A Korean AI startup just proved that true general intelligence doesn't need a law degree — or even a law textbook.
TL;DR: VIDRAFT, a Korean AI startup, has built a model called Darwin-180B-RSI that ranked first on the LEXam and LEXam-hard legal reasoning leaderboards on Hugging Face — despite never being trained on legal data. The 180-billion-parameter model now holds first place on seven official Hugging Face leaderboards, outscoring GPT-5, Claude-4.5-Sonnet, and Gemini-2.5-Pro. The achievement is attributed to the model's recursive self-improvement capability, which lets it refine its own reasoning without domain-specific training.
VIDRAFT, the Korean AI startup making waves in the global AI research community, has achieved something that would have seemed implausible just a few years ago: its flagship model, Darwin-180B-RSI, has claimed the top spot on one of the most rigorous legal reasoning benchmarks in existence — without ever having been trained on a single piece of legal data. The result, reported by WondTech on October 1, 2026, is sending ripples through the AI research community and raising urgent questions about how far self-improving AI can actually go.
Darwin-180B-RSI is a 180-billion-parameter model developed by VIDRAFT. According to WondTech's reporting, the model recently reached first place on both the LEXam and LEXam-hard leaderboards, which are official benchmarks for legal reasoning ability hosted on Hugging Face. In doing so, it surpassed well-known frontier models including GPT-5, Claude-4.5-Sonnet, and Gemini-2.5-Pro.
The numbers are concrete: Darwin-180B-RSI scored 68.94 points on LEXam, compared to GPT-5's score of 62.65. Beyond the legal benchmarks, the model now holds first place on seven official Hugging Face leaderboards in total — more than any other single organization currently listed on the platform.
What makes this especially striking is the context around those numbers. LEXam was developed collaboratively by researchers at ETH Zurich, the University of Zurich, and the Max Planck Institute. It draws from 340 real law-school examinations sourced from Swiss universities, available in both German and English. Critically, the benchmark is not designed to reward rote memorization of statutes or legal codes. Instead, it evaluates whether a model can take a set of facts, apply relevant legal principles, and reason its way logically to a conclusion — the same cognitive process expected of a human law student or practicing attorney. Darwin-180B-RSI had never been trained on legal datasets, yet it topped this benchmark anyway.
The mechanism behind the result, as described in the source article, is what VIDRAFT calls model-level recursive self-improvement. In practical terms, this means the model works through practice problems and autonomously refines its own performance in response — without requiring human curators to supply domain-specific training data. The model learns by doing, iterates on its own outputs, and improves itself in a closed loop.
The implications of this result stretch well beyond the legal domain. Historically, strong AI performance on a specialized benchmark has been treated as evidence of strong specialized training — a model trained on medical literature does well on medical tests, a model trained on code does well on coding benchmarks. Darwin-180B-RSI's performance on LEXam disrupts that assumption in a meaningful way.
If a model that has never studied law can reason through Swiss law-school examinations well enough to beat purpose-built competitors, it suggests that genuine generalization — the ability to transfer reasoning skills into entirely unfamiliar domains — may be further along than the research community broadly anticipated. The benchmark's design specifically guards against the shortcut of pattern-matching on memorized legal text, which makes the result harder to dismiss as a data-contamination artifact.
For the broader AI industry, this also signals that Korean AI startup VIDRAFT is competing at the very top tier of foundation model development, placing itself alongside — and in this benchmark, ahead of — the most prominent labs in the world.
Q: What is LEXam, and why is it considered a difficult benchmark?
A: LEXam is a legal reasoning benchmark created by researchers from ETH Zurich, the University of Zurich, and the Max Planck Institute. It is built from 340 real Swiss law-school exams in German and English, and it specifically tests applied legal reasoning — not memorization of statutes — making it difficult to game through data exposure alone.
Q: How did Darwin-180B-RSI perform so well on a legal benchmark without legal training data?
A: According to WondTech's reporting, the model uses a method VIDRAFT describes as model-level recursive self-improvement, in which it solves problems and autonomously improves its own performance over time without needing domain-specific datasets.
Q: How does Darwin-180B-RSI compare to other leading AI models?
A: On the LEXam benchmark, Darwin-180B-RSI scored 68.94, outperforming GPT-5 (62.65), Claude-4.5-Sonnet, and Gemini-2.5-Pro. It also currently leads seven official Hugging Face leaderboards — more than any other single organization.
Source: WondTech (글로벌) (2026-10-01) — original article