Tag
RecMem is a recurrent memory compression framework that compresses long user interaction histories into fixed memory slots for efficient generative recommendation, achieving 87.9% of HSTU-1200's Recall@10 with only 16 memory slots and a 216-token decoder context.
The paper proves that group-blind recommendation policies incur linear regret when serving intersecting demographic groups, and proposes three group-aware UCB-based algorithms with sublinear regret guarantees, evaluated on Goodreads and BoardGameGeek datasets.
New Meta paper introduces PAMO, a training method for personal AI agents that use cross-app user history to improve app recommendations by learning when to defer to the platform's ranking. On real Steam users with Nintendo and Xbox history, a 4B PAMO agent hit 47.9% top-10 rate vs. 40.8% for the platform baseline.
The paper introduces LoopRec, an item-level adaptive recursion framework for sequential recommendation that reuses a weight-shared backbone block with learned gates to allocate per-item computation depth, achieving consistent gains across 3 Amazon benchmarks and 4 backbones including SASRec and NextItNet.
StoBasket 提出联合随机篮筐编辑(joint stochastic basket editing)框架,将升级销售(upsell)与交叉销售(cross-sell)统一优化以最大化贡献毛利,并在五个公开购物篮数据集上的仿真中几乎捕获全部可实现利润。研究显示联合选择在 Instacart 上比独立选择提升 10.2% 的可实现边际,且精确求解器在毫秒级内完成优化。
Walmart Global Tech 研究员探讨在电商搜索页面中使用语言模型离线评估页面级布局决策(如插入次级产品栈的位置),结果表明基于表示的方法在预测用户参与度方面持续优于直接提示判断。
Yandex Music introduces Sona, a single transformer that replaced 15+ candidate generators plus pre-ranker and ranker in production A/B tests, using History Compression to halve inference cost while delivering +4.53% Active Users and +6.30% Listening Time.
This Samsung Semiconductor paper proposes LINEUP, which decouples personalization from per-user adaptation by learning a bank of reusable low-rank factors and restricting each user to only 8 trainable scalars instead of 4.19M per-user LoRA parameters, outperforming baselines on all 12 metrics across six personalized LLM tasks.
This arXiv preprint develops a theory of context sufficiency for generative AI personalization, proposing a Context-Sufficiency Frontier that distinguishes four states (insufficiency, sufficiency, saturation, interference) and showing via a full-factorial experiment with a generative recommender at a home-furnishing retailer that relevant context improves appropriateness while irrelevant context harms it and destabilizes retrieval.
SimTrace is an open-source framework that generates faithful, fine-grained synthetic multimodal user clickstreams using a computer-use agent grounded in anonymized real trajectories and simulated web environments. It outperforms baselines on 7 of 8 fidelity metrics and improves next-action prediction by 11.0% when augmenting real data.
This paper applies Andrej Karpathy's AutoResearch paradigm to production-scale machine learning, identifying five recurring failure modes and proposing a multi-agent framework that achieves significant performance improvements over hand-tuned baselines.
The paper proposes PsyBer-Agent, a psychology-inspired LLM user simulator for recommender systems that models evolving latent states like fatigue and shifting preferences, calibrated via Unbalanced Gromov-Wasserstein optimal transport, along with the WebSim platform and PsyBer Benchmark for evaluating behavioral fidelity at scale.
This paper proposes a framework to learn cross-task relationships in multi-task models by approximating joint label distributions, improving performance in YouTube's recommendation systems through transfer learning.
This paper introduces a dual-model masking metric to benchmark ten explainable methods for temporal attribution in sequential recommendation systems, finding that gradient-based methods like GradientSHAP and Integrated Gradients yield the most faithful and robust attributions.
This paper introduces the Lightweight Ranking Heads framework to accelerate multi-task experimentation in production recommender systems by enabling dynamic task injection without retraining backbone models, reducing iteration cycles from weeks to days at YouTube scale.
AutoRecLab is an autonomous Python-based system that automates recommender systems experiments from natural-language prompts, using retrieval-augmented generation (RAG), static verification, and tree search to generate and validate executable code.
Auto-RecSys is an autonomous research system for automating long-horizon experimentation on industry-scale recommendation models, using distributed execution, centralized memory, and cognitive-procedural separation to improve efficiency and reliability.
NVIDIA reports significant speedups with PyTorch AOTI backend in HSTU inference tests, up to 2.38x in ideal scenarios, and highlights tools like recsys-examples and nv-embedding-cache for generative recommender systems at scale.
This paper explores using a model's semantic ID hierarchy for off-policy evaluation in generative recommenders, showing that coarsening to code-prefix clusters improves estimation accuracy under production logging constraints.
This paper introduces TH-GNN, a heterogeneous temporal graph neural network that fuses graph structure and textual semantics to detect shilling attacks generated by LLM agents in recommender systems, achieving superior performance over existing methods.