SciAgents

Overview

SciAgents explores how graph-guided reasoning and specialized language-model agents can support hypothesis development in bio-inspired materials research. The repository supplies notebook implementations of two workflows: a predefined sequence of agent interactions and a dynamically coordinated automated framework. It is research code for generating and refining proposals, rather than evidence that the proposed materials or mechanisms have been experimentally validated. The automated implementation uses AG2, and the project is included in the Build with AG2 collection.

The described workflow starts with selected keywords or random graph exploration. A sampled path supplies concepts and relationships that agents use to construct a research hypothesis. An Ontologist clarifies the concepts, two Scientist roles develop and extend the proposal, and a Critic identifies weaknesses and possible improvements. The automated approach additionally includes a Planner and an Assistant that checks hypothesis novelty. Structured JSON captures the hypothesis, expected outcomes, mechanisms, design principles, unexpected properties, comparisons and novelty; successive expansion produces a longer research draft with critical analysis and proposed modeling or experimental priorities.

Running the supplied workflows requires the separate GraphReasoning package, OpenAI and Semantic Scholar APIs, and graph and embedding files identified in the README. These dependencies distinguish SciAgents from a self-contained discovery service. Its outputs are candidate research directions for further investigation: novelty assessments, proposed mechanisms and suggested validation tasks should be independently evaluated before informing scientific conclusions.

Key Features

• Provides separate notebooks for predefined and automated multi-agent graph-reasoning workflows.
• Uses keyword selection or random graph exploration followed by path sampling to supply hypothesis-generation context.
• Assigns concept definition, proposal development, refinement and critique to specialized agent roles.
• Adds planning and hypothesis-novelty checking roles in the automated framework implemented with AG2.
• Produces structured hypothesis components in JSON and expands them into a research draft with critical analysis.
• Incorporates retrieved research information and identifies modeling, simulation and experimental priorities for follow-up.

Use Cases

• Suggested evaluation: explore connections between bio-inspired material concepts and assess whether graph-guided proposals offer useful research directions.
• Suggested evaluation: compare the predefined and automated workflows on the same research question, examining proposal coherence and the usefulness of agent critiques.
• Suggested evaluation: use generated mechanisms and design principles as starting points for a researcher-led literature review and validation plan.

How to Use

1. Read the SciAgents README to understand the graph-guided workflow, agent roles and bio-inspired materials examples before choosing a reproduction task.
2. In the repository, inspect SciAgents_ScienceDiscovery_GraphReasoning_non-automated.ipynb and SciAgents_ScienceDiscovery_GraphReasoning_automated.ipynb in the Notebooks directory. Select the predefined or automated interaction model.
3. Follow the README requirements for the separate GraphReasoning package and access to OpenAI and Semantic Scholar APIs. Inspect the notebook configuration rather than assuming a ready-to-run service.
4. Obtain large_graph_simple_giant.graphml and embeddings_simple_giant_ge-large-en-v1.5.pkl from lamm-mit/bio-graph-1K using the retrieval procedure supplied in the README. Check how the selected notebook loads these inputs.
5. Evaluate a keyword-based or exploratory graph task. Review the resulting hypothesis components, expanded draft and critiques against retrieved literature, then select claims for independent modeling or experimental investigation; generated proposals are not validation results.