lamm-mit/Graph-Preflexor-8b_12292025

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Dec 29, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The lamm-mit/Graph-Preflexor-8b_12292025 is an 8 billion parameter language model developed by lamm-mit, based on Qwen3-8B. It is specifically trained for graph-native scientific reasoning, producing structured intermediate representations and machine-readable knowledge graphs. This model excels at transparently generating scientific hypotheses and explanations, making its reasoning process inspectable and analyzable.

Loading preview...

Graph-Preflexor-8b_12292025: Structured Scientific Reasoning

This model, developed by lamm-mit, is an 8 billion parameter language model built upon Qwen3-8B, uniquely designed for scientific reasoning with explicit, structured intermediate representations. It was trained in two stages: first using ORPO (Offline Reinforcement Preference Optimization) to establish a graph-centric reasoning style with specific "sentinel" blocks, and then fine-tuned with GRPO (Generative Reinforcement Preference Optimization) using an external LLM judge.

Key Capabilities

  • Structured Reasoning: Generates explicit reasoning traces using distinct sentinel blocks like <brainstorm>, <graph>, <graph_json>, <patterns>, and <synthesis>. This allows for segmented and interpretable reasoning.
  • Machine-Readable Knowledge Graphs: Emits valid, parseable knowledge graphs in JSON format (<graph_json>), enabling programmatic extraction and downstream tooling.
  • Graph-Native Explanations: Optimized to encode scientific reasoning where the generated graphs themselves carry significant explanatory power, balancing correctness with structural richness.
  • External Judge Fine-tuning: Utilized grok-4-1-fast-non-reasoning as an external judge during GRPO, with a multi-component reward function emphasizing correctness, format compliance, graph utility, validity, diversity, and structural quality.

Good For

  • AI-for-Science Applications: Ideal for tasks requiring transparent and structured scientific hypothesis generation and knowledge discovery.
  • Graph-Native Reasoning Workflows: Suitable for scenarios where the output needs to be an analyzable knowledge graph alongside natural language explanations.
  • Interpretability & Auditability: Provides inspectable reasoning steps, crucial for applications where understanding "how" the model arrived at an answer is as important as the answer itself.
  • Complex Problem Solving: Designed to tackle scientific and reasoning questions by breaking them down into explicit exploration, formalization, abstraction, and explanation phases.