EVAN-8B

EVAN-8B is Qwen3-VL-8B-Instruct trained with EVAN (EVidence-Ablated Negatives), the method of our NeurIPS 2026 paper Grounding Multimodal Reasoning with Evidence-Ablated Negatives.

Paper: arXiv:2610.39237 | Code: github.com/DawnLIU35/EVAN | Models: EVAN-3B · EVAN-7B · EVAN-8B

Method

RLVR rewards typically check only the final answer, not whether it came from the image or from language priors and dataset shortcuts. EVAN masks the visual evidence a question needs, lets the current policy answer without it, and penalizes plausible wrong answers under the clean image. EVAN-8B is trained with DAPO on ViRL39K. The GitHub repository has the full method, data pipeline, and training scripts.

Results

Accuracy (%) from Table 1 of the paper, with sampling at temperature 1.0. DAPO uses the same data and training recipe without EVAN.

Usage

Requires transformers>=4.57, accelerate, and torchvision. Add the instruction below after the question, as in training. The default generation config holds the sampling settings we used for evaluation.

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "DawnLiu/EVAN-8B"
model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)

question = "Which object is the largest?"
prompt = (
    f"{question}\n\n"
    "You first think through the reasoning process as an internal monologue, enclosed within <think> </think> tags. Then, provide your final answer enclosed within \\boxed{}."
)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "path/to/image.png"},
            {"type": "text", "text": prompt},
        ],
    }
]
inputs = processor.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt"
).to(model.device)
output = model.generate(**inputs)
print(processor.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Citation

@misc{liu2026grounding,
  title         = {Grounding Multimodal Reasoning with Evidence-Ablated Negatives},
  author        = {Chenxi Liu and Tianyi Xiong and Yanshuo Chen and Heng Huang},
  year          = {2026},
  eprint        = {2610.39237},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.39237},
}

License

Apache 2.0, the same as Qwen3-VL-8B-Instruct. Relative to the base model, the weights were changed by RL training with EVAN, generation_config.json sets our evaluation sampling defaults, and the other configuration, tokenizer, and processor files were re-saved (with dtype set to bfloat16 in config.json).