EAN-7B

EAN-7B is Qwen2.5-VL-7B-Instruct trained with EAN (Evidence-Ablated Negatives), the method of our NeurIPS 2026 paper Grounding Multimodal Reasoning with Evidence-Ablated Negatives.

Code: github.com/DawnLIU35/EAN | Models: EAN-3B · EAN-7B · EAN-8B

Method

RLVR rewards typically check only the final answer, not whether it came from the image or from language priors and dataset shortcuts. EAN masks the visual evidence a question needs, lets the current policy answer without it, and penalizes plausible wrong answers under the clean image. EAN-7B 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 EAN.

Model	Geo3K	MathVista	MathVerse	We-Math	DynaMath	MathVision	MMK12	MMMU-Pro	LogicVista	Avg.
Qwen2.5-VL-7B-Instruct	36.1	63.9	38.6	46.6	56.1	18.6	43.4	25.9	42.8	41.3
DAPO	45.5	66.2	69.2	68.4	65.8	29.0	81.9	37.0	45.8	56.5
EAN-7B	46.0	69.0	69.1	69.7	66.5	31.3	82.2	38.8	48.8	57.9

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/EAN-7B"
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))

License

Apache 2.0, the same as Qwen2.5-VL-7B-Instruct. Relative to the base model, the weights were changed by RL training with EAN, 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).