Title: Med-V1: Lightweight Language Models Enable Scalable, Zero-Shot Biomedical Evidence Attribution

Verifying whether source material substantiates specific claims is a critical step in detecting hallucinations and validating assertions. Although large language models (LLMs) offer the capability to automate this verification process, achieving high performance typically necessitates the use of cutting-edge models like GPT-5. However, the prohibitive cost of deploying such frontier models at scale presents a significant barrier. To address this challenge in biomedical evidence attribution, we introduce Med-V1, a new family of small language models comprising just three billion parameters.

Developed using high-quality synthetic data created specifically for this research, Med-V1 demonstrates substantial improvements over its base models, with performance gains ranging from 27.0% to 71.3% across five biomedical benchmarks reformatted for verification tasks. Remarkably, despite its reduced size, Med-V1 delivers performance comparable to leading frontier LLMs, including GPT-5, while also providing high-quality explanations to support its predictions.

We leveraged Med-V1 for a pioneering use case study designed to quantify hallucinations in LLM-generated responses under varying citation instructions. The findings indicate that instruction formatting significantly influences both citation validity and hallucination rates. Notably, while GPT-5 produced a higher volume of claims, its hallucination rate remained similar to that of GPT-4o.

In a second application, we demonstrated that Med-V1 can automatically detect high-stakes evidence misattributions within clinical practice guidelines. This capability highlights potentially adverse public health impacts that are otherwise difficult to identify on a large scale. Ultimately, Med-V1 serves as an efficient, accurate, and lightweight alternative to expensive frontier LLMs, offering a practical solution for real-world biomedical evidence attribution and verification. Med-V1 is accessible at https://github.com/ncbi-nlp/Med-V1.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC