VIDRAFT has released Darwin-180B-RSI, a 180-billion-parameter language model that has achieved top rankings on two legal reasoning benchmarks, LEXam and LEXam-hard. Notably, the model accomplished this without any specific legal training data, demonstrating emergent reasoning capabilities from general math and science problem-solving. This achievement is attributed to VIDRAFT's novel techniques of diagnostics-driven model merging, Recursive Self-Improvement (RSI) with a Zero-Token Confidence (ZTC) mechanism, which modified only a small fraction of the model's parameters. AI
IMPACT Challenges assumptions about domain specialization and suggests reasoning generalization can transfer effectively to specialized fields.
RANK_REASON New benchmark results for a language model, with technical details on its methodology. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Claude-4.5-Sonnet
- Darwin-180B-RSI
- DeepSeek-R1
- Gemini 2.5 Pro
- GPT-5
- Hugging Face
- LEXam-hard
- South Korea
- VIDRAFT
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →