When Does Synthetic Patent Data Help? Volume-Fidelity Trade-offs in Low-Resource Multi-Label Classification
The study investigates the effectiveness of LLM-generated synthetic data in low-resource multi-label patent classification. It finds that while larger augmented datasets can improve performance, the true value of synthetic data is context-dependent. The research highlights the importance of fidelity metrics and suggests optimal mixing strategies for real and synthetic data.
- ▪The headline BERT-for-Patents micro-F1 score improved from 0.120 to 0.702, primarily driven by volume.
- ▪Controlled synthetic gain was only +0.024 over a real-only control but +0.219 over the strongest non-augmentation baseline.
- ▪Fidelity metrics change meaning with scale, showing a positive correlation with classification gain at extreme scarcity.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.24296 (cs) [Submitted on 22 May 2026] Title:When Does Synthetic Patent Data Help? Volume-Fidelity Trade-offs in Low-Resource Multi-Label Classification Authors:Amirhossein Yousefiramandi, Ciaran Cooney View a PDF of the paper titled When Does Synthetic Patent Data Help? Volume-Fidelity Trade-offs in Low-Resource Multi-Label Classification, by Amirhossein Yousefiramandi and 1 other authors View PDF HTML (experimental) Abstract:We study when LLM-generated synthetic data helps low-resource multi-label patent classification, separating true synthetic value from the confound that larger augmented sets can win by volume alone.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
