NAS-PO-GRPO-Qwen3-VL-8B

NAS-PO-GRPO-Qwen3-VL-8B

Qwen3-VL-8B-Instruct trained for 202 RLVR steps with GRPO and NAS-PO: Native Attention-Strategy Policy Optimization for Vision-Language Models.

NAS-PO decomposes native final-layer attention into a visual-mass policy and a conditional visual policy. NAS-AS uses detached visual mass for trajectory-level advantage scaling, while AAC consolidates successful token-level visual routing through a positive-advantage JSD objective.

Training

• Data: ViRL39K, 38,870 multimodal reasoning examples
• Rollouts per prompt: 8
• Rollout / actor global batch: 384 / 128
• Maximum prompt / response: 4096 / 256 tokens
• Sampling: temperature 1.0, top-p 0.99
• Optimizer: AdamW bf16, learning rate 1e-6, weight decay 0.01
• Reference-policy KL: disabled
• NAS-AS: drop quantile 0.25, minimum scale 0.9, square-root excess
• AAC: JSD over q/rho, target loss ratio 0.3, coefficient range [0, 10]
• Hardware: 8 NVIDIA H100 80 GB GPUs

Evaluation

The paper reports 68.04 overall mean Accuracy@8 over eight configurations from seven out-of-domain multimodal reasoning benchmarks. Accuracy@8 is mean response accuracy, not pass@8.

Use the public NAS-PO evaluator with 8 samples, temperature 1.0, top-p 0.99, and 256 new tokens.

Usage

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "ANke121/NAS-PO-GRPO-Qwen3-VL-8B",
    torch_dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("ANke121/NAS-PO-GRPO-Qwen3-VL-8B")

This is a research checkpoint. Outputs may be incorrect; verify them before consequential use. Citation metadata will be added after the paper is public.