mgsm_cot: by models

Home Paper Code


SE predicted by accuracy

The typical standard errors between pairs of models on this dataset as a function of the absolute accuracy.

SE vs accuracy difference

The standard error of each model pair against their observed accuracy difference. Pairs below a reference line differ by more than the corresponding number of standard errors, i.e. they are statistically distinguishable at that level.

CDF of question level accuracy

Results table by model

model pass1 pass@count win_rate count SE(A) SE_x(A) SE_pred(A)
google_gemma_3_27b_it 89.2 93.1 43.5 3 0.59 0.5 0.32
qwen3-32b 89 92.6 43.3 2.3 0.6 NaN NaN
google_gemma_3_12b_it 86.2 91.4 41.2 3.4 0.66 NaN NaN
deepseek_r1_distill_llama_70b 85.8 91.7 41.2 2.2 0.67 NaN NaN
qwen3-14b 84.4 90.1 39.6 3 0.69 0.58 0.38
google_gemma_2_27b_it 82.5 82.5 38.8 1 0.73 NaN NaN
qwen3-8b 80.1 87.8 36.5 3 0.76 0.62 0.44
qwen2-72b-instruct 79.4 85.7 35.9 2.2 0.77 NaN NaN
deepseek_r1_distill_qwen_32b 79.4 86.1 36.1 2.2 0.77 NaN NaN
google_gemma_2_9b_it 77.8 87.5 35.5 3 0.79 0.62 0.49
qwen2.5-coder-32b-instruct 77.3 84.2 34.6 2.2 0.8 NaN NaN
llama-3.1-70B-instruct 77 77 35.8 4 0.8 0.8 0
qwen3-4b 73.7 83.1 32.1 4 0.84 0.71 0.44
qwen2-math-72b-instruct 73.7 83.4 32.7 2.2 0.84 NaN NaN
deepseek_r1_distill_qwen_14b 73.7 84 32.3 3.4 0.84 NaN NaN
google_gemma_3_4b_it 71.2 85.1 31.3 4 0.86 0.66 0.55
mistralai_mixtral_8x22b_instruct_v0.1 68 78.7 28.9 2.3 0.89 NaN NaN
qwen2.5-coder-14b-instruct 64.9 79.6 26.9 3 0.91 0.67 0.61
qwen1.5-72b-chat 63.7 74.7 26 2.2 0.92 NaN NaN
qwen1.5-32b-chat 58.6 70.7 23.5 2.3 0.94 NaN NaN
deepseek_r1_distill_qwen_7b 58 73.9 22.9 3.4 0.94 NaN NaN
qwen3-1.7b 54.9 70.2 21.1 4 0.95 0.77 0.56
qwen2-7b-instruct 52.6 70.1 19.9 3.4 0.95 NaN NaN
mistralai_ministral_8b_instruct_2410 51 69.1 19.2 3.4 0.95 NaN NaN
mistralai_mathstral_7b_v0.1 50.9 72.9 20.1 3.4 0.95 NaN NaN
qwen2.5-coder-7b-instruct 50.8 69.1 18.9 3.4 0.95 NaN NaN
mistralai_mixtral_8x7b_instruct_v0.1 44.4 64.4 16.7 3 0.95 0.64 0.7
llama-3.1-8B-instruct 44 44 16.8 7 0.95 0.95 0
qwen1.5-14b-chat 43.5 61.5 15.6 3 0.95 0.67 0.66
qwen2-math-7b-instruct 43.3 65.1 15.4 4 0.94 0.68 0.66
deepseek_r1_distill_llama_8b 40.4 62.6 14.2 4 0.94 0.66 0.67
google_codegemma_1.1_7b_it 38.4 62.4 14 4 0.93 0.63 0.68
deepseek_v2_lite_chat 37.2 52.5 13 2.3 0.92 NaN NaN
llama-3.2-3B-instruct 35.2 35.2 12.6 10 0.91 0.91 0
qwen2.5-coder-3b-instruct 34.3 57.6 11.5 4 0.91 0.61 0.67
qwen1.5-7b-chat 29.4 46.6 9.3 3 0.87 0.59 0.63
deepseek_r1_distill_qwen_1.5b 27.2 46 8.45 4 0.85 0.61 0.59
google_gemma_3_1b_it 26 45.5 8.38 4 0.84 0.6 0.59
mistralai_mistral_7b_instruct_v0.3 24.5 44.7 7.91 3.4 0.82 NaN NaN
mistralai_mistral_7b_instruct_v0.2 23.9 43.2 7.99 3.4 0.81 NaN NaN
qwen3-0.6b 21.9 39.9 6.84 4 0.79 0.55 0.57
qwen2-math-1.5b-instruct 20 36.6 5.99 4 0.76 0.54 0.54
google_gemma_7b_it 17.8 33.2 5.79 4 0.73 0.51 0.52
qwen2.5-coder-1.5b-instruct 17 36.1 4.89 4 0.72 0.43 0.57
mistralai_mistral_7b_instruct_v0.1 12.1 26.3 3.48 3.4 0.62 NaN NaN
qwen2-1.5b-instruct 8.15 21.9 2.3 4 0.52 0.26 0.46
google_gemma_2b_it 5.05 13.9 1.89 4 0.42 0.21 0.36
qwen2.5-coder-0.5b-instruct 3.46 10.7 1.08 4 0.35 0.14 0.32
qwen1.5-1.8b-chat 3.37 8.44 1.09 3 0.34 0.13 0.32
qwen2-0.5b-instruct 2.97 9.67 0.949 4 0.32 0.12 0.3
qwen1.5-0.5b-chat 1.42 5.09 0.6 4 0.23 0.055 0.22