What is BiPS?
Many VLMs fail on multimodal reasoning because they look at the wrong visual evidence (especially for charts, thin lines, intersections, and small regions).
BiPS transforms visual cues into training-time policy constraints by constructing evidence-preserving and evidence-ablated views, enabling the model to internalize fine-grained perception without inference-time overhead. This formulation can also be interpreted from the perspective of visual on-policy distillation (OPD) / on-policy self-distillation, where information-asymmetric visual views provide dense policy-shaping signals for GRPO training.
Key idea
BiPS trains a VLM with two complementary view transformations:
-
Evidence-Preserving View: keep only the visual evidence needed to answer, reduce distractions.
→ enforce consistency between predictions from the original image and the preserved view.
-
Evidence-Ablated View: remove the key evidence so the image no longer supports the answer.
→ enforce separation so the model cannot rely on shortcuts.
These constraints are typically implemented with KL-based objectives and can be integrated into GRPO training.
Why it matters
- Better fine-grained evidence alignment
- Less “guessing” from language priors
- No additional inference overhead (views are used only during training)
How to use
BiPS is mainly a training recipe. To apply it:
- Follow the official repo to set up dependencies and scripts.
- Train your base VLM with BiPS-generated preserve/ablate views.
- Use the resulting checkpoint as a standard VLM at inference time (no extra steps).
Citation
@article{zhang2025bips,
title={See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning},
author={Zhang, Shuoshuo and Zhang, Yizhen and Fu, Jingjing and Song, Lei and Bian, Jiang and Yang, Yujiu and Wang, Rui},
journal={arXiv preprint arXiv:2512.22120},
year={2025}
}