VLM-CapCurriculum-InternVL3-8B-Staged

VLM-CapCurriculum-InternVL3-8B-Staged

A vision-language model post-trained from OpenGVLab/InternVL3-8B-hf with the staged, capability-dimension curriculum from "From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models" (ICML 2026).

This release is the Stage-3 step-186 checkpoint, which gave the largest gain over the merged baseline among the four backbones we tried.

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.

Headline numbers (extended benchmark suite, AVG over 10 benchmarks)

Δ over merged: +3.77 — the largest staged-vs-merged gap among the four backbones. Most striking is WeMath (+9.90 over merged), evidence that decoupling perception and reasoning is especially impactful for weaker base models. See Appendix Table 9 of the paper for the full per-benchmark breakdown.

How it was trained

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

1. Stage 1 — visual perception on UCSC-VLAA/VLM-CapCurriculum-Perception (synthesised + filtered DOCCI MCQs).
2. Stage 2 — textual reasoning on UCSC-VLAA/VLM-CapCurriculum-TextReasoning (ORZ-Math-13k).
3. Stage 3 — visual reasoning on UCSC-VLAA/VLM-CapCurriculum-VisualReasoning (CLEVR-Math + GeoQA170K + Math PUMA + DocVQA + ArxivQA mix). Released checkpoint: step 186 of Stage 3.

InternVL3 needs damped optimisation in Stage 2 to avoid entropy explosion (lr=3e-7, kl=5e-2, clip=0.15, max_grad_norm=0.5) — these are baked into the launch scripts.

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

Detailed launch scripts: training/examples/internvl3_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 LMDeploy (the InternVL family is served via LMDeploy in our setup):

lmdeploy serve api_server UCSC-VLAA/VLM-CapCurriculum-InternVL3-8B-Staged \
  --server-port 23342 --tp 4

For VLMEvalKit-style benchmark eval, plug it in via the InternVL3_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 InternVL3-8B backbone; we have not added additional alignment fine-tuning. Not recommended for high-stakes deployments without further evaluation.

Trained at the 8B parameter scale with 4096-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:

@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 InternVL family.