S1-NexusAgent: a Self-Evolving Agent Framework for Multidisciplinary Scientific Research

Modern scientific research relies on large-scale data, complex workflows, and specialized tools, which existing LLMs and tool-based agents struggle to handle due to limitations in long-horizon planning, robust goal maintenance, and continual learning from execution. To address these issues, in this work, we propose S1-NexusAgent, a self-evolving agent framework designed for multidisciplinary scientific research. S1-NexusAgent adopts a hierarchical Plan-and-CodeAct execution paradigm, decoupling global scientific planning from subtask-level tool execution through a dual-loop architecture, thereby enabling stable modeling of complex research workflows. The system natively supports the Model Context Protocol (MCP), integrates up to thousands of cross-disciplinary scientific tools, and achieves efficient orchestration of heterogeneous research tools via intention-aware dynamic tool retrieval and hot-plug mechanisms. To address long-context and large-scale data challenges in scientific settings, S1-NexusAgent introduces object-reference-based sparse context management, which enables sub-task context isolation and intermediate result compression. Building on this, a Critic Agent automatically evaluates complete execution trajectories and distills high-quality research paths into reusable Scientific Skills, forming a closed loop for continuous self-evolution, which is valuable for sustainable and long-horizon scientific research. Experiments on authoritative scientific benchmarks involving long-horizon planning and complex specialized tool orchestration, including biomini-eval (biology), ChemBench (chemistry), and MatSciBench (material science), demonstrate that S1-NexusAgent achieves state-of-the-art performance, validating its effectiveness and generalization capability in complex scientific tasks.

💡 Research Summary

S1‑NexusAgent is presented as a general‑purpose, self‑evolving scientific agent designed to tackle the growing complexity of modern multidisciplinary research, which involves massive datasets, long‑horizon experimental pipelines, and thousands of specialized tools. The core of the system is a hierarchical Plan‑and‑CodeAct paradigm implemented through an inner‑outer dual‑loop architecture. The outer loop performs high‑level planning: it parses a user’s scientific intent, generates a global roadmap of research stages, and continuously monitors task progress. The inner loop couples reasoning with executable code via the CodeAct module, allowing the agent to invoke real scientific tools, iteratively refine actions based on tool feedback, and terminate sub‑tasks when objectives are met.

To manage the vast tool ecosystem, S1‑NexusAgent integrates the Model Context Protocol (MCP) and introduces an Intent‑Aware Dynamic Hot‑Plugging (DHP) mechanism. Rather than loading all available tools into the reasoning context, DHP filters the tool repository based on the current sub‑task’s intent, loading only the most relevant instruments on demand. This reduces context bloat, preserves reasoning accuracy, and enables seamless scaling to thousands of heterogeneous tools across biology, chemistry, materials science, astronomy, mathematics, and scientific computing.

Long‑context and large‑data challenges are addressed with an object‑reference‑based sparse context management framework. Instead of retaining raw data in the LLM’s token window, the system stores key experimental artifacts and metadata as lightweight object references. Four complementary mechanisms—object‑level referencing, sub‑task context isolation, execution‑trajectory compression, and planning‑aware context augmentation—ensure that only decision‑relevant information is kept in memory, suppressing noise and preventing token‑limit violations during extended investigations.

Self‑evolution is realized through a Trajectory‑Evaluation‑based Self‑Evolution (TE‑SE) loop. After a full execution trajectory is completed, a dedicated Critic Agent evaluates its quality, extracts high‑performing pathways, and distills them into reusable “Scientific Skills.” These skills are stored in a skill library and can be invoked by the Planner in future tasks, effectively allowing the agent to learn from its own successes and improve decision‑making efficiency over time.

Empirical evaluation spans three authoritative benchmarks that stress long‑horizon planning and complex tool orchestration: biomini‑Eval (biology), ChemBench (chemistry), and MatSciBench (materials science). Across all metrics—goal‑drift reduction, tool selection precision, overall task success rate, and computational efficiency—S1‑NexusAgent outperforms existing state‑of‑the‑art domain‑specific and general scientific agents. Additional real‑world case studies demonstrate end‑to‑end automation of full research pipelines, from data preprocessing and hypothesis generation to experimental execution and manuscript drafting.

The paper’s contributions are fourfold: (1) a dual‑loop Plan‑and‑CodeAct architecture that stabilizes long‑horizon scientific workflows, (2) a dynamic, MCP‑compatible tool ecosystem with intent‑aware hot‑plugging, (3) a sparse context management strategy tailored to scientific data scales, and (4) a trajectory‑evaluation self‑evolution framework that continuously enriches the agent’s skill repertoire. Limitations include reliance on sandboxed tool execution, pending integration with real laboratory hardware, and the need for robust security and multi‑agent conflict resolution mechanisms. Future work aims to extend live instrument APIs, scale multi‑agent collaboration, and automate meta‑learning of domain‑specific skills, moving toward truly autonomous, lifelong scientific assistants.