Analysing Lightweight Large Language Models for Biomedical Named Entity Recognition on Diverse Ouput Formats
Smaller models could cut costs in healthcare NLP without sacrificing accuracy.
A new study from researchers at HeKA (Inria) and IP Paris, presented at LREC 2026, challenges the assumption that only massive language models can handle specialized tasks like Biomedical Named Entity Recognition (NER). The team, led by Pierre Epron, tested lightweight LLMs on extracting entities such as diseases, drugs, and genes from medical texts, comparing their performance across different output formats.
Surprisingly, the lightweight models achieved competitive accuracy against their larger counterparts, suggesting that smaller, more efficient architectures can deliver practical value in resource-constrained environments. The study also revealed that fine-tuning on many distinct output formats does not improve performance; instead, a few specific formats consistently yielded better results. This finding simplifies deployment by reducing the need for extensive format experimentation. For healthcare organizations with strict privacy rules or limited budgets, these lightweight LLMs offer a viable path to automate information extraction without the high computational costs or cloud dependency of larger models.
- Lightweight LLMs achieve competitive performance vs. larger models on Biomedical NER tasks.
- Instruction tuning across many output formats does not improve performance, but specific formats consistently boost results.
- Smaller models are better suited for privacy-sensitive and budget-constrained healthcare settings.
Why It Matters
Enables cost-effective, private biomedical NLP, making AI accessible for smaller healthcare providers.