emergence_summary.md 675 B

E3 能力涌现谱突变检测摘要

分析时间:2026-03-14 22:41

检测到的涌现点

能力 涌现模型 MMLU@涌现 跃变幅度 显著性
cot_reasoning gemma-2-9b-it 82.0% 3.225 2.56σ
arithmetic gemma-2-9b-it 82.0% 9.431 2.67σ
code_generation llama3.2-3b-instruct 58.0% 7.316 2.33σ
instruction_following llama3.2-3b-instruct 58.0% 9.399 2.81σ

模型序列 (按 MMLU 排序)

  1. llama3.2-3b-instruct (MMLU: 58.0%)
  2. Mistral-7B-v0.3 (MMLU: 62.5%)
  3. llama3-8b (MMLU: 68.4%)
  4. gemma-2-9b-it (MMLU: 82.0%)
  5. Qwen2.5-7B-Instruct (MMLU: 86.3%)