2025), SciAgent achieved a total score of 36 out of 42, surpassing the average human gold-medalist score. In the International Mathematics Competition (IMC 2025), it attained a perfect score of 100, matching the highest human performance.
- Physics Olympiads: The system consistently reached gold-level performance in the International Physics Olympiad (IPhO 2024, IPhO 2025) and the Chinese Physics Olympiad (CPhO 2025), outperforming average human gold medalists.
- Chemistry Olympiads and General Scientific Reasoning: While still under development for full gold-medalist comparison due to data availability, SciAgent has shown strong generalization capabilities on the International Chemistry Olympiad (IChO 2025) and selected problems from the challenging Humanity’s Last Exam (HLE) benchmark, covering mathematics, physics, chemistry, and biology tasks.
These results highlight SciAgent’s ability to handle abstract symbolic reasoning, complex conceptual modeling, quantitative derivation, and multimodal inputs across different scientific fields.
Also Read:
- AI Agents Collaborate to Uncover New Scientific Machine Learning Methods
- The Station: A New AI Environment for Autonomous Scientific Discovery
A Step Towards General Scientific Intelligence
The researchers view SciAgent not just as a specialized multi-agent system but as a concrete step toward generalistic scientific intelligence. By orchestrating reasoning as a dynamic, collaborative process among specialized agents, SciAgent moves beyond traditional model scaling to an architecture that can autonomously select, compose, and revise reasoning strategies appropriate to each problem’s structure.
Future work aims to expand SciAgent to additional scientific domains like biology, integrate more multimodal reasoning capabilities (e.g., visual diagram analysis), develop self-improving feedback loops, and ultimately deploy it as a partner in real-world scientific research workflows. This ambitious project, detailed in their paper available at arXiv:2511.08151, lays a foundation for the next generation of AI systems that can not only reproduce expert reasoning but also contribute original insights to scientific discovery.
