README.md · antgroup/HumanSense_Omni

metadata

license: apache-2.0
datasets:
  - antgroup/HumanSense_Benchmark
language:
  - en
metrics:
  - accuracy
library_name: transformers
base_model:
  - Qwen/Qwen2.5-Omni-7B
pipeline_tag: video-text-to-text

Zheng Qin¹, Ruobing Zheng^*², Yabing Wang¹, Tianqi Li², Yi Yuan², Jingdong Chen², Le Wang^†¹

^*Co-first authors. Project Lead. ^†Corresponding Author.
¹Xi’an Jiaotong University. ²Ant Group.

While Multimodal Large Language Models (MLLMs) show immense promise for achieving truly human-like interactions, progress is hindered by the lack of fine-grained evaluation frameworks for human-centered scenarios, encompassing both the understanding of complex human intentions and the provision of empathetic, context-aware responses. Here we introduce HumanSense, a comprehensive benchmark designed to evaluate the human-centered perception and interaction capabilities of MLLMs, with a particular focus on deep understanding of extended multimodal contexts and the formulation of rational feedback. Our evaluation reveals that leading MLLMs still have considerable room for improvement, particularly for advanced interaction-oriented tasks. Supplementing visual input with audio and text information yields substantial improvements, and Omni-modal models show advantages on these tasks. Furthermore, we argue that appropriate feedback stems from a contextual analysis of the interlocutor's needs and emotions, with reasoning ability serving as the key to unlocking it. Accordingly, we devise a multi-stage, modality-progressive reinforcement learning approach, resulting in HumanSense-Omni-Reasoning, which substantially enhances performance on higher-level understanding and interactive tasks. Additionally, we observe that successful reasoning processes exhibit highly consistent thought patterns. By designing corresponding prompts, we also enhance the performance of non-reasoning models in a training-free manner.

Release

2025-08-27 :hearts: We release both the training code and dataset!
2025-08-27 :hearts: We released Benchmark and code!
2025-08-15 :rocket: We released our paper!

Quickstart

Below, we provide simple examples to show how to use HumanSense_Omni_Reasoning with 🤗 Transformers.

pip uninstall transformers
pip install transformers==4.52.0
pip install accelerate
pip install qwen-omni-utils
pip install qwen-omni-utils[decord] -U

import torch
from transformers import Qwen2_5OmniForConditionalGeneration, Qwen2_5OmniProcessor
from qwen_omni_utils import process_mm_info


model_path = "antgroup/HumanSense_Omni_Reasoning"


model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="flash_attention_2",
    )
model.disable_talker()
processor = Qwen2_5OmniProcessor.from_pretrained(model_path)

conversation = [
    {
        "role": "user",
        "content": [
                    {
                        "type": "video",
                        "video": "file:///path/to/xxx.mp4",
                        "max_pixels": 151200
                    },
                    {
                        "type": "text",
                        "text": "xxxxxxxxxxxxxxxxxx\n"
                        }
                ],
    }
]

USE_AUDIO_IN_VIDEO=True

text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
audios, images, videos = process_mm_info(conversation, use_audio_in_video=USE_AUDIO_IN_VIDEO)
inputs = processor(text=text, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True,padding_side="left",add_special_tokens=False, use_audio_in_video=USE_AUDIO_IN_VIDEO)
inputs = inputs.to(model.device).to(model.dtype)

# Inference: Generation of the output text and audio
text_ids = model.generate(**inputs,return_audio=False, use_audio_in_video=USE_AUDIO_IN_VIDEO)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, text_ids)
]
text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)



response = text[0]
print('*'*30)
print(response)

Examples of Reasoning:

These cases cover four high-level perception and interaction tasks, including both video-based and audio-based questions. The reasoning processes all demonstrate thinking that integrates characteristics, emotions, and context, and then provides appropriate feedback.

BibTeX:

@article{qin2025humansense,
  title={HumanSense: From Multimodal Perception to Empathetic Context-Aware Responses through Reasoning MLLMs},
  author={Qin, Zheng and Zheng, Ruobing and Wang, Yabing and Li, Tianqi and Yuan, Yi and Chen, Jingdong and Wang, Le},
  journal={arXiv preprint arXiv:2508.10576},
  year={2025}
}