Pretrained ViTs provide general-purpose representations for diverse downstream tasks. However, artifacts in ViTs are widely observed across different supervision paradigms. In the first part of the seminar, I present a through analysis of artifacts in ViTs, and conclude that these artifacts originate from a lazy aggregation behavior: ViT uses semantically irrelevant background patches as shortcuts to represent global semantics, driven by global attention and coarse-grained semanticsupervision. Our solution selectively integrates patch features into the CLS token, and substantially improves performance across 12 benchmarks.

Existing image editing reward models usually give overall scores withoutdetailed checks, causing biased rewards. To address this, we move from a simple scorer to a reasoning verifier. In the second part of the seminar, I introduce Edit-R1, a framework that builds a chain-of-thought (CoT) verifier-based reasoning reward model (RRM) and then leverages it for image editing. To build such an RRM, we introduce Group Contrastive Preference Optimization (GCPO), a reinforcement learning algorithm that leverages human pairwise preference data to reinforce our pointwise RRM. Extensive experiments demonstrate that our Edit-RRM surpasses powerful VLMs as an editing-specific reward model with a scaling trend from 3B to 7B parameters.