🚀 ReasonFlux: 階層的LLM推論における思考テンプレートの拡張
革新的なテンプレート拡張推論パラダイムにより、32Bモデルが推論タスクにおいてo1-miniやDeepSeek-R1蒸留モデルを上回る性能を発揮します。
タスク/Pass@1 |
ReasonFlux-F1-32B |
ReasonFlux-Zero-32B |
R1-Distill-32B |
o1-mini |
LIMO -32B |
s1-32B |
MATH500 |
96.0 |
91.2 |
94.3 |
90.0 |
90.6 |
93.0 |
AIME 2024 |
76.7 |
56.7 |
72.6 |
56.7 |
50.0 |
56.7 |
AIME 2025 |
53.3 |
37.2 |
46.67 |
50.8 |
37.2 |
49.3 |
GPQA-Diamond |
67.2 |
61.2 |
62.1 |
60.0 |
65.2 |
59.6 |
📦 モデル情報
プロパティ |
詳細 |
ライブラリ名 |
transformers |
ライセンス |
other |
ベースモデル |
deepseek-ai/DeepSeek-R1-Distill-Qwen-32B |
タグ |
llama-factory, full, generated_from_trainer |
モデル名 |
ReasonFlux-F1-32B |
✨ 主な機能
ReasonFlux-F1-32Bは、ReasonFlux-Zeroからのテンプレート拡張推論軌道を活用して微調整されたSOTAレベルの推論LLMです。
📊 評価
以下は、ReasonFlux-F1-32BのAIME2024、AIM2025、MATH500、GPQA-Diamondなどの難しい推論タスクでの評価結果です。公平な比較のため、ReasonFlux-F1の評価スクリプトを使用しています。
モデル |
AIME2024@pass1 |
AIME2025@pass1 |
MATH500@pass1 |
GPQA@pass1 |
QwQ-32B-Preview |
46.7 |
37.2 |
90.6 |
65.2 |
LIMO-32B |
56.3 |
44.5 |
94.8 |
58.1 |
s1-32B |
56.7 |
49.3 |
93.0 |
59.6 |
OpenThinker-32B |
66.0 |
53.3 |
94.8 |
60.1 |
R1-Distill-32B |
70.0 |
46.7 |
92.0 |
59.6 |
ReasonFlux-Zero-32B |
56.7 |
37.2 |
91.2 |
61.2 |
ReasonFlux-F1-32B |
76.7 |
53.3 |
96.0 |
67.2 |
💻 使用例
基本的な使用法
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
model_id = 'Gen-Verse/ReasonFlux-F1'
model = LLM(
model_id,
tensor_parallel_size=8,
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
sampling_params = SamplingParams(
max_tokens=32768,
)
question = """Let \(x, y\), and \(z\) be positive real numbers satisfying the system of equations:
\[
\begin{array}{c}
\sqrt{2 x-x y}+\sqrt{2 y-x y}=1 \\
\sqrt{2 y-y z}+\sqrt{2 z-y z}=\sqrt{2} \\
\sqrt{2 z-z x}+\sqrt{2 x-z x}=\sqrt{3} .
\end{array}
\]
Then \(\left[(1-x)(1-y)(1-z)\right]^{2}\) can be written as \(\frac{m}{n}\), where \(m\) and \(n\) are relatively prime positive integers. Find \(m+n\)."""
ds_prompt="<|User|>\n" + question + "<|Assistant|>\n"
output = model.generate(ds_prompt, sampling_params=sampling_params)
print(output[0].outputs[0].text)
📄 引用
@article{yang2025reasonflux,
title={ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates},
author={Yang, Ling and Yu, Zhaochen and Cui, Bin and Wang, Mengdi},
journal={arXiv preprint arXiv:2502.06772},
year={2025}
}