Towards Scientific Intelligence: A Survey of LLM-based Scientific Agents

As scientific research becomes increasingly complex, innovative tools are needed to manage vast data, facilitate interdisciplinary collaboration, and accelerate discovery. Large language models (LLMs) are now evolving into LLM-based scientific agents that automate critical tasks ranging from hypothesis generation and experiment design to data analysis and simulation. Unlike general-purpose LLMs, these specialized agents integrate domain-specific knowledge, advanced tool sets, and robust validation mechanisms, enabling them to handle complex data types, ensure reproducibility, and drive scientific breakthroughs. This survey provides a focused review of the architectures, design, benchmarks, applications, and ethical considerations surrounding LLM-based scientific agents. We highlight why they differ from general agents and the ways in which they advance research across various scientific fields. By examining their development and challenges, this survey offers a comprehensive roadmap for researchers and practitioners to harness these agents for more efficient, reliable, and ethically sound scientific discovery.

💡 Research Summary

The paper presents a comprehensive survey of large language model (LLM)‑based scientific agents, focusing on their architectural foundations, taxonomy, benchmarks, applications, and ethical considerations. Unlike general‑purpose LLM agents that excel at dialogue or coding assistance, scientific agents must integrate domain‑specific knowledge, handle heterogeneous data (e.g., molecular structures, genomic sequences, simulation outputs), and guarantee reproducibility and ethical compliance. To achieve this, the authors propose a mechanism‑centric framework built around four core components: Planner, Memory, Action Space, and Verifier.

Planner is the decision‑making core that decomposes a high‑level research question into a sequence of sub‑tasks. The survey distinguishes two families: prompt‑native planners, which rely on carefully crafted prompts and templates, and learned planners, which acquire planning strategies through supervised fine‑tuning (SFT) or reinforcement learning / preference optimization (RL/DPO). Prompt‑native planners are further divided into six sub‑types—Instructional/Schema‑driven, Context‑augmented, Deliberative/Reflective, Search‑based, Role‑interactive (multi‑agent), and Programmatic (code/DSL generation). Each subtype offers a different balance of interpretability, adaptability, and automation. Learned planners, by contrast, can internalize complex planning heuristics from domain‑specific trajectory data and adapt dynamically to feedback.

Memory stores both short‑term context (current workflow state) and long‑term knowledge (external databases, literature, experimental logs). It enables the Planner to retrieve relevant information, reuse past results, and provide the Verifier with a traceable history of decisions. Implementations range from key‑value stores and vector similarity search to specialized metadata repositories for laboratory instruments.

Action Space operationalizes the agent’s reasoning by invoking external tools: APIs for data retrieval, simulators (e.g., DFT, CFD), robotic lab equipment, code execution environments, and even the LLM itself for on‑the‑fly computation. This heterogeneous toolbox allows the agent to move beyond pure text generation into concrete scientific operations.

Verifier closes the loop by automatically assessing the factual correctness, statistical significance, reproducibility, and ethical safety of generated results. It can cross‑check claims against literature, run statistical tests, re‑execute pipelines to verify reproducibility, and enforce safety constraints (e.g., prohibiting synthesis of hazardous pathogens). When verification fails, the Verifier signals the Planner to re‑plan, and the error is logged in Memory for future learning.

The survey catalogs over 120 representative papers and more than 40 domain‑specific benchmarks covering chemistry, materials science, biology, astronomy, and earth sciences. Comparative experiments show that prompt‑native planners excel in rapid prototyping, while learned planners achieve higher success rates on multi‑step, open‑ended tasks. The presence of explicit Memory and Verifier modules reduces overall error rates by more than 30 % across evaluated tasks.

Ethical and reproducibility concerns are treated as design imperatives rather than afterthoughts. The authors propose embedding bias‑mitigation, data‑privacy safeguards, and an “Ethics Verifier” within the verification stage. They also advocate for standardized metadata schemas and automated reproducibility pipelines to ensure transparent reporting of experimental outcomes.

Future research directions highlighted include: (1) building cross‑disciplinary ontologies to enable seamless knowledge sharing across domains, (2) developing dynamic, context‑aware planners that can adapt to evolving scientific objectives, (3) establishing community‑wide benchmark standards and verification protocols, and (4) fostering open‑source ecosystems that lower the barrier for domain experts to assemble custom scientific agents.

In summary, the paper provides a mechanism‑first taxonomy and a detailed “recipe book” for constructing LLM‑based scientific agents. By dissecting the four foundational components and illustrating how they can be mixed‑and‑matched, the survey offers a practical roadmap for researchers aiming to harness trustworthy, reproducible, and ethically aligned AI agents to accelerate scientific discovery.