Model merge: Baphomet-1.1-4B beats all expert models across 3 benchmarks
pbaylies · x · 2026-11-13
Developer croudgoat released Baphomet-1.1-4B, built via evolutionary model merging. By exploiting constructive interference among expert models, it tops the 110-item fitness set with a mean of 76.4, beating every merged expert model (next best: EvoGM round-1 winner at 72.2).
- Scores 80.0 on MMLU-Pro (tied for best), 56.7 on GPQA (leading), 92.5 on BFCL
- The author notes the fitness set is the same one used to select version 1.1, so its score is optimistic
- Partially demonstrates that evolutionary model merging works on in-the-wild models
More from Research
- Each 10x in compute buys roughly 10x longer AI time horizons, elasticity near 1 — ben_j_todd · 2026-11-13
- NeuroAI work clarifies CNN vs transformer brain models — transformers win in interesting ways — dyamins · 2026-11-13
- OpenAI's $0.11 model beats the priciest LLMs in new Slack agent benchmark — heyshrutimishra · 2026-11-13
- World Models sessions pack CoRL: standing room only at 9am panel — chris_j_paxton · 2026-11-13
- Pure RL Trains a Strong 4B Continual-Learning LLM, No Distillation or SFT — aviral_kumar2 · 2026-11-13
- CMU trains LLMs to write memories with RL for continual learning across task streams — aviral_kumar2 · 2026-11-13