VIDRAFT's Darwin-180B-RSI Tops Legal Reasoning Benchmarks LEXam and LEXam-hard — Without Legal Training Data

VIDRAFT's Darwin-180B-RSI model has achieved top rankings on two challenging legal reasoning benchmarks, LEXam and LEXam-hard, without any training data specific to law. This remarkable feat is achieved through a technique called Recursive Self-Improvement (RSI), which involves the model solving problems autonomously, verifying its own solutions, and learning from confirmed correct solutions.

The resulting reasoning capabilities can be applied across various domains, including legal reasoning. The Darwin-180B-RSI is part of the Darwin family of models developed by Korean AI startup VIDRAFT, which leverages two key technologies: Darwin, a model evolution framework that combines high-performing components from multiple source models, and RSI, a self-improvement training approach that allows the model to learn and improve itself.

These advancements have positioned Darwin-180B-RSI at the forefront of legal reasoning benchmarks, demonstrating impressive performance without the need for domain-specific legal training.