VERSE-3B

VERSE-3B

VERSE-3B is a 3B-scale vision-language checkpoint from VERSE (Spatial Self-Evolve Vision Language Agents). It starts from Qwen2.5-VL-3B-Instruct and is trained with the VERSE self-evolve GRPO stage, including an engine-verified VSI object-grounding process reward. This release is the merged step-100 policy.

VERSE teaches a VLM to draw on images (Object Mapper / Path Tracer) and then self-evolve: the model proposes spatial questions over 3D contexts, a deterministic engine verifies each question and supplies ground truth, the model solves them with agentic think→draw→observe reasoning, and RL updates on the result.

Target benchmarks: VSI-Bench, SPAR-Bench, MAZE.

Files

Merged Hugging Face weights (model-00001-of-00002.safetensors + model-00002-of-00002.safetensors) plus tokenizer / processor configs.

Usage

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "Aeronix-zzz/VERSE-3B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="bfloat16", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

Training

• Base model: Qwen2.5-VL-3B-Instruct
• Stage: self-evolve GRPO with grounding process reward
• Checkpoint: merged global_step_100

License

Weights follow the Qwen2.5-VL / Apache-2.0 terms. Please also respect upstream licenses for ViLaSR, EasyR1/verl, and the evaluation benchmarks.