FuseO1-Preview is our initial endeavor to enhance the System-II reasoning capabilities of large language models (LLMs) through innovative model fusion techniques. By employing our advanced SCE merging methodologies, we integrate multiple open-source o1-like LLMs into a unified model. Our goal is to incorporate the distinct knowledge and strengths from different reasoning LLMs into a single, unified model with strong System-II reasoning abilities, particularly in mathematics, coding, and science domains.
To achieve this, we conduct two types of model merging:
Long-Long Reasoning Merging: This approach involves model fusion across LLMs that utilize long-CoT reasoning, with the goal of enhancing long-CoT reasoning capabilities. The resulted FuseAI/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview achieves a Pass@1 accuracy of 74.0 on AIME24, demonstrating significant performance improvements compared to the OpenAI o1-preview (44.6) and OpenAI o1-mini (63.4), even approaching OpenAI o1 (79.2).
Long-Short Reasoning Merging: This approach involves model fusion between long-CoT and short-CoT LLMs, aiming to improve reasoning capabilities in both long and short reasoning processes. The resulted FuseAI/FuseO1-DeepSeekR1-Qwen2.5-Instruct-32B-Preview is capable of utilizing both long and short reasoning processes and demonstrates relatively strong performance in long reasoning tasks.
1cd FuseAI/FuseO1-Preview/mergekit
2pip3 install -e .
3model_save_dir=xx # your path to save the merged models
4mergekit-yaml fuseo1_configs/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview.yaml ${model_save_dir}/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview --cudas
1cd FuseAI/FuseO1-Preview/mergekit
2pip3 install -e .
3model_save_dir=xxx # your path to save the merged models
4mergekit-yaml fuseo1_configs/FuseO1-DeepSeekR1-QwQ-32B-Preview.yaml ${model_save_dir}/FuseO1-DeepSeekR1-QwQ-32B-Preview --cuda
We provide the example code to use FuseAI/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview.
python3
1from vllm import LLM, SamplingParams
23llm = LLM(model="FuseAI/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview", tensor_parallel_size=8)
4sampling_params = SamplingParams(max_tokens=32768, temperature=0.7, stop=["<|im_end|>", "<｜end▁of▁sentence｜>"], stop_token_ids=[151645, 151643])
56conversations = [
7 [
8 {"role": "system", "content": "Please reason step by step, and put your final answer within \\boxed{{}}."},
9 {"role": "user", "content": "Quadratic polynomials $P(x)$ and $Q(x)$ have leading coefficients $2$ and $-2,$ respectively. The graphs of both polynomials pass through the two points $(16,54)$ and $(20,53).$ Find $P(0) + Q(0).$."},
10 ],
11]
1213responses = llm.chat(messages=conversations, sampling_params=sampling_params, use_tqdm=True)
1415for response in responses:
16 print(response.outputs[0].text.strip())
Long-Short Reasoning Merging
We conduct experiments on these folloing long-cot and short-cot LLMs.
1cd FuseAI/FuseO1-Preview/mergekit
2pip3 install -e .
3model_save_dir=xxx # your path to save the merged models
4mergekit-yaml fuseo1_configs/FuseO1-DeepSeekR1-Qwen2.5-Instruct-32B-Preview.yaml ${model_save_dir}/FuseO1-DeepSeekR1-Qwen2.5-Instruct-32B-Preview --cuda
1cd FuseAI/FuseO1-Preview/mergekit
2pip3 install -e .
3model_save_dir=xxx # your path to save the merged models
4mergekit-yaml fuseo1_configs/FuseO1-DeepSeekR1-Qwen2.5-Coder-32B-Preview.yaml ${model_save_dir}/FuseO1-DeepSeekR1-Qwen2.5-Coder-32B-Preview --cuda
We provide the code to use FuseAI/FuseO1-DeepSeekR1-Qwen2.5-Instruct-32B-Preview.
python3
1from vllm import LLM, SamplingParams
23llm = LLM(model="FuseAI/FuseO1-DeepSeekR1-Qwen2.5-Instruct-32B-Preview", tensor_parallel_size=8)
4sampling_params = SamplingParams(max_tokens=32768, temperature=0.7, stop=["<|im_end|>", "<｜end▁of▁sentence｜>"], stop_token_ids=[151645, 151643])
56conversations = [
7 [
8 {"role": "system", "content": "Please reason step by step, and put your final answer within \\boxed{{}}."},
9 {"role": "user", "content": "Quadratic polynomials $P(x)$ and $Q(x)$ have leading coefficients $2$ and $-2,$ respectively. The graphs of both polynomials pass through the two points $(16,54)$ and $(20,53).$ Find $P(0) + Q(0).$."},
10 ],
11]
1213responses = llm.chat(messages=conversations, sampling_params=sampling_params, use_tqdm=True)
1415for response in responses:
16 print(response.outputs[0].text.strip())
Evaluation Results
We test the resulted models on three kinds of benchmarks, including Math Reasoning, Code Reasoning , and Scientific Reasoning.
