HKU seminar: ViT shortcut-patch artifacts and Edit-R1 RL reward model for image editing
YiMaTweets · x · 2026-11-19
Prof. Yizhou Yu (HKU IDS) presented two research lines at the HKU IDS Seminar Series:
- ViT artifact analysis: Artifacts in pretrained ViTs stem from 'lazy aggregation'—the model uses semantically irrelevant background patches as shortcuts for global semantics. Selectively integrating patch features into the CLS token substantially improves performance across 12 benchmarks.
- Edit-R1: Existing image editing reward models give overall scores without detailed checks, causing biased rewards. Edit-R1 builds a chain-of-thought verifier-based reasoning reward model (RRM), trained with Group Contrastive Preference Optimization (GCPO) using human pairwise preferences. Edit-RRM outperforms powerful VLMs as an editing-specific reward model, with scaling trends from 3B to 7B parameters.
More from Multimodal
- Turning Sonnet 3's glossolalia into hyperpop songs with Suno — repligate · 2026-11-20
- Dev builds pixel arcade shooter with Astra 6.2 and Scenario in a few rounds — AIandDesign · 2026-11-20
- Synthesia lands in Claude: agentic video generation inside chat — synthesiaIO · 2026-11-20
- AI cover of 'Kiss from a Rose' impresses past the first seconds of jank — rickasaurus · 2026-11-20
- Pure-vision retrieval tops ViDoRe V3 leaderboard: GLM agent + EVIE hits 70.01 nDCG@10 — CShorten30 · 2026-11-20
- Krea Bots beta draws 50,000+ applications; new testers get $100 in credits — angrypenguinPNG · 2026-11-20