NVILA-Lite-2B-DATA_SCALE_EXP_2M-20260205_003632

NVILA-Lite-2B — Data-Scale Fine-tune (2M)

This model is a fine-tune of Efficient-Large-Model/NVILA-Lite-2B on a 5-dataset spatial-reasoning mixture (SAT + RoboSpatial + SPAR-7M + RefSpatial + PRISM), with the training-sample budget set to 2M.

It was released as part of the paper "Why Far Looks Up: Probing Spatial Representation in Vision-Language Models". You can find other data-scale variants in the HF Collection.

Paper Summary

Vision-language models (VLMs) achieve strong performance on spatial reasoning benchmarks, yet it remains unclear whether this reflects structured 3D understanding or reliance on statistical shortcuts. This paper introduces a representation-level analysis framework using minimal contrastive pairs to measure spatial organization in VLM embeddings. The authors identify a consistent "vertical-distance entanglement" where models conflate vertical image position with distance, mirroring perspective biases in natural photography.

• 📄 Paper: arXiv:2605.30161
• 🌐 Project page: https://cheolhong0916.github.io/whyfarlooksup.github.io/
• 💻 Probing / evaluation code: https://github.com/cheolhong0916/contrastive-probing
• 📊 Diagnostic benchmark: cubec/spatialtunnel

Loading

This checkpoint inherits the architecture of Efficient-Large-Model/NVILA-Lite-2B. You can load it using the standard transformers library interface:

from transformers import AutoModelForVision2Seq, AutoProcessor

model_id = "ch-min/NVILA-Lite-2B-DATA_SCALE_EXP_2M-20260205_003632"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForVision2Seq.from_pretrained(model_id, trust_remote_code=True)

Citation

If you use this checkpoint, please cite both the probing paper and the original NVILA-Lite-2B paper.

Our paper (this checkpoint family):

@article{min2026whyfarlooksup,
  title   = {Why Far Looks Up: Probing Spatial Representation in Vision-Language Models},
  author  = {Min, Cheolhong and Jung, Jaeyun and Lee, Daeun and Jeon, Hyeonseong and
             Su, Yu and Tremblay, Jonathan and Song, Chan Hee and Park, Jaesik},
  journal = {arXiv preprint arXiv:2605.30161},
  year    = {2026},
}

Original NVILA-Lite-2B (NVILA: Efficient Frontier Visual Language Models):

@article{liu2024nvila,
  title   = {NVILA: Efficient Frontier Visual Language Models},
  author  = {Liu, Zhijian and Zhu, Ligeng and Shi, Baifeng and others},
  journal = {arXiv preprint arXiv:2412.04468},
  year    = {2024},
}