Overview
Vision-Language Models (VLMs) often answer from language priors and dataset shortcuts instead of actually looking at the image. Show a pair of glasses with one arm removed, and the model still says "two arms" because glasses usually have two. Evidence-RL-9B is post-trained with Counterfactual Evidence Disentanglement (CED), a counterfactual visual evidence reward that ensures each answer causally depends on the relevant object region, reducing hallucination and improving fine-grained multimodal visual reasoning.
How CED Works
During reinforcement learning post-training (GRPO), CED tests whether the model's answer truly depends on the visual evidence that supports it:
- Identifies an object-centric Evidence Region in the image
- Neutralizes that region in feature space (mean replacement) and measures the log-likelihood drop
- Compares this drop against matched non-evidence regions
Answers that are visually grounded, whose correctness causally depends on the right part of the image, earn higher reward. Answers that survive even without looking at the relevant region (relying on language priors or irrelevant context) are penalized. This counterfactual visual grounding reward uses only weak object-level proposals, requires no question-specific region annotations and adds zero inference-time overhead.
Evidence-RL consistently improves performance across nine benchmarks (counting, spatial reasoning, hallucination detection, fine-grained perception, and general multimodal reasoning) and four backbones, with the largest gains on perception-heavy tasks where language-prior shortcuts are most exposed.
Quickstart
from transformers import AutoModelForImageTextToText, AutoProcessorfrom PIL import Image model_id = "hhj-ai/Evidence-RL-9B"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained( model_id, torch_dtype="bfloat16", device_map="auto") image = Image.open("example.jpg")messages = [ {"role": "user", "content": [ {"type": "image", "image": image}, {"type": "text", "text": "How many wheels are visible in the image?"} ]}] inputs = processor.apply_chat_template(messages, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=256)print(processor.decode(outputs[0], skip_special_tokens=True))
Resources
Citation
@article{huang2026evidence, title={Evidence-RL: Towards Evidence-intensive Visual Reasoning}, author={Huang, Haojie and Yu, Xinlei and Xu, Chengming and Chen, Zhangquan and Yang, Cheng and He, Qingdong and Yang, Yu and Zhang, Jiangning and Hu, Xiaobin}, journal={arXiv preprint arXiv:2608.08021}, year={2026}}