8B Model Beats 70B Giant

🔥 What happened Researchers independently re-implemented Direct Nash Optimization (DNO) and tested it on Llama 3.1 8B and Mistral 7B. Three iterations lift the length-controlled win rate against GPT-4o from 31.7% to 45.3% (Llama) and from 6.8% to 18.4% (Mistral). 💡 Why it matters The tuned 8B model beats Llama 3.1 70B on every reported win rate — a 9x size gap. With open-weight annotators, runs retain 62–82% of the gain, so the rating step doesn't need a proprietary model. ⚡ Our take Impressive, but GPT-4o is both teacher and judge — a self-grading bias the authors admit. Anyone calling this independent proof skipped the footnote.