VIDRAFT's Darwin-180B-RSI Tops Legal Reasoning Benchmarks LEXam and LEXam-hard — Without Legal Training Data
TL;DR: VIDRAFT's Darwin-180B-RSI reasoning model scored 1st place on both LEXam (68.94) and LEXam-hard (45.72), two Hugging Face-certified legal reasoning benchmarks, despite never being trained on legal data. The model achieves this through Recursive Self-Improvement (RSI) on math and science problems — and the resulting reasoning generalizes across domains. This is now VIDRAFT's 7th #1 ranking across Hugging Face-certified leaderboards.
What it is
Darwin-180B-RSI is a 180-billion-parameter reasoning model developed by VIDRAFT (비드래프트), a Korean deep-tech AI startup. It is built on two compounding technologies:
- Darwin — VIDRAFT's model evolution framework that constructs new models by diagnosing the strengths and weaknesses of existing model layers and expert modules, then selectively combining high-performing components from multiple source models. This avoids full pretraining from scratch.
- RSI (Recursive Self-Improvement) — A training approach in which the model solves problems autonomously, automatically verifies its own solutions (on problems with deterministic answers, such as math and science), learns from confirmed correct solutions, and repeats the cycle with the improved version of itself. No human-written solution annotations are used.
The result is a model that has accumulated strong general reasoning capabilities — without domain-specific legal fine-tuning — that transfer to fields like law.
The Darwin family now spans ~50 official Hugging Face models and ~400 community-derived variants. The core methodology is documented in the Darwin Family paper (arXiv: 2605.14386).
How it works
At a high level, Darwin-180B-RSI is built in two stages:
Model evolution (Darwin): Instead of pretraining a large model from scratch, Darwin analyzes existing open models at the layer and expert level, identifies which components perform best on which tasks, and assembles a new model by combining those components. This is a diagnostics-driven merging process, scaled up to models with hundreds of billions of parameters.
Recursive Self-Improvement (RSI): The evolved model is then trained using a self-bootstrapping loop — it attempts to solve problems, uses automatic verification (possible only for problems with checkable answers, like math), and trains on the solutions it verified as correct. The updated model then becomes the starting point for the next iteration. This is explicitly a model-weight-level improvement: the learned reasoning is baked into the model weights themselves, requiring no external scaffolding or tool calls at inference time.
VIDRAFT distinguishes this from prompt- or workflow-level self-improvement approaches (such as modifying prompts, tools, or task graphs at runtime), noting the two paradigms are complementary rather than competing.
Additionally, Darwin-180B-RSI incorporates ZTC (Zero-Token Confidence), a technique that estimates answer confidence by inspecting internal model states before generating output — used to guide the self-improvement verification process.
Only approximately 0.02% of total parameters were modified during the RSI stage; the 512 knowledge-carrying expert modules, routers, and vision components were kept intact.
Benchmarks & results
All scores below are from Hugging Face-certified leaderboards:
| Benchmark | Darwin-180B-RSI | Previous #1 | Notes |
|---|---|---|---|
| LEXam (MCQ, 1,655 Qs) | 68.94 | DeepSeek-R1: 52.41 | 4-attempt majority vote |
| LEXam-hard (open-ended, 518 Qs) | 45.72 | Inkling: 40.82 | Scored by benchmark's official grader model |
| LEXam single-attempt (no majority vote) | 60.54 | — | Same model, 1 inference pass |
Context on LEXam difficulty: Random-choice baseline is 25 points (4-choice MCQ). Most prior models scored around 50. LEXam-hard is composed specifically of items where existing models struggled, with the previous top score in the low 40s.
Other models on LEXam (from benchmark developers' separate release): GPT-5: 62.65 · Claude-4.5-Sonnet: 58.01 · Gemini-2.5-Pro: 55.72 · Qwen3-235B-Thinking: 48.19 · gpt-oss-120b: 47.71.
Full #1 leaderboard record for Darwin-180B-RSI across 7 Hugging Face-certified benchmarks:
- AIME 2026 — 100
- HMMT 2026 — 100
- GPQA Diamond — 94.44
- MMLU-Pro — 88.12
- MMMU-Pro — 79.48
- LEXam — 68.94
- LEXam-hard — 45.72
VIDRAFT also reports that reasoning chain length decreased by ~11% compared to the parent model while maintaining equivalent accuracy.
How to try it
VIDRAFT states that Darwin-180B-RSI and its full evaluation conditions are publicly available on Hugging Face for independent verification. You can explore the model and the Darwin family at huggingface.co by searching for VIDRAFT or Darwin-180B-RSI. The core methodology paper is publicly accessible:
arXiv: 2605.14386 (Darwin Family) No additional API endpoint or SDK installation has been announced publicly at this time. Check VIDRAFT's official Hugging Face organization page for the latest model cards and access instructions.
FAQ
Q: Darwin-180B-RSI is 180B parameters — how does it beat DeepSeek-R1 (685B) on LEXam?
A: Parameter count is not the only determinant of benchmark performance. Darwin-180B-RSI uses a diagnostics-driven model merging approach (Darwin) combined with recursive self-improvement that strengthens generalizable reasoning. The model was not trained on legal data at all; its strong performance on LEXam is attributed to domain-general reasoning capabilities built up through RSI on math and science problems.
Q: What exactly is LEXam, and why does it matter for real-world legal AI?
A: LEXam was developed by researchers at ETH Zurich, the University of Zurich, and the Max Planck Institute. It draws from 340 actual final exam sessions from Swiss university law faculties. Critically, it tests legal reasoning — applying law to given facts to reach conclusions — rather than statute memorization. Switzerland and Korea both follow civil law systems, so VIDRAFT suggests these results offer a meaningful signal for AI applicability in Korean legal and administrative contexts.
Q: Does RSI require human-annotated solutions?
A: No. RSI as implemented by VIDRAFT uses only problems with automatically verifiable answers (math and science) so the model can confirm solution correctness without human review. Human-written solutions are explicitly not used in the self-improvement loop.
Originally reported by AI이코노미뉴스 (2026-10-01) — source article.
