An open-weight 180B model named Darwin-180B-RSI, developed by the Korean startup VIDRAFT, has achieved top rankings on legal reasoning benchmarks, despite never being trained on legal data. The model utilizes a novel approach called model-level recursive self-improvement (RSI), where it learns by solving practice problems and then trains only on its own correct solutions. This method appears to foster a general reasoning ability that transfers to new domains, as evidenced by its superior performance on Swiss law exams compared to leading models like GPT-5, Claude-4.5-Sonnet, and Gemini 2.5 Pro. AI
IMPACT Demonstrates that general reasoning skills can be trained without domain-specific data, potentially accelerating AI adaptation to new fields.
RANK_REASON The item details a novel AI model training methodology and its performance on a specific benchmark, aligning with research findings. [lever_c_demoted from research: ic=1 ai=1.0]
- Claude-4.5-Sonnet
- Darwin-180B-RSI
- DeepSeek
- ETH Zurich
- Gemini 2.5 Pro
- GPT-5
- Hugging Face
- LEXam-hard
- Max Planck Institute
- Moonshot AI
- University of Zurich
- VIDRAFT
- Xiaomi
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
