Darwin-180B-RSI: VIDRAFT AI Tops LEXam Legal Benchmarks
Korean AI startup VIDRAFT's 180B-parameter Darwin-180B-RSI model claims 1st place on LEXam legal benchmarks without legal training data.
Stock photo for illustration only, not from the actual event
- Darwin-180B-RSI ranked 1st on Hugging Face-certified LEXam (68.94) and LEXam-hard (45.72)
- Developed by Korean deep-tech AI startup VIDRAFT with 180 billion parameters
- Trained via recursive self-improvement on math and science without legal domain data
- Marks VIDRAFT's 7th top ranking across Hugging Face-certified leaderboards
The artificial intelligence community is witnessing a major milestone as Darwin-180B-RSI, a 180-billion-parameter reasoning model developed by South Korean deep-tech startup VIDRAFT, secured first place on both Hugging Face-certified legal reasoning benchmarks: LEXam with a score of 68.94 and LEXam-hard with 45.72. This achievement represents the company's seventh #1 ranking on Hugging Face-certified leaderboards.
What makes this feat remarkable is that Darwin-180B-RSI was never trained on legal datasets. Despite this, it outperformed heavyweights such as GPT-5, which scored 62.65, and Claude-4.5-Sonnet, which scored 58.01. The LEXam benchmark was designed by researchers from ETH Zurich, the University of Zurich, and the Max Planck Institute, utilizing 340 actual final exam sessions from Swiss university law faculties to test legal reasoning and the application of law to facts rather than mere statute memorization.
Stock photo for illustration only, not from the actual event
The LEXam benchmark presents a formidable challenge, given that the random-choice baseline sits at 25 points while prior models typically hovered around the 50-point mark. The fact that the Darwin family model achieved such high performance by leveraging generalized reasoning acquired through math and science problems highlights how abstract problem-solving capabilities can transfer seamlessly across domains, reducing the need for costly domain-specific fine-tuning.
Under the hood, Darwin-180B-RSI relies on two core compounding technologies:
- Model evolution (Darwin): Bypassing the need to pretrain a massive model from scratch, it analyzes existing open models at the layer and expert level, combining top-performing components through a diagnostics-driven merging process.
- Recursive Self-Improvement (RSI): The evolved model undergoes a self-bootstrapping training loop using verifiable math and science problems. During this phase, only about 0.02% of total parameters were modified while keeping expert modules and routers intact.
Additionally, the model integrates ZTC (Zero-Token Confidence), a technique that estimates answer confidence by inspecting internal model states before generating output. VIDRAFT has made Darwin-180B-RSI and its full evaluation conditions publicly available on Hugging Face for independent verification.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment