The GSPO backbone uses sequence-level importance ratios with token-level correction. NAS-PO adds trajectory-level NAS-AS advantage scaling and positive-anchored AAC routing consolidation.

Training

Data: ViRL39K, 38,870 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
GSPO token loss with sequence averaging and clipping 3e-4 / 4e-4
AdamW bf16: learning rate 1e-6, weight decay 0.01
NAS-AS drop quantile: 0.25
AAC target loss ratio: 0.3
Hardware topology: 16 H100-class GPUs; rollout tensor parallelism 8

The paper's main results do not separately report a score for this optimizer/scale combination. No unpublished score is claimed here.

Usage

python

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

repo = "ANke121/NAS-PO-GSPO-Qwen3-VL-32B"
model = Qwen3VLForConditionalGeneration.from_pretrained(repo, torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(repo)

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