A vision-language model post-trained from Qwen/Qwen3-VL-8B-Instruct with the staged, capability-dimension curriculum from "From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models" (ICML 2026).

TL;DR. Visual perception — not reasoning length — is the dominant bottleneck for visual reasoning in VLMs. We fix this by post-training along a capability axis (perception → textual reasoning → visual reasoning) rather than mixing all data together.

Resource	Link
📄 Paper	https://arxiv.org/abs/2605.20177
💻 Code	https://github.com/UCSC-VLAA/VLM-CapCurriculum
🌐 Project page	https://ucsc-vlaa.github.io/VLM-CapCurriculum
🤗 Collection (model + data + eval)	https://huggingface.co/collections/UCSC-VLAA/vlm-capcurriculum-from-seeing-to-thinking-icml-2026-6a07691f944148ccb2b183b8

Headline numbers

Setting	Visual Math AVG	Perception AVG	Overall AVG
Qwen3-VL-8B (base)	45.17	79.21	62.19
Qwen3-VL-8B + Merged training	49.64	79.71	64.67
Qwen3-VL-8B + Staged (this model)	51.10	80.44	65.77
OneThinker-8B (concurrent baseline)	51.10	78.64	64.87

Visual math = MathVista / MathVision / MathVerse(VI) / WeMath. Perception = A-OKVQA / RealWorldQA / MMStar / POPE.

Compared with the merged baseline on the same backbone, this model also produces 20.8% shorter reasoning traces — better perception lets the model think less.

How it was trained

Three RLVR stages with GRPO (on top of EasyR1):

Stage 1 — visual perception on UCSC-VLAA/VLM-CapCurriculum-Perception (synthesised + filtered DOCCI MCQs).
Stage 2 — textual reasoning on UCSC-VLAA/VLM-CapCurriculum-TextReasoning (ORZ-Math-13k).
Stage 3 — visual reasoning on UCSC-VLAA/VLM-CapCurriculum-VisualReasoning (CLEVR-Math + GeoQA170K + Math PUMA + DocVQA + ArxivQA mix).

All three stages share one system / format prompt — see Inference below.

Detailed launch scripts: training/examples/qwen3_vl_8b/ in the code repo.

Inference

The model expects the unified system prompt that it was trained against:

You FIRST think about the reasoning process as an internal monologue and then
provide the final answer. The reasoning process MUST BE enclosed within
<think> </think> tags. The final answer MUST BE put in \boxed{}.
i.e. <think> reasoning here </think> \boxed{final answer here}

Quick start with vLLM:

bash

vllm serve UCSC-VLAA/VLM-CapCurriculum-Qwen3-VL-8B-Staged \
  --tensor-parallel-size 4 --gpu-memory-utilization 0.9 --port 23341

Then send chat completions including the system prompt above.

For VLMEvalKit-style benchmark eval, plug it in via the Qwen3_VL_8B_Staged alias defined in evaluation/configs/models.py.

Intended use & limitations

Intended for research on vision-language reasoning, post-training methodology, and capability-dimension curriculum learning. Inherits the safety / bias profile of the underlying Qwen3-VL-8B-Instruct backbone; we have not added additional alignment fine-tuning. Not recommended for high-stakes deployments without further evaluation.

The model was trained at the 8B parameter scale with 2048-token max prompt length and a fixed group size of 5. Behaviour at much longer contexts or substantially different prompt formats has not been characterised.

License & citation

Released under Apache-2.0, matching the upstream backbone. If you use this model, please cite:

bibtex

@inproceedings{vlmcapcurriculum2026,
  title  = {From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models},
  author = {Juncheng Wu and Hardy Chen and Haoqin Tu and Xianfeng Tang and Freda Shi and Hui Liu and Hanqing Lu and Cihang Xie and Yuyin Zhou},
  booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
  year   = {2026},
  eprint = {2605.20177},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url    = {https://arxiv.org/abs/2605.20177}
}

Acknowledgements

Built on top of EasyR1, VLMEvalKit, and the Qwen3-VL family.