Skip to content

๐ŸŽ Recommender Systems

๐Ÿ’ฌ EMNLP2026 ยท 4 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿง  NeurIPS2026 (24) ยท ๐Ÿ”ฌ ICLR2026 (24) ยท ๐Ÿ’ฌ ACL2026 (22) ยท ๐Ÿงช ICML2026 (11) ยท ๐Ÿค– AAAI2026 (27) ยท ๐Ÿง  NeurIPS2025 (24)

๐Ÿ”ฅ Top topics: Recommendation ร—4

Illuminating the Path: Reachability-Aware Supervision Curation for LLM-Based Recommendation

GEARec overcomes context-utilization bias and severe exploration collapse in LLM-based sequential recommendation through a reachability-aware data curation pipelineโ€”extracting dual-perspective anchors, filtering unreachable reasoning trajectories via full-pool retrieval, and aligning with dynamic fine-tuningโ€”achieving state-of-the-art accuracy across full-item candidate pools and boosting long-tail recommendation.

LLaDA-Rec: Discrete Diffusion for Parallel Semantic ID Generation in Generative Recommendation

Addressing the core weaknesses of autoregressive generative recommendationโ€”unidirectional constraints that undermine global representation and early token mistakes that cascade along the decoding sequenceโ€”LLaDA-Rec introduces discrete diffusion for parallel semantic ID generation via Multi-Head VQ-VAE tokenization, dual-level masking, and adaptive-order beam search.

Re2A: Situated Conversational Recommendation via Rubric-based Preference Reasoning and Alignment

Re2A formulates situated conversational recommendation as a structured reason-then-align process, using automated rubric-guided policy optimization to infer explicit preference states and heterogeneous preference-conditioned DPO to resolve the dual alignment dilemma between user needs and physical scene constraints.

SafeGEO: Understanding Generative Engine Optimization Risks in Recommendation Agents

SafeGEO reframes Generative Engine Optimization (GEO) from visibility hacking into a utility-alignment failure for recommendation agents, evaluating 22 attack variants across 600 cases to show that realistic seller rewrites inflate flawed product recommendations by up to 83.2 pp while pre-ranking evidence breakdown mitigates this by up to 39.2 pp.