Evaluation-Verification Reward for Consistent Multi-Reference Image Editing

  • 类型:arxiv
  • 标识:2607.29025
  • 链接:https://arxiv.org/abs/2607.29025
  • 主分类:evaluation
  • 形态:method
  • 被引:0
  • 被引来源:Semantic Scholar + OpenAlex
  • S2被引:0
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:A Multi-dimensional Evaluation-Verification Reward (EVR) decomposes evaluation into distinct visual criteria; for each criterion, an MLLM Evaluator generates multiple candidate hypotheses, and a Verifier grounds each claim in concrete visual evidence to accept or reject it, producing reliable and fine-grained reward signals.
  • OpenAlex ID:W7172326985
  • OpenAlex DOI:10.48550/arxiv.2607.29025
  • DOI:10.48550/arxiv.2607.29025
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2607.29025
  • OpenAlex更新:2026-08-26
  • 副分类:multimodal
  • 待LLM分类:否
  • 标题中文:用于一致多参考图像编辑的评估-验证奖励
  • TLDR中文:多维评估-验证奖励(EVR)将评估分解为独立视觉准则;针对每个准则,MLLM Evaluator生成多个候选假设,Verifier在具体视觉证据中grounding每个claim以接受或拒绝,产生可靠且细粒度的奖励信号。
  • 来源文件
  • /inbox/tom/_candidates/2026-08-03-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]