Self-Supervised Visual On-Policy Distillation Boosts Small VLMs
UCSanDiego · hf · 2026-08-17
UC San Diego proposes Self-Supervised Visual On-Policy Distillation. This method improves small vision-language models by distilling from original images into strongly augmented student views without privileged annotations or larger teachers.
More from Research
- IBM Open Sources Docling-Graph: Converting PDFs to Knowledge Graphs — aigleeson · 2026-08-17
- 何为 Benchmaxxing？AI 基准测试能否造假？ — Former-Towel9004 · 2026-08-17
- ZipSplat: Fewer Gaussians, Better 3D Scene Reconstruction — rsasaki0109 · 2026-08-17
- Petition calls for mandatory quantization labels in model评测 posts — Su1tz · 2026-08-17
- A Pathway to General-Purpose Scientific AI: New Multimodal Benchmark — SciKnowOrg · 2026-08-17
- Subjective Logic: A Foundational Book on Reasoning Under Uncertainty — FrnkNlsn · 2026-08-17