AIAny
Icon for item

Visual Contrastive Self-Distillation

Converts image-content removal into a contrastive on-policy self-distillation signal: the EMA teacher produces next-token distributions with and without image content, uses their log-probability differences to sharpen visual-grounded candidates, and distills that full-distribution target into the student—no external teacher or extra inference cost.

Introduction

Why this matters Multimodal LLMs often struggle to assign probability specifically to tokens grounded in the visual input rather than generic, language-only completions. VCSD reframes that problem: instead of relying on an external teacher or privileged labels, it contrasts teacher outputs conditioned on the original image versus a content-erased control to highlight tokens whose likelihood is driven by instance-level visual content. That contrast becomes the training signal used to sharpen and distill a full-token distribution back into the student.

Key Findings
  • Contrastive on-policy self-distillation: At each student-generated prefix, an EMA teacher computes two next-token distributions—one with the original image and one with a content-erased control. The token-wise log-probability difference identifies candidates specifically boosted by the image, which are then used to reweight and sharpen the teacher distribution before distillation.
  • No external supervision or inference overhead: VCSD requires no external teacher model, privileged answers, extra visual-evidence annotations, reasoning traces, or additional inference-time cost for deployment.
  • Empirical gains on multimodal LLMs: On the ViRL39K dataset, VCSD consistently improves Qwen3-VL benchmarks: for example, 2B model aggregate 62.27% → 67.04%, 4B 71.30% → 73.16%, 8B 72.51% → 76.26% (reported by the authors), showing the method scales across model sizes.
Who it's for and trade-offs

Great fit if you are fine-tuning or distilling multimodal LLMs and need a simple, data-driven way to strengthen visual grounding without collecting extra labels or maintaining a stronger external teacher. It is especially relevant when you can generate or obtain content-erased control images and run EMA-teacher-based training loops. Look elsewhere if you cannot produce reliable content-erased controls for your images, need methods that explicitly output interpretable visual evidence, or if minimizing training-time complexity (EMA teacher bookkeeping and additional forward passes during training) is a higher priority than avoiding external supervision.

How it works (brief)

The core mechanism is contrastive: by subtracting the teacher's log-probabilities on a control image from those on the original image, VCSD highlights token candidates whose probability gain is attributable to instance-level visual content. The teacher's original-image distribution is then sharpened within its plausible support using that contrast signal, and the resulting full-distribution target is distilled into the student via standard distributional distillation losses. The pipeline was evaluated on ViRL39K and applied to Qwen3-VL and Qwen3.5-family models in the paper.

Information

  • Websitearxiv.org
  • AuthorsYijun Liang, Yunjie Tian, Yijiang Li, Yuqi Jia, Furong Huang, Tianyi Zhou, Di Fu
  • Published date2026/07/23

More Items

Evaluates spatial cognition of image-generation models by eliciting protocol-constrained visual answers and parsing pixel outputs into structured predictions compatible with existing metrics. Introduces the ProVisE framework and SpatialGen-Bench (470 samples) to compare image-generation models and text-output VLMs on unified spatial tasks.

Selects a referred target from candidate bounding boxes, then decodes tracking waypoints for single-camera embodied visual tracking. Injects past selected-bbox geometry via sliding-window TVBI tokens and is co-trained on a Refer‑QA dataset; achieves SOTA on EVT‑Bench and demonstrates sim-to-real on legged and humanoid robots.

Quantifies active visual observation in multimodal LLMs with ActiveVision, a 17-task benchmark that forces repeated perception rather than one-shot description. Finds frontier MLLMs fail badly (top model 10.6% vs humans 96.1%) and that model-generated vision code does not close the gap.