2026-10-06 周二 · Generative

独立报告页 · 在站内浏览(含往期切换)→
Generative生成式推荐 · 2 篇Personalization个性化推荐 · 2 篇AI SearchAI 搜索 · 6 篇Agentic智能体搜索 · 10 篇Infra系统与索引 · 10 篇

今日场景社区搜索 (UGC 短文本搜索): 用户在社区平台搜索帖子、问答、评论等 UGC 内容。 核心指标是搜索相关性 (nDCG) 和点击率 (CTR)。 关注:语义匹配、query 理解、短文本召回、排序模型、LLM 在搜索中的应用。 — 评分 2 篇,入选 2 篇,分数 3.65 – 4.30

本期目录 · 2 篇
  1. 01When History Misleads: Asymmetric Margin Supervision for Instruction-Guided LLM Generative Recommendation
  2. 02Reasoning with Evidence, Not Merely Rationales: Verifiable Preference Proofs for LLM-Based Recommendation
LEAD STORY · 今日必读 · Meta / Duke

When History Misleads: Asymmetric Margin Supervision for Instruction-Guided LLM Generative Recommendation

AIMS解决当前指令被历史偏好覆盖的问题:冻结模型逐条删除历史事件,筛选同时提高目标得分与边界竞争间隔的删除结果,再用完整历史学习该间隔,辅助梯度仅抑制竞争项。在6个模型骨干上,Recall@10较继续交叉熵训练提高4.0%–10.9%;Llama-8B工业数据从6.902%升至7.223%。

相关 5新颖 4实用 4证据 4时效 4
4.30
综合分 / 5
偏好调整 4.60
KEY INSIGHT · 对业务意味着什么社区UGC搜索中,用户当前查询的类别、时长等约束可能与点击历史冲突。可借鉴其离线反事实间隔监督,保留完整历史与现有解码器,避免在线逐条删历史的成本;Qilin上Llama-8B的nDCG@10从22.92%升至24.41%。公开基准采用候选池排序,不能直接当作全库召回收益。
Yin, Ming, Yang, Yuhan, Chen, Chen, Lin, Xinyu, Shi, Wentao, Yin, Fangcong, Yang, Chaofei,... · 2026-10-01 · arXiv
独立页面阅读 ↗

今日其余入选

1 篇 · 按综合排名
02 3.65

Reasoning with Evidence, Not Merely Rationales: Verifiable Preference Proofs for LLM-Based Recommendation

YonseiKyungpook National UniversitySungkyunkwan University

PROVE-Rec将历史评论压缩为带证据编号的正向与回避偏好,再仅用偏好及选中证据推荐。训练通过证据遮蔽、偏好删除与排序保持目标检验推理链。在Grocery、Tools和Yelp三个数据集上,相对最强基线各指标提高2.01%–7.45%,Yelp的nDCG@5由0.0161升至0.0173。

INSIGHT社区帖子、问答和评论可作为用户兴趣证据,但流畅解释并不保证真正参与排序。可把偏好抽取与推荐分开,用相同证据下的偏好删除检验排序影响;文中未缓存双阶段延迟约313毫秒,缓存后约63毫秒。优化器和训练规模披露不足,迁移到UGC的nDCG与CTR仍需单独验证。

Hou, Yu, Kang, Nathaniel, Wang, Pengkai, Li, Hua
arXiv·2026-10-02·Relevance 4 · Novelty 4 · Practical 3 · Evidence 3 · Timely 4

其他赛道今日

点击切换
Personalization
2
个性化推荐 · 篇入选
AI Search
6
AI 搜索 · 篇入选
Agentic
10
智能体搜索 · 篇入选
Infra
10
系统与索引 · 篇入选

开源雷达

GitHub / HuggingFace 热门 · 6 个项目
explainers-by-googlers/decisions-apiAn explainer exploring fast, type-safe, on-device ML decisions, classification, scoring, and rankingBikeshed · ★ 5 · 2026-09-30 · GitHubzai-org/Vision2Web-LeaderboardVision2Web Leaderboard Submissions This repository accepts leaderboard submissions for Vision2Web, a benchmark for visual website development agents. Submissions should contain the inference outputs (i.e., the generated code/files) for each task. Evaluation is conducted by the ma; tags: task_categories:text-generation, language:en, license:apache-2.0, region:us, codedataset · ★ 7 · 2026-03-29 · HuggingFacedeepseek-ai/DeepSeek-V4.1-Flashpipeline: image-text-to-text; tags: transformers, safetensors, deepseek_v41, text-generation, image-text-to-text, conversational, license:mit, eval-resultsmodel · ★ 4117 · 2026-09-10 · HuggingFaceQwen/Qwen-Image-2.1pipeline: text-to-image; tags: diffusers, safetensors, qwen, image-generation, image-editing, rgba, text-to-image, license:othermodel · ★ 2983 · 2026-09-14 · HuggingFaceAlibaba-DAMO-Academy/RynnWorld-Latenttags: safetensors, cosmos3_omni, custom_code, region:usmodel · ★ 3 · 2026-10-05 · HuggingFaceinternlm/AdvancedMathBenchAdvancedMathBench Paper · HF Paper · GitHub · Dataset · AutoVerifier AdvancedMathBench evaluates natural-language mathematical proof generation and verification at the undergraduate (UGD) and qualifying-exam (QE) levels. Benchmark Task Problems Proofs UGD / QE ProverBench Generat; tags: task_categories:text-generation, language:en, language:zh, size_categories:1K<n<10K, format:json, modality:text, library:datasets, library:pandasdataset · ★ 1 · 2026-09-29 · HuggingFace