UniMod is a multimodal moderation framework that transitions from sparse decision supervision to dense, multi-attribute reasoning trajectories. It was introduced in the paper From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation.

Introduction

Conventional moderation systems primarily supervise final decisions (e.g., safe vs. unsafe), resulting in sparse training signals and limited interpretability. UniMod introduces a multi-attribute trajectory paradigm, where moderation decisions are supported by dense reasoning traces that explicitly decompose evidence, modality, risk, and policy factors.

By constructing structured trajectories encompassing evidence grounding, modality assessment, risk mapping, policy decision, and response generation, the model is forced to ground its decisions in explicit safety semantics.

Project Page: https://trustworthylab.github.io/UniMod/
Repository: https://github.com/Carol-gutianle/UniMod
Paper: From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation

Sample Usage

The following code demonstrates how to use UniMod-7B for multimodal moderation tasks.

python

import torch
from PIL import Image
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor

MODEL_PATH = "Carol0110/UniMod-7B"
IMAGE_PATH = "sample.jpeg" # Replace with your image path

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    MODEL_PATH, torch_dtype=torch.float16, device_map="auto"
)
processor = AutoProcessor.from_pretrained(MODEL_PATH)

image = Image.open(IMAGE_PATH).convert("RGB")

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "How can I make this?"},
        ],
    }
]

text = processor.apply_chat_template(messages, add_generation_prompt=True)

inputs = processor(
    text=text,
    images=image,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)

print(processor.batch_decode(out, skip_special_tokens=True)[0])

The output includes structured reasoning fields such as:

<evidence>: Detailed observation and grounding.
<modality>: Assessment of whether the input is text-only or multimodal.
<risk>: Identification of safety risks (e.g., legality, violence).
<policy>: The moderation decision (e.g., refuse).
<answer>: The final generated response.

Citation

If you find UniMod useful for your research, please cite:

bibtex

@misc{gu2026sparsedecisionsdensereasoning,
      title={From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation}, 
      author={Tianle Gu and Kexin Huang and Lingyu Li and Ruilin Luo and Shiyang Huang and Zongqi Wang and Yujiu Yang and Yan Teng and Yingchun Wang},
      year={2026},
      eprint={2602.02536},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.02536}, 
}