A Korean AI startup just proved that true reasoning generalizes across domains — even law.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, announced on October 1, 2026 that its reasoning model Darwin-180B-RSI has claimed first place on both the LEXam and LEXam-hard legal reasoning benchmarks on Hugging Face — despite never being trained on legal data. The model learned to reason using only math and science problems through a self-improvement process, then outperformed larger, well-known models including DeepSeek-R1, GPT-5, and Claude 4.5 Sonnet on legal reasoning tasks.
VIDRAFT, the Korean Pre-AGI AI startup, made headlines on October 1, 2026 when it revealed that its proprietary reasoning model, Darwin-180B-RSI, had ranked first on two globally recognized legal reasoning benchmarks — all without having seen a single piece of legal training data. The result challenges long-held assumptions about how AI models acquire domain-specific expertise, suggesting that a sufficiently strong general reasoner can transfer its capabilities across fields it has never formally studied.
VIDRAFT's Darwin-180B-RSI model achieved top scores on both the LEXam and LEXam-hard benchmarks, which are officially recognized on Hugging Face and were developed by researchers including those at ETH Zurich. These evaluations are grounded in real university-level law school examinations and are designed to test not rote memorization of statutes, but the ability to apply facts to legal principles and reach well-reasoned conclusions.
On the LEXam multiple-choice evaluation, Darwin-180B-RSI scored 68.94, surpassing DeepSeek-R1 (52.41), Qwen-235B-Thinking (48.19), and gpt-oss-120b (47.71). Notably, Darwin-180B-RSI achieved this with 180 billion parameters — significantly fewer than DeepSeek-R1's 685 billion. On the more demanding LEXam-hard open-ended assessment, the model scored 45.72, overtaking the previous leader, Thinking Machines' Inkling, which had scored 40.82. According to scores published by the LEXam development team, Darwin-180B-RSI also outperformed GPT-5 (62.65) and Claude 4.5 Sonnet (58.01) on the LEXam multiple-choice benchmark.
The driving force behind these results is VIDRAFT's Recursive Self-Improvement (RSI) methodology. Rather than relying on human-authored answer keys, the model works through problems autonomously, verifies its own solutions, and then uses the confirmed correct solutions as new training signal. Critically, this self-improvement loop was conducted exclusively using mathematics and science problems — no legal content was involved at any stage.
The underlying Darwin architecture is equally notable. Instead of training a new model from scratch — a process that can cost hundreds of millions of Korean won — Darwin merges and combines the strengths of existing models. VIDRAFT supplements this with a self-improvement mechanism that modifies only 0.02% of the model's total parameters, alongside a proprietary ZTC (Zero-Token Confidence) technique that calculates the model's internal confidence in its answers. Together, these approaches maintain accuracy comparable to the parent model while trimming reasoning output length by 11%.
With this latest achievement, VIDRAFT now holds first place in seven categories across 47 officially recognized Hugging Face leaderboard divisions, covering mathematics, science, general knowledge, visual reasoning, and now law. This places the company ahead of Zhipu AI (four top positions), and DeepSeek, Xiaomi, and Moonshot AI (three each), making VIDRAFT the Korean company with the most Hugging Face leaderboard first-place rankings.
The LEXam benchmark's design — emphasizing applied legal reasoning over statute memorization — makes VIDRAFT's result especially meaningful. Because South Korea shares a civil law legal tradition with the European jurisdictions that informed LEXam's construction, this benchmark performance is seen as a credible indicator of how well such a model could support real-world legal, public-sector, and administrative applications in Korea.
More broadly, the findings carry implications for how the AI industry thinks about domain specialization. Historically, practitioners assumed that excelling at legal reasoning required extensive legal fine-tuning. VIDRAFT's approach suggests that robust general reasoning — trained on math and science — can generalize in ways that rival or exceed models built with explicit domain knowledge. CEO Min-sik Kim commented in the announcement that a model topping a legal exam without ever studying law is evidence that it has learned how to reason, not merely how to recall.
The model and its evaluation conditions are publicly accessible on Hugging Face, allowing independent verification by the research community.
Q: How did VIDRAFT's Darwin-180B-RSI top legal benchmarks without legal training data?
A: The model used a Recursive Self-Improvement process — autonomously solving and verifying math and science problems — to develop general reasoning skills strong enough to transfer to legal reasoning tasks without any domain-specific legal training.
Q: How does Darwin-180B-RSI compare to larger models like DeepSeek-R1?
A: Despite having 180 billion parameters compared to DeepSeek-R1's 685 billion, Darwin-180B-RSI scored significantly higher on LEXam (68.94 vs. 52.41), demonstrating that efficient reasoning architecture can outperform sheer model scale.
Q: What is the LEXam benchmark and who developed it?
A: LEXam is a Hugging Face-recognized legal reasoning benchmark developed by researchers including those at ETH Zurich. It is based on real law school examinations and tests applied legal reasoning — fact analysis and conclusion-drawing — rather than memorization of legal codes.
Source: AI타임스 (2026-10-01) — original article