README.md

metadata

title: Graph-PrefLexOR Scientific Reasoning
emoji: 🧠
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.19.0
app_file: app.py
short_description: Graph-native LLM scientific reasoning with graph viz
python_version: '3.12'
startup_duration_timeout: 1h

Graph-PrefLexOR: Graph-Native Reinforcement Learning for Scientific Hypothesis Generation

This Space demonstrates lamm-mit/Graph-Preflexor-8b_12292025 — a Qwen3-8B model fine-tuned with graph-native reinforcement learning (ORPO + Graph-GRPO) to produce structured, traceable scientific reasoning.

How it works

The model reasons through scientific questions using explicit structured blocks:

1. <brainstorm> — divergent exploration of hypotheses and mechanisms
2. <graph> — verbal sketch of the conceptual graph
3. <graph_json> — machine-readable knowledge graph (nodes + edges)
4. <patterns> — abstract motifs and design rules
5. <synthesis> — coherent narrative assembled from the graph

The Space parses the <graph_json> block and renders it as an interactive Mermaid diagram alongside the chat.

Paper

Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination — GitHub

Tips

• Ask materials science, biology, or interdisciplinary scientific questions for best results.
• The graph panel updates after generation completes.
• Adjust temperature and max tokens in the Advanced settings.