BioRENLI: enhancing large language models for biomedical relation extraction through preference-aligned natural...

Biomedical relation extraction (RE) aims to identify typed relations between biomedical entities from text. Most existing RE approaches rely on text classification frameworks, which struggle to fully exploit the generative capabilities of large language models (LLMs). Moreover, when LLMs are directly fine-tuned for biomedical RE, they often confuse semantically similar relation types, leading to the incorrect prediction of plausible but unsupported relations. To address these limitations, we propose BioRENLI, an LLM-based biomedical RE framework driven by preference-aligned natural language inference (NLI). BioRENLI reformulates RE as an NLI task by converting each candidate relation into a natural language hypothesis, prompting the model to predict either Entailment or Contradiction given the input context. To further mitigate label ambiguity, we augment supervised fine-tuning with Direct Preference Optimization (DPO). Using label-derived preference pairs, DPO encourages the model to prefer gold NLI decisions over non-gold alternatives, thereby reducing unsupported entailment predictions. Experimental results on the ChemProt and DDI datasets show that BioRENLI consistently outperforms supervised fine-tuned LLM variants in both full-supervision and low-resource settings, while remaining competitive with strong BERT-based baselines, highlighting the effectiveness of enhancing biomedical RE with preference-aligned NLI.