ProvMind: AI Framework for Materials Synthesis Reasoning
Researchers have developed a novel AI framework named ProvMind, which enhances reasoning in materials synthesis processes through the use of provenance-grounded graphs. To assess seven process-reasoning tasks, they created MatProcBench, a benchmark derived from MatPROV graphs found in scientific publications. These tasks encompass route continuity, step-level variable inference, and global causal consistency, evaluated through a strict dual-OOD split that integrates both temporal and material-class shifts. ProvMind identifies similar training processes, translates them into provenance-aware option-level compatibility scores, and employs a language model for constrained decision-making. It achieves an accuracy of 52.84% on the dual-OOD split, surpassing various baselines, including prompting and retrieval-augmented methods. This research is available on arXiv with ID 2605.28487.
Key facts
- ProvMind is a process-memory reasoning framework for materials synthesis.
- MatProcBench is a provenance-grounded benchmark from literature-mined MatPROV graphs.
- Seven process-reasoning tasks are evaluated.
- Dual-OOD split combines temporal and material-class shift.
- ProvMind achieves 52.84% accuracy on dual-OOD split.
- Outperforms prompting, retrieval-augmented, and supervised fine-tuning baselines.
- Published on arXiv with ID 2605.28487.
- Uses provenance-aware option-level compatibility scores.
Entities
Institutions
- arXiv