RGE

Reasoning Guided Embeddings (RGE)

Reasoning Guided Embeddings: Leveraging MLLM Reasoning for Improved Multimodal Retrieval.

Chunxu Liu, Jiyuan Yang, Ruopeng Gao, Yuhan Zhu, Feng Zhu, Rui Zhao, Limin Wang

The proposed model is trained based on Qwen2.5-VL-3B-Instruct.

💻 Code | 📄 Paper

TL; DR. We introduce Reasoning Guide Embedding (RGE) model, which takes advantage of MLLMs’ structured reasoning during embedding extraction, using generated rationales with contrastive training to produce more context-aware representations, improving embedding quality.

Eval data:

We use MMEB-eval for multimodal retrieval evaluation.

Experimental Results

<img width="900" alt="abs" src="https://raw.githubusercontent.com/MCG-NJU/RGE/refs/heads/main/assets/results.png">

Usage

Transformers

Below is an example we adapted from VLM2Vec and MoCa.

git clone https://github.com/MCG-NJU/RGE.git
cd RGE
pip install -r requirements.txt

from transformers import AutoProcessor, AutoConfig, Qwen2_5_VLForConditionalGeneration
import torch
from PIL import Image
import torch.nn.functional as F

def compute_similarity(q_reps, p_reps):
    return torch.matmul(q_reps, p_reps.transpose(0, 1))

def construct_prompt(text):
        return [
            {
                "role": "user",
                "content": [
                    {"type": "text", "text": text}
                ]
            }
        ]

@torch.no_grad()
def encode_and_generate(model, processor, inputs: dict):
    # baseline -> caption 
    generation_output = model.generate(
        **inputs,
        return_dict_in_generate=True,
        output_hidden_states=True,
        max_new_tokens=128,
        eos_token_id=processor.tokenizer.convert_tokens_to_ids("<emb>"),
        use_cache=True,
    )

    prompt_len = inputs['input_ids'].shape[1]
    generated_ids = generation_output.sequences[0][prompt_len:]
    generated_text = processor.tokenizer.decode(generated_ids, skip_special_tokens=False)
    print(f"generated_text: {generated_text}")

    past_seen = generation_output.past_key_values.get_seq_length()

    new_inputs = {
        'input_ids': generation_output.sequences[:, -1:],
        'past_key_values': generation_output.past_key_values,
        'cache_position': torch.arange(past_seen, past_seen + 1, device=generation_output.sequences.device),
    }

    outputs = model(**new_inputs, output_hidden_states=True, use_cache=True)

    pooled_output = outputs.hidden_states[-1][:, -1, :]
    return pooled_output, generated_text

@torch.no_grad()
def ar_encode(model, processor, inputs: dict):
    def _ar_pooling(input_ids, hidden_states):
        emb_id = processor.tokenizer.convert_tokens_to_ids("<emb>")
        embed_indices = torch.argmax((input_ids == emb_id).int(), dim=1)
        embed_features = hidden_states[torch.arange(len(embed_indices)), embed_indices]
        return embed_features
    outputs = model(**inputs, output_hidden_states=True)
    pooled_output = _ar_pooling(inputs['input_ids'], outputs.hidden_states[-1])
    return pooled_output
        
model_name = "lcxrocks/RGE"
processor_name = "lcxrocks/RGE"

# Load Processor and Model
processor = AutoProcessor.from_pretrained(processor_name)
config = AutoConfig.from_pretrained(model_name)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_name, config=config, 
    torch_dtype=torch.bfloat16,
    attn_implementation='flash_attention_2',
).to("cuda")
model.eval() 

string = "<|vision_start|><|image_pad|><|vision_end|> Answer the question based on the given image: What is in the image?\n"
instruction = processor.apply_chat_template(construct_prompt(string), tokenize=False, add_generation_prompt=True)
inputs = processor(text=instruction, images=[Image.open('assets/example.jpg')], return_tensors="pt").to("cuda")
qry_output, generated_text = encode_and_generate(model, processor, inputs)
qry_output = F.normalize(qry_output, p=2, dim=-1)

string = 'Summarize the following answer in one word: A cat and a dog.'
instruction = processor.apply_chat_template(construct_prompt(string), tokenize=False, add_generation_prompt=True) + "<emb>"
inputs = processor(text=instruction, images=None, return_tensors="pt").to("cuda")
tgt_output = ar_encode(model, processor, inputs)
tgt_output = F.normalize(tgt_output, p=2, dim=-1)
print(string, '=', compute_similarity(qry_output, tgt_output))
# tensor([[0.7031]], device='cuda:0', dtype=torch.bfloat16)

string = 'Summarize the following answer in one word: A cat and a tiger.'
instruction = processor.apply_chat_template(construct_prompt(string), tokenize=False, add_generation_prompt=True) + "<emb>"
inputs = processor(text=instruction, images=None, return_tensors="pt").to("cuda")
tgt_output = ar_encode(model, processor, inputs)
tgt_output = F.normalize(tgt_output, p=2, dim=-1)
print(string, '=', compute_similarity(qry_output, tgt_output))
# tensor([[0.6250]], device='cuda:0', dtype=torch.bfloat16)

Citation

If you use this model in your research, please cite the associated paper.

@misc{liu2025reasoningguidedembeddingsleveraging,
      title={Reasoning Guided Embeddings: Leveraging MLLM Reasoning for Improved Multimodal Retrieval}, 
      author={Chunxu Liu and Jiyuan Yang and Ruopeng Gao and Yuhan Zhu and Feng Zhu and Rui Zhao and Limin Wang},
      year={2025},
      eprint={2511.16150},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.16150}, 
}