recommender-systems

This paper proposes residual dominance as a structural explanation for last-item reliance in causal self-attention based sequential recommenders, using prediction-time diagnostics and norm-based analysis to link this behavior to residual addition in transformer models.

This paper introduces Content Depth Score (CDS) and SCOPE-Bench to evaluate short-video recommender systems for cognitive depth, revealing that existing systems prioritize shallow content over deeper engagement.

The paper proposes an Inverse Theory of Mind (IToM) pipeline that infers user beliefs, preferences, and decision-making traits from observed interactions, using LLM-driven counterfactual reasoning to synthesize structured user personas for adaptive content recommendation across modalities, including a VisionOS spatial banking app.

This paper presents RecSys Factory, an LLM-agent platform deployed at Tencent that confines agent autonomy to decision points rather than full pipelines, balancing autonomy, determinism, and efficiency across three industrial recommender business lines.

MARCO is a Meta AI framework that decomposes clicks by intent to improve ads conversion prediction, correcting per-intent calibration bias and lifting conversions per click by +2.80% and topline metrics by +0.98% in production.

The paper argues that detecting an average effect of an acquired LLM-derived signal is not the same as learning per-instance acquisition policies, and establishes a reward-SNR floor (ρ* ≈ 2.8/√N) below which offline routing is impossible. It introduces Structured Hypothesis Embeddings (SHE) and shows across three datasets that learned per-example acquisition collapses below this floor, recommending design-time regime gates instead.

An analysis of Twitter's community structure using domain co-occurrence data and PCA/t-SNE, revealing tight right-wing and diffuse left-wing clusters, with implications for recommender systems and random walk mixing times.

This paper proposes GOAL, a constraint-aware generative framework for incentivized advertising that formulates incentive allocation as conditional sequence generation, and introduces SCPO to learn a single generative policy that generalizes across ROI constraints. Experiments show improved long-term revenue and user retention with reduced ROI violations.

This paper proposes A/B Agent, a closed-loop agent framework that organizes historical A/B testing knowledge into a hierarchical experience tree, retrieves transferable strategies via multi-path Tree-RAG, and self-evolves through online experiment feedback, achieving a 4.829% GMV improvement in a short-video e-commerce recommendation system.

RecHarness is a bandit-routed agentic harness that automates recommender model optimization by separating direction selection from hypothesis generation, achieving stable improvements and significant gains in an online A/B test.

This paper systematically investigates Semantic IDs (SIDs) in generative recommendation, finding that while SIDs preserve coarse item organization, they lose fine local structure from the encoder. The authors propose Item-Supported Decoding (ISD), a lightweight inference-time method that improves NDCG@10 by up to 31.2% without additional parameters or retraining.

This paper introduces Proximity Features, a privacy-compliant system that uses aggregated geo-IP data to personalize recommendations for cold-start users at Airbnb, achieving significant booking lifts in production experiments.

RouteRec is a framework for evaluating request-level hard selection versus item-level learned aggregation of heterogeneous recommender agents, including an LLM reranker, under cost constraints. Experiments on MovieLens-1M show that item-level aggregation significantly outperforms request-level selection.

This paper analyzes the early training collapse phenomenon in deep neural models for click-through rate prediction and proposes mitigation strategies such as sparse feature removal and value filtering, demonstrating improvements on large-scale industrial datasets.

This paper argues that agents should help users construct preferences rather than assuming well-formed ones, proposing the CoPref model and CoShop benchmark. Evaluations show even frontier models achieve only 56% accuracy due to poor preference expansion.

Proposes a multi-objective reinforcement learning framework combining semantic embeddings with Pareto-DQN to balance engagement, diversity, and fairness in recommendations, mitigating filter bubbles.

OneRank proposes a Transformer-native multi-task ranking framework that integrates feature encoding and prediction to reduce inter-task interference and improve ranking performance in recommender systems.

This paper proposes a carbon-aware re-ranking strategy for e-commerce recommendations, using a retrieval-augmented pipeline to estimate product carbon footprints and trading off predicted engagement against sustainability. Evaluated on Amazon Reviews data, substantial carbon reductions are achievable with minimal engagement loss.

τ-Rec is a verifiable benchmark for agentic recommender systems that replaces subjective LLM-as-a-judge evaluations with verifiable rewards and controlled dialogue constraints, revealing steep reliability cliffs across leading models where even the best achieves only ~57% pass@1.

This paper proposes RankElastor, a novel architecture that mitigates embedding collapse in dense scaling of recommendation models by introducing parameterized full mixing and GLU-improved P-FFNs, achieving robust scaling and improved performance on large-scale datasets.