Math Reasoning
AIME24
MATH500
OlympiadBench
Scientific Reasoning
GPQA-Diamond
MMLU-Pro
MMLU
Code Reasoning
LiveCodeBench (2408-2502)
Important Note: We manully set "add_bos_token": false in tokenizer_config.json for all the evaluated LLMs to prevent the bos_token to be added twice for each prompt. Please download and modify to ensure consistency.
Math Reasoning
The evaluation code is modified from Qwen2.5-Math. In our evaluation, we set the temperature to 0.6, the top-p to 0.95 and the max_tokens to 32768. We provide the example to reproduce our results in math_evaluation.
The system prompt for evaluation is set to:
Please reason step by step, and put your final answer within \\boxed{{}}.
The evaluation results are shown in the table below:
In our evaluation of AIME24, we follow the method from DeepSeek-R1, wherein Pass@1 is computed by averaging the results across 32 sampled responses per prompt, while Cons@32 is determined through self-consistency analysis of the same 32 sampled responses for each prompt. For other benchmarks, we only sample 1 response and report the Pass@1.
We show that our merged FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview demonstrate superior performance improvements comparet to DeepSeek-R1-Distill-Qwen-32B, QwQ-32B-Preview, and Sky-T1-32B-Preview on math reasoning. Specifically, our model achieves an accuracy of 74.0 Pass@1 and 86.7 Cons@32 on AIME24, demonstrating significant performance improvements compared to DeepSeek-R1-Distill-Qwen-32B (69.2 Pass@1 and 83.3 Cons@32), OpenAI o1-preview (44.6 Pass@1) and OpenAI o1-mini (63.4 Pass@1), even approaching OpenAI o1 (79.2 Pass@1).
Scientific Reasoning
The evaluation code is modified from SkyThought. In our evaluation, we set the temperature to 0.7 and the max_tokens to 32768. We provide the example to reproduce our results in evaluation.
The system prompt for evaluation is set to:
You are a helpful and harmless assistant. You should think step-by-step.
The evaluation results are shown in the table below:
We show that our merged FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview demonstrate superior performance improvements comparet to DeepSeek-R1-Distill-Qwen-32B, QwQ-32B-Preview, and Sky-T1-32B-Preview on scientific reasoning. Specifically, our model achieves an accuracy of 62.1 on GPQA-Diamond and 70.8 on MMLU-Pro, demonstrating significant performance improvements compared to DeepSeek-R1-Distill-Qwen-32B (57.6 on GPQA-Diamond and 68.7 on MMLU-Pro).
Code Reasoning
The evaluation code is modified from Qwen2.5-Coder. In our evaluation, we set the temperature to 0.6, the top-p to 0.95 and the max_tokens to 32768. We provide the example to reproduce our results in code_evaluation.
The system prompt for evaluation is set to:
A conversation between User and Assistant. The user asks a question, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer. The reasoning process and answer are enclosed within <think> </think> and <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think> <answer> answer here </answer>.
In our evaluation of LiveCodeBench, we follow the method from DeepSeek-R1 and make a slight modification. The Pass@1 is computed by averaging the results across 16 sampled responses per prompt.
The evaluation results are shown in the table below:
We show that our merged FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview demonstrate superior performance improvements comparet to DeepSeek-R1-Distill-Qwen-32B, QwQ-32B-Preview, and Sky-T1-32B-Preview on scientific reasoning. Specifically, our model achieves an accuracy of 57.9 on LiveCodeBench and 25.5 on LiveCodeBench-Hard, demonstrating significant performance improvements compared to DeepSeek-R1-Distill-Qwen-32B (56.1 on LiveCodeBench and 23.4 on LiveCodeBench-Hard), OpenAI o1-preview (42.7 on LiveCodeBench and 9.8 on LiveCodeBench-Hard) and OpenAI o1-mini (52.0 on LiveCodeBench and 19.5 on LiveCodeBench-Hard Pass@1).
Future Works
This work is our first attempt effort to achieve knowledge fusion of System-II reasoning LLMs through a model merging approach, which is limited to LLMs with identical scale and architecture. In future work, we plan to employ our explicit model fusion method, based on multi-teacher knowledge distillation, and our implici model fusion method, which utilizes weighted-reward preference optimization for LLMs with different scales and architectures.
Furthermore, we intend to explore the combination of knowledge fusion with reinforcement learning (RL) methods, which have been demonstrated as the most effective approach for enhancing reasoning abilities. Stay tuned for the next version of FuseO1!
Citations
@article{wan2024fusechat,
title={Fusechat: Knowledge fusion of chat models},
author={Wan, Fanqi and Zhong, Longguang and Yang, Ziyi and Chen, Ruijun and Quan, Xiaojun},
journal={arXiv preprint arXiv:2408.07990},
year={2024}
}
