🎁 Recommender Systems

💬 ACL2025 · 7 paper notes

📌 Same area in other venues: 🔬 ICLR2026 (24) · 💬 ACL2026 (22) · 🧪 ICML2026 (11) · 🤖 AAAI2026 (27) · 🧠 NeurIPS2025 (24) · 🧪 ICML2025 (17)

🔥 Top topics: Recommendation ×6 · LLM ×4

Beyond Single Labels: Improving Conversational Recommendation through LLM-Powered Data Augmentation

To address the false negative problem in conversational recommender systems (where items users might like are incorrectly labeled as negative samples), an LLM-powered data augmentation framework is proposed. It generates synthetic labels through semantic retrieval and relevance scoring, and balances semantic relevance with collaborative information via a two-stage training strategy.

Laser: Bi-Tuning with Collaborative Information for Controllable LLM-Based Sequential Recommendation

This paper proposes the Laser framework, which inserts trainable virtual tokens as prefixes and suffixes to a frozen LLM (Bi-Tuning) to inject user-item collaborative information, and designs an MoE-based M-Former to capture diverse characteristics of different users, achieving parameter-efficient sequential recommendation.

CoVE: Compressed Vocabulary Expansion Makes Better LLM-based Recommender Systems

The CoVE framework is proposed to expand the LLM vocabulary by assigning a unique token ID and embedding to each item, which converts sequential recommendation into a next-token prediction task. Compared to existing methods, CoVE improves recommendation accuracy by up to 62% and achieves an approximate 100x speedup in inference, while addressing memory constraints in large-scale scenarios via hashed embedding compression.

GRAM: Generative Recommendation via Semantic-aware Multi-granular Late Fusion

This work proposes GRAM, a generative recommendation framework. By utilizing semantic-to-lexical translation, it encodes implicit hierarchical taxonomic and collaborative relationships of items into the LLM vocabulary space. Employing multi-granular late fusion, it independently encodes different-grained prompts and fuses them at the decoder side, yielding Recall@5 improvements of 11.5–16.0% and NDCG@5 improvements of 5.3–13.6% across four benchmarks.

KERL: Knowledge-Enhanced Personalized Recipe Recommendation using Large Language Models

Proposes KERL, a unified food recommendation system that combines the FoodKG knowledge graph with Multi-LoRA fine-tuning of Phi-3-mini. It accomplishes three functions: personalized recipe recommendation (F1=0.973), recipe generation, and micronutrient estimation, performance-wise significantly outperforming baseline LLMs and traditional embedding methods.

LOTUS: A Leaderboard for Detailed Image Captioning from Quality to Societal Bias and User Preferences

Proposed the LOTUS leaderboard, which uniformly evaluates the detailed image captioning capabilities of Large Vision-Language Models across three dimensions: description quality (alignment, descriptiveness, language complexity), side effects (hallucinations, toxicity), and societal bias (gender, skin tone), while supporting customized evaluation based on user preferences.

RecLM: Recommendation Instruction Tuning

This work proposes RecLM, a model-agnostic recommendation instruction tuning framework. It injects collaborative filtering signals into user/item profiles generated by an LLM via two-round conversational instruction tuning, and refines profile quality using RLHF (PPO). Serving as a plug-and-play component, it consistently improves performance for BiasMF, NCF, LightGCN, SGL, and SimGCL on MIND, Netflix, and industrial datasets, demonstrating significant efficacy particularly in cold-start scenarios.