SafeWork-R1

📂 GitHub · 📜Technical Report · 💬Online Chat

Overview

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model demonstrating the coevolution of safety and general intelligence under the guiding principle of the AI-45° Law.

SafeWork-R1 is built upon the SafeLadder framework, which integrates large-scale, progressive, safety-oriented reinforcement learning post-training supported by multi-principled verifiers. Unlike conventional RLHF that simply learns human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, leading to emergent safety “aha” moments.

Model Zoo

Model Variant	Parameters	Base Model	Link
SafeWork-R1	72B	Qwen2.5-VL-72B	🤗 link
SafeWork-R1-InternVL3-78B	78B	InternVL3-78B	🤗 link
SafeWork-R1-DeepSeek-70B	70B	Deepseek-R1-DistillLlama-70B	🤗 link
SafeWork-R1-Qwen2.5VL-7B	7B	Qwen2.5-VL-7B	🤗 link

Performance

Safety Benchmarks

Model	MM-SafetyBench	MSSBench	XSTest-Safe	SIUO	Avg.
Gemini 2.5 pro	79.3	70.5	100.0	76.7	81.6
Claude Opus 4	82.1	59.6	96.8	62.8	75.3
GPT-4.1	78.2	69.1	96.4	92.9	84.1
GPT-4o	70.2	58.8	94.0	51.8	68.7
Qwen2.5-VL-72B	70.4	53.8	91.2	38.2	63.4
SafeWork-R1	92.0↑21.6	74.8↑21.0	99.2↑8.0	90.5↑52.3	89.2↑25.8

Value Benchmarks

Model	FLAMES	M³oralBench (Judge)	M³oralBench (Classification)	M³oralBench (Response)	Avg.
Gemini 2.5 Pro	16.8	70.0	66.2	86.8	44.7
Claude Opus 4	38.1	70.7	74.7	72.5	52.2
GPT-4.1	33.3	74.4	62.7	61.7	53.0
GPT-4o	36.6	72.4	65.9	79.7	55.5
Qwen2.5-VL-72B	39.1	58.4	48.1	75.7	49.9
SafeWork-R1	65.3↑26.2	68.1↑9.7	54.6↑6.5	70.9↓4.8	64.9↑15.0

General Benchmarks

Model	MMMU	MathVista	Olympiad	GPQA Diamond	GAOKAO-MM	Avg.
Gemini 2.5 Pro	82.0	83.0	81.8	86.9	87.2	84.2
Claude Opus 4	73.0	73.0	68.5	74.7	73.7	72.6
GPT-4.1	72.4	72.0	49.0	69.2	60.2	64.6
GPT-4o	70.6	61.6	33.7	46.9	33.8	49.3
Qwen2.5-VL-72B	67.2	74.8	40.4	50.5	73.1	61.2
SafeWork-R1	70.9↑3.7	76.1↑1.3	59.9↑19.5	59.6↑9.1	78.2↑5.1	68.9↑7.7

Quick Start

from transformers import AutoProcessor, AutoModelForCausalLM
import torch

model_name = "AI45Research/SafeWork-R1"
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype="auto", trust_remote_code=True)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "file:///path/to/image",
            },
            {"type": "text", "text": "Prompt containing harmful content."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=8192)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

License

This project is released under the Apache 2.0 license.

Citation

If you find this work useful, feel free to give us a cite.

@misc{lab2025safework,
  title={SafeWork-R1: Coevolving Safety and Intelligence under the AI-45 Law},
  author={Lab, Shanghai AI and Bao, Yicheng and Chen, Guanxu and Chen, Mingkang and Chen, Yunhao and Chen, Chiyu and Chen, Lingjie and Chen, Sirui and Chen, Xinquan and Cheng, Jie and others},
  journal={arXiv preprint arXiv:2507.18576},
  year={2025}
}