From Knowledge to Action: A Panoramic Analysis of LLM Hackathon Results in Materials Science and Chemistry for 2025

This article systematically analyzes community projects from the 2025 LLM Hackathon in materials science and chemistry, identifies two major application paradigms—knowledge infrastructure and action systems, reveals the evolutionary trend from single tools to multi-agent workflows, and provides a practical classification framework for understanding LLM applications throughout the full lifecycle of scientific research.

Large Language Models (LLMs) are permeating the entire chain of scientific discovery, from literature retrieval to hypothesis generation. The first 2025 LLM Hackathon in materials science and chemistry provides a unique window to observe LLM applications, with global teams submitting dozens of innovative projects that demonstrate the diverse uses of LLMs in traditional scientific fields.

Knowledge Infrastructure

Core tasks include structuring, retrieving, synthesizing, and verifying scientific information, focusing on static knowledge. Typical applications include intelligent literature review systems, knowledge graph construction tools, hypothesis verification assistants, and multilingual scientific translation. Their characteristics are enhancing information acquisition capabilities, acting as an 'external brain' rather than 'hands'.

Action Systems

The goal is to execute, coordinate, or automate scientific work, intervening in dynamic research processes. Typical applications include experimental design optimizers, computational workflow orchestration, laboratory automation interfaces, and real-time data analysis pipelines. Their characteristics are blurring the boundary between cognition and action, shifting from 'advisors' to 'executors'.

Rise of Multi-Agent Architecture

Adopting multi-agent division of labor and collaboration, such as retrieval, reasoning, tool use, and verification agents, each focusing on tasks they excel at to achieve complex goals.

RAG as Foundational Infrastructure

Retrieval-Augmented Generation (RAG) has evolved from an optional optimization to a core component, meeting the strict requirements for accuracy and traceability in scientific fields.

Persistent Structured Knowledge Representation

Persisting reasoning results into structured forms like knowledge graphs improves verifiability, reusability, composability, and incremental update capabilities.

Visual Understanding Capabilities

Combining computer vision with LLMs to recognize microstructures from SEM images, analyze crystal structures from XRD patterns, and extract phase boundary data from phase diagrams.

Multilingual Scientific Literature Processing

Automatically translating non-English literature while maintaining terminological accuracy, extracting key findings for localization, and building multilingual knowledge retrieval systems to break language barriers.

Dimension	Description	Example
Knowledge Depth	Processing surface-level information or deep knowledge	Keyword retrieval vs. Mechanistic reasoning
Action Capability	Ability to influence the physical world	Pure information retrieval vs. Experimental control
Autonomy	Amount of human intervention required	Auxiliary tool vs. Autonomous agent
Verifiability	Whether output can be independently verified	Black-box generation vs. Traceable citation
Collaboration Mode	Interaction method with human researchers	Q&A interface vs. Workflow integration
This framework helps position projects, identify improvement directions, and facilitate comparative learning.

Key Challenges

• Tension between accuracy and creativity: Limited ability to generate novel hypotheses
• Interpretability and credibility: Need to improve reasoning transparency
• Data quality and bias: Publication bias in literature may be amplified
• Computational cost and sustainability: High cost of large LLM calls

In the future, there is a need to balance accuracy and creativity, enhance interpretability, mitigate data bias, and optimize model efficiency.

The 2025 Hackathon marks a turning point in AI applications in materials science and chemistry, where LLMs have evolved from general-purpose conversational assistants to composable scientific reasoning infrastructure. Core features include shifts from Q&A to workflows, single-modality to multimodality, information to action, and general-purpose to domain-specific. Although challenges exist, the direction is clear—LLMs will accelerate scientific discovery and change the way researchers interact in the future.