iFAN: Inference-Aware Learning for Plain Mask Transformers

  • 类型：arxiv
  • 标识：2608.03216
  • 链接：https://arxiv.org/abs/2608.03216
  • 主分类：llm-infra
  • 形态：method
  • 被引：0
  • 被引来源：Semantic Scholar
  • S2被引：0
  • 影响力被引：0
  • TLDR：iFAN introduces Adjusted Probability-Mask Ranking (APMR), which aligns query competition with predicted mask quality and suppresses high-confidence but inaccurate competitors, and employs Cross-Layer Self-Distillation (CLSD) to transfer stronger intermediate predictions to the final layer.
  • 副分类：engineering
  • 待LLM分类：否
  • 标题中文：iFAN：面向普通 Mask Transformer 的推理感知学习
  • TLDR中文：iFAN 提出 Adjusted Probability-Mask Ranking (APMR)，将查询竞争与预测的掩码质量对齐，抑制高置信度但不准确的竞争者；同时采用 Cross-Layer Self-Distillation (CLSD) 将更强的中间预测传递至最终层。
  • 来源文件：
  • /inbox/tom/_candidates/2026-08-12-agent-rag-longcontext-candidates.json
  • [S2 enrich]