Model Intelligence Score

A model intelligence score is a composite metric designed to summarize overall model capability across multiple benchmarks (MMLU, GSM8K, MBPP, etc.).

  Example Weighting Scheme

Overall score = Σ weight × normalized benchmark.

  Design Trade-offs

• Simple weighted average easy to communicate but hides domain weaknesses.
• Too many benchmarks dilutes signal.
• Proprietary weights reduce transparency.

  Current Trends (2025)

• Adaptive weights adjust per user domain (code vs chat).
• Leaderboards publish raw scores alongside composite.
• Bootstrapped CIs show ±2 % noise intervals.

  Implementation Tips

1. Normalize each benchmark to 0–100 before weighting.
2. Publish YAML of weight config for reproducibility.
3. Update weights annually to reflect task importance shifts.