A self-improving Korean AI model outperformed GPT-5, Claude, and DeepSeek on one of the hardest legal reasoning tests in the world — without ever training on legal data.
TL;DR: VIDRAFT, a Korean AI startup, announced on October 1, 2026 that its self-improving reasoning model Darwin-180B-RSI ranked first on both the LEXam and LEXam-hard legal benchmarks on Hugging Face — without any prior legal domain training. The model now holds first place in seven official Hugging Face benchmark leaderboard categories, more than any other AI company globally. It outscored models from DeepSeek, Alibaba, OpenAI, Anthropic, and Google on the legal reasoning evaluations.
VIDRAFT, the Korean Pre-AGI AI startup led by CEO Minsik Kim, claimed a landmark achievement on October 1, 2026: its flagship reasoning model, Darwin-180B-RSI, secured the top position on both the LEXam and LEXam-hard legal reasoning benchmarks hosted on the global AI platform Hugging Face — and did so without having been trained on any legal domain data whatsoever.
Darwin-180B-RSI now holds first place across seven officially recognized Hugging Face benchmark leaderboard categories. The two new legal reasoning victories added to five categories the model had already topped in mathematics, science, general knowledge, and visual reasoning. Across all 47 benchmarks officially recognized by Hugging Face, no other AI company currently holds more first-place titles. By comparison, Zhipu AI holds four, while DeepSeek, Xiaomi, and Moonshot AI each hold three.
The legal benchmark in question — LEXam — was developed jointly by researchers at ETH Zurich, the University of Zurich, and the Max Planck Institute. It is built from 340 real final examinations administered by Swiss law school programs, and it tests applied legal reasoning: the ability to interpret facts, apply relevant statutes, and reach a legal conclusion. It is explicitly not a test of memorization. The LEXam dataset consists of 1,655 multiple-choice questions, while LEXam-hard is composed of 518 open-ended questions drawn from items that most leading models failed to answer correctly.
The difficulty of these benchmarks is notable. Despite being structured as four-choice multiple-choice questions — where random guessing would yield a baseline score of 25 — most of the world's top large language models average around 50 points on LEXam. DeepSeek-R1, a model with 685 billion parameters, previously led the leaderboard with 52.41 points. Darwin-180B-RSI, a 180-billion-parameter model, scored 68.94 points using a four-pass majority-vote approach, and 60.54 points in a single-pass inference — still above the prior leaderboard record. On LEXam-hard, Darwin-180B-RSI scored 45.72 points, ahead of Inkling by Thinking Machines (40.82) and DeepSeek-V4-Pro (38.93). The benchmark's developers separately measured GPT-5 at 62.65, Claude-4.5-Sonnet at 58.01, and Gemini-2.5-Pro at 55.72; Darwin-180B-RSI's majority-vote score of 68.94 exceeds all three.
The results are attributable to Darwin-180B-RSI's underlying training architecture. Rather than injecting human-written explanations or legal commentary, the model was built using a Recursive Self-Improvement (RSI) framework. In this approach, the model autonomously attempts problems, mechanically verifies whether its answers are correct, and then selectively retrains on only the logical pathways that produced correct answers. The improved model then becomes the computational agent for the next iteration. During foundational training, only problems with objectively verifiable answers — mathematics and science — were used. No human-written intermediate reasoning steps were included. VIDRAFT describes this as evidence that RSI develops a generalizable reasoning process, not just domain-specific knowledge — enabling the model to transfer its multi-step logical capabilities into fields like law that it never explicitly studied.
VIDRAFT also notes that Darwin-180B-RSI maintains the same precision as its base model while reducing inference token length by 11%, meaning the model has optimized its reasoning paths rather than simply computing more. The model's checkpoints and evaluation environment are publicly available on Hugging Face for independent verification, and VIDRAFT has disclosed the specific benchmark parameters used — including pass count, ensemble voting method, and inference token budget.
On the architectural distinction: Google has demonstrated system-level self-improvement that adjusts prompts, external tools, and workflows while keeping base weights fixed. VIDRAFT's RSI approach directly updates the model's internal parameter weights, meaning performance improvements are permanently encoded in a single model file, with no auxiliary systems required at inference time. VIDRAFT has noted that both approaches are complementary rather than competing.
Swiss civil law belongs to the same continental legal tradition as Korean law, making LEXam performance a credible proxy for AI readiness in Korean legal and administrative contexts. A Korean startup outperforming significantly larger models from well-funded global AI labs — using a 180-billion-parameter model to beat a 685-billion-parameter DeepSeek-R1 — signals that architectural innovation in training methodology can overcome raw scale advantages. It also validates RSI as a pathway toward domain-general reasoning rather than narrow specialization.
Q: What is Darwin-180B-RSI and who built it?
A: Darwin-180B-RSI is a self-improving large language model developed by VIDRAFT, a Korean AI startup. It uses a Recursive Self-Improvement (RSI) architecture that retrains the model on verified correct reasoning paths rather than on human-written explanations.
Q: How did a model with no legal training data score first on a legal reasoning benchmark?
A: VIDRAFT's RSI framework trained Darwin-180B-RSI exclusively on mathematics and science problems with objectively verifiable answers. The company says the multi-step logical reasoning skills developed through that process transferred to legal reasoning, demonstrating that RSI builds generalizable reasoning ability rather than domain-specific knowledge.
Q: What is LEXam and why is it significant?
A: LEXam is an official legal reasoning benchmark developed by ETH Zurich, the University of Zurich, and the Max Planck Institute, based on 340 real Swiss law school final exams. It tests applied legal reasoning rather than rote memorization. Because Swiss law belongs to the same continental legal tradition as Korean law, strong performance on LEXam is considered a meaningful indicator of AI suitability for Korean legal and administrative applications.
Source: 로이슈 (2026-10-01) — original article