2026-11-17 周二 · 个性化推荐

今日场景今天的个性化赛道围着模型该怎么把有限容量用在真正有用的信号上展开。一篇针对端侧序列推荐,指出同一套压缩比例会让简单请求浪费算力、复杂请求又不够用,于是按排序是否稳定来分配通道和历史长度。另一篇则回到表格模型里的特征交叉:普通多层感知机把组合埋在隐层里,作者把乘积写进残差块,并在科学回归基准上检查模型是否把力气花在了已知的交叉上。对做推荐的人来说,前者关乎端侧排序怎么按请求调容量,后者提醒显式交叉不仅是为了把误差压下去,也是为了让人能检查模型到底抓住了哪些组合。 — 评分 16 篇,入选 2 篇,分数 3.25 – 4.05

本期目录 · 2 篇
  1. 01RACE: Ranking-Uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation
  2. 02Learning nonlinear interactions and cross-feature couplings with product-unit residual networks
LEAD STORY · 今日必读

RACE: Ranking-Uncertainty Aware Capacity-Adaptive Customization for Efficient Sequential Recommendation

端侧做序列推荐时,模型压缩常常用同一套比例对待所有请求,兴趣简单的浪费算力,兴趣复杂的又不够用。现有剪枝或分解方法多半先给模型和输入定死一个尺寸,很难跟着这次行为序列的难度调整容量。作者用模型对当前排序结果稳不稳来估计这次有多难,再据此决定保留多少通道、截掉多长历史。以 SASRec 为骨干,在常用数据上平均来看,排序质量指标 NDCG 比未压缩的原模型提高 40.87%,计算量同时减半还多。

相关 5新颖 4实用 3证据 3时效 5
4.05
综合分 / 5
偏好调整 4.09
Tianqi Liu, Kaipu Sun, Kairui Fu, Wenkai Wang, Sihao Liu, Hui Zang, Ran Chen, Tao Wang 等 9 人 · 2026-11-16 · arXiv

今日其余入选

1 篇 · 按综合排名
02 3.25

Learning nonlinear interactions and cross-feature couplings with product-unit residual networks

表格模型要预测由多个因素交叉决定的结果时,普通多层感知机往往把交叉关系埋在隐层里,误差不一定差,但很难判断它是不是抓住了真正起作用的组合。推荐里常用的 FM 和交叉网络也是在显式做特征相乘,这篇把乘积运算写进残差块,并用复数权重同时表示幅度和相位,让优化更稳。在经典的 Friedman 函数上,均方根误差比普通多层感知机低 53.9%。交互图里有 60.7% 的强度集中在那个已知的成对关系上,普通网络只有 31.0%。预测变准,是因为模型把力气花在了真正的交叉上。这些对比做在科学回归基准上,对做点击率或转化这类也靠特征交叉的模型来说,显式乘法和残差提供了一种可检查的结构,而不是只让隐层自己去学组合。

Ziyuan Li, Uwe Jaekel, Babette Dellen
arXiv·2026-11-16·Relevance 3 · Novelty 3 · Practical 4 · Evidence 4 · Timely 5

其他赛道今日

点击切换

开源雷达

GitHub / HuggingFace 热门 · 8 个项目
aituringflow/lingerAn open-source personal desktop agent for macOS and Windows. Tell it what you need by text or voice, then follow the progress and results in conversation — themes and wallpapers, multiple agent frameworks (WorkBuddy, Codex, DeepSeek Harness), a local knowledge base, and collaboration built in.TypeScript · ★ 14 · 2026-11-12 · GitHubbertholomus/glm-5.3-exlr8-4x-sparkGLM-5.3 (full) EXLR8 on four DGX Sparks with gr8: bit275 (~2.75 bpw) manifest, concurrency dial CONC 1/2/4/8, busy guard (BUSY=1, v2), optional SSD KV store and spark-atticShell · ★ 10 · 2026-11-11 · GitHubmicrosoft/eureka-ml-insightsA framework for standardizing evaluations of large foundation models, beyond single-score reporting and rankings.Python · ★ 187 · 2024-07-18 · GitHubinternlm/WildClawBench-TrajectoriesWildClawBench Trajectories What 12 frontier models did on 60 real-world agent tasks, step by step, with the score of every run 720 graded trajectories. 12 models. 15,678 model steps. 17,206 tool calls. 90 hours of agent time. Every model on this list ran the full WildClawBench su; tags: task_categories:text-generation, language:en, language:zh, license:mit, size_categories:n<1K, format:json, modality:tabular, modality:textdataset · ★ 8 · 2026-07-30 · HuggingFacedeepseek-ai/DeepSeek-V4.1-Flash-1017pipeline: image-text-to-text; tags: transformers, safetensors, deepseek_v41, text-generation, image-text-to-text, conversational, license:mit, eval-resultsmodel · ★ 5785 · 2026-10-17 · HuggingFaceinternlm/WildClawBenchWildClawBench Hard, practical, end-to-end evaluation for AI agents — in the wild. WildClawBench is an agent benchmark that tests what actually matters: can an AI agent do real work, end-to-end, without hand-holding? We drop agents into a live OpenClaw environment — the same open-; tags: benchmark:official, task_categories:visual-question-answering, task_categories:image-text-to-text, task_categories:question-answering, language:en, language:zh, license:mit, size_categories:n<1Kdataset · ★ 73 · 2026-03-24 · HuggingFaceinternlm/Sage-S2.1-Flashtags: safetensors, glm5_next, region:usmodel · ★ 46 · 2026-11-15 · HuggingFaceinternlm/Sage-S2.1-Flash-FP8tags: safetensors, glm5_next, fp8, region:usmodel · ★ 5 · 2026-11-13 · HuggingFace