MultAttnAttrib: Training-Free Multimodal Attribution in Long Document Question Answering

  • 类型:arxiv
  • 标识:2607.01420
  • 链接:https://arxiv.org/abs/2607.01420
  • 主分类:multimodal
  • 形态:application
  • 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:This work introduces MultAttnAttrib, a training-free attribution-generation method that leverages a model's prefill pass, selected attention heads, and calibrated thresholds to locate source evidence within a document, and consistently outperforms a variety of attribution-generation methods.
  • OpenAlex ID:W7167262568
  • OpenAlex DOI:10.48550/arxiv.2607.01420
  • DOI:10.48550/arxiv.2607.01420
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2607.01420
  • OpenAlex更新:2026-07-19
  • 副分类:engineering
  • 待LLM分类:否
  • 标题中文:MultAttnAttrib:长文档问答中的免训练多模态归因
  • TLDR中文:本工作提出 MultAttnAttrib,一种免训练的归因生成方法,利用模型的预填充过程、选定的注意力头以及校准阈值在文档中定位源证据,且在多种归因生成方法上一致地表现更优。
  • 来源文件
  • /inbox/tom/_candidates/2026-07-07-rag-retrieval-reranking-candidates.json
  • /inbox/tom/_candidates/2026-07-06-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]