Evidence-RL-9B (Answer-CED)

Official model checkpoint for Evidence-RL: Towards Evidence-intensive Visual Reasoning (NeurIPS 2026).

Model Overview

Base Model: Qwen/Qwen3.5-9B
Training Method: Post-trained using Gated GRPO with Answer-level Counterfactual Evidence Disentanglement (Answer-CED) reward.
Objective: Mitigating visual hallucinations by reinforcing reliance on localized visual evidence via counterfactual region interventions.

Quickstart

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

model_id = "hhj-ai/Evidence-RL-9B"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

image = Image.open("example.jpg").convert("RGB")
prompt = "Look at the image carefully and count the objects. Answer with just a number, without any additional text."

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": prompt},
        ],
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=256)

response = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)[0]
print(response)

Citation

@misc{huang2026evidencerlevidenceintensivevisualreasoning,
      title={Evidence-RL: Towards Evidence-intensive Visual Reasoning}, 
      author={Haojie Huang and Xinlei Yu and Chengming Xu and Zhangquan Chen and Cheng Yang and Qingdong He and Yu Yang and Jiangning Zhang and Xiaobin Hu},
      year={2026},
      eprint={2608.08021},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.08021}, 
}