---
title: "DREAM：让大模型自己'猜答案'训练稠密检索器，无需标注正负样本"
source: wechat
source_url: https://mp.weixin.qq.com/s/OaChk2wvXSY2s2QrkgOObw
author: ChallengeHub
feed_name: ChallengeHub
review_value: 8
review_confidence: 8
review_recommendation: worth-reading
review_stars: 4
date: 2026-06-29
created: 2026-06-29
updated: 2026-06-29
tags: [dense-retrieval, rag, information-retrieval, llm, attention-head, training, contrastive-free, ntp-loss]
type: article
provenance_state: extracted
sha256: b37c3bac7dabaac60979f035a0d28fbb0df3c007aa561cab7a624012c73c0fc1