zhangkangning/VAD_for_Qwen3.5-4b

VAD for Qwen3.5-4B

This repository contains the merged inference checkpoint of Visual Attribution Distillation (VAD) built on Qwen3.5-4B.

This is the main 4B VAD model trained with the symmetric Jensen--Shannon divergence (JSD) objective. It is the global_step_80 checkpoint reported in the paper, with an unweighted six-benchmark average of 78.32 under the Vision-OPD official evaluation pipeline and GPT-OSS-120B judge.

Model description

Multimodal on-policy distillation normally transfers the complete correction proposed by a privileged teacher. That correction can mix visual evidence with language preferences, formatting behavior, and other teacher-specific effects. VAD instead asks which part of the correction is attributable to a controlled change in the relevant visual evidence.

At each student-generated prefix, VAD evaluates the same fixed teacher with the evidence present and removed. The resulting distributional response provides a signed proxy direction for visual support and refutation. VAD projects the teacher correction onto this direction, allocates separate budgets to supported and refuted candidate tokens, and reconstructs a target around the full-image student distribution. The reconstructed target supplies the primary JSD supervision, while the original privileged teacher contributes only a weak stability regularizer.

The auxiliary teacher views are used only during training. Inference requires a single image and uses the standard Qwen3.5-4B architecture.

Checkpoint identity

Item	Value
Base model	Qwen/Qwen3.5-4B
Internal run	VC-OPD-Qwen3.5-4B-BASVCtau08-fixedInitialTeacher-anchorEta01-norhogate-nocap-8gpu-bs96-vllm18944-eval10-2epoch-20260623_223705
Released checkpoint	global_step_80
Primary supervision	Symmetric JSD, $\alpha=0.5$
VAD target	Branch-separated, budgeted visual support and refutation
Positive-branch cap	$\tau_+=0.8$
Weak teacher-regularizer weight	$\lambda=0.1$
Projection stabilizer	$\zeta=10^{-3}$
Coordinate shift bound	$c=20$
Teacher	Frozen copy of the initial Qwen3.5-4B student
Training schedule	2 epochs / 130 steps; checkpoint selected at step 80
Learning rate	$2\times10^{-6}$
Train batch size	96
Rollouts per prompt	8
Maximum prompt / response length	8192 / 1024 tokens
Training context length	18,944 tokens
Training GPUs	8
Random seed	42

Training used 6,241 synthetic visual question-answering examples. Each example contains a full image, an evidence-present crop, and a spatially matched evidence-degraded crop. The primary token-level target is evaluated over the student top-100 support plus a tail bucket. During training, VStar, ZoomBench, HRBench-4K, and HRBench-8K were monitored every 10 optimization steps.

Evaluation

We evaluated this checkpoint using the Vision-OPD official inference and accuracy pipeline with openai/gpt-oss-120b as judge. Inference used max_model_len=32768, MAX_TOKENS=8192, ENABLE_THINKING=False, and seed 42. The judge used temperature 0. All answer and judge files passed the completeness audit, with no non-canonical final judge labels.

Benchmark	Accuracy
VStar	92.15
ZoomBench	60.59
HRBench-4K	85.38
HRBench-8K	83.38
MME-RealWorld-EN	76.97
MME-RealWorld-CN	71.47
Unweighted average	78.32

These results are specific to this checkpoint, benchmark versions, inference configuration, and judge protocol.

Serving with vLLM

Use a recent vLLM release with Qwen3.5 support:

vllm serve zhangkangning/VAD_for_Qwen3.5-4b \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name VAD-Qwen3.5-4B \
  --trust-remote-code \
  --tensor-parallel-size 1 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.90

Example OpenAI-compatible multimodal request:

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="VAD-Qwen3.5-4B",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/image.jpg"},
                },
                {
                    "type": "text",
                    "text": "Describe the fine-grained visual details.",
                },
            ],
        }
    ],
    max_tokens=1024,
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

print(response.choices[0].message.content)

Replace the placeholder image URL with an accessible image URL or a supported data URL.

Intended use and limitations

• This checkpoint is intended for research on multimodal perception, fine-grained visual question answering, and on-policy distillation.
• Performance outside the reported visual benchmarks, including broad language capability, safety, bias, and robustness, has not been comprehensively evaluated.
• Like other vision-language models, it may produce incorrect or unsupported answers. Independently verify outputs in high-stakes settings.
• The model inherits the capabilities and limitations of Qwen3.5-4B.

License

The checkpoint is released under the Apache License 2.0, consistent with the base model. Users must also comply with the Qwen3.5-4B license.

Citation

The public paper identifier and final citation will be added when available. No provisional bibliographic entry is provided here.