🔍 LLM-based Recommendation

What: This topic covers the direct use of Large Language Models as recommender engines, encompassing prompt-based approaches, fine-tuning and reinforcement learning methods, and generative recommendation techniques that leverage LLM reasoning and world knowledge for personalized item suggestions.

Why: Traditional collaborative filtering and content-based methods struggle with cold-start users, lack explainability, and cannot leverage the vast world knowledge embedded in LLMs. Integrating LLMs promises more intelligent, conversational, and transparent recommendation systems.

Baseline: Conventional approaches use collaborative filtering (e.g., matrix factorization, graph neural networks like LightGCN) with learned user/item ID embeddings, relying purely on interaction patterns without semantic understanding or external knowledge.

• Bridging the semantic-collaborative gap: LLMs understand language but miss behavioral co-occurrence patterns encoded in user-item interaction graphs
• Inference efficiency: Autoregressive decoding and large model sizes create prohibitive latency for real-time recommendation serving at scale
• Hallucination and grounding: LLMs may generate non-existent items or recommendations misaligned with the actual catalog
• Fairness and bias: LLMs inherit social stereotypes from pre-training data that can amplify unfair treatment across demographic groups

What: Prompt-based recommendation leverages large language models by encoding user interaction histories and item metadata as natural language prompts to generate personalized recommendations, typically without modifying the LLM's internal parameters.

Why: LLMs possess extensive world knowledge and reasoning capabilities that, when properly prompted, can address cold-start problems, enable cross-domain transfer, and provide explainable recommendations without expensive task-specific model training.

Baseline: Traditional approaches either fine-tune task-specific models on user-item interaction matrices (collaborative filtering) or use ID-based embeddings that lack semantic understanding, requiring separate architectures for each recommendation task.

• Bridging the semantic gap between natural language and collaborative signals: LLMs understand text but lack direct access to user-item interaction patterns that drive recommendation quality.
• Context window limitations: User histories and item catalogs are too large to fit in a single prompt, requiring intelligent selection and compression strategies.
• Prompt sensitivity and position bias: Minor changes in prompt phrasing or item ordering can significantly alter recommendations, undermining reliability.
• Balancing generalization and personalization: Zero-shot LLM prompting captures world knowledge but misses user-specific behavioral nuances that fine-tuned models exploit.

What: Post-training for recommendation encompasses techniques that adapt pre-trained large language models to recommendation tasks through fine-tuning, instruction tuning, preference alignment (e.g., DPO, RLHF), and reinforcement learning, bridging the gap between general language understanding and personalized item prediction.

Why: General-purpose LLMs possess rich world knowledge and reasoning abilities but fail at recommendation tasks due to a fundamental misalignment between their pre-training objectives (next-token prediction on text) and the goals of recommendation (modeling user-item interactions and collaborative signals). Post-training methods are essential to unlock LLM potential for personalized, accurate recommendations.

Baseline: The conventional baseline is either (a) using LLMs with in-context learning (zero-shot prompting with user history), which typically underperforms even simple collaborative filtering methods like matrix factorization, or (b) standard supervised fine-tuning (SFT) with LoRA on recommendation data formatted as instructions, which improves over prompting but suffers from popularity bias, hallucination, and inability to capture collaborative signals.

• Bridging the semantic gap between natural language token spaces and collaborative filtering signal spaces, since LLMs lack inherent understanding of user-item interaction patterns.
• Avoiding hallucination where the LLM generates plausible but invalid item identifiers or logically inconsistent recommendations.
• Adapting to evolving user preferences over time without catastrophic forgetting of stable long-term interests, especially given the prohibitive cost of full LLM retraining.
• Effectively leveraging negative signals during training, since standard SFT only learns from positive examples and DPO-based methods amplify popularity bias.

What: Generative recommendation reformulates the traditional retrieve-and-rank paradigm into a generation task, where models autoregressively produce item identifiers or descriptions directly from user interaction histories using language model architectures.

Why: By casting recommendation as generation, these systems can leverage LLMs' world knowledge for cold-start items, enable cross-domain transfer without shared IDs, and unify multiple recommendation tasks (retrieval, ranking, explanation) within a single model.

Baseline: Traditional recommendation systems use discriminative models that score each candidate item independently using collaborative filtering (e.g., matrix factorization, SASRec) or content-based features, requiring separate stages for recall, pre-ranking, and ranking.

• Bridging the semantic gap between natural language tokens and item identifiers, since items lack inherent linguistic meaning and must be encoded into formats compatible with language models
• Achieving real-time inference latency for autoregressive generation, which is inherently sequential and much slower than single-pass discriminative scoring used in production systems
• Preventing hallucination of non-existent items during generation, especially when the model produces free-form text that may not map to any real item in the catalog
• Scaling generative models to industrial catalogs with millions of items while maintaining the collaborative filtering signals that ID-based models capture effectively

What: This topic covers methods that leverage Large Language Models to enhance existing recommendation systems through generated signals such as textual profiles, semantic embeddings, synthetic training data, and reasoning-based feature augmentation, without replacing the core recommendation architecture.

Why: Traditional recommender systems rely on sparse interaction signals (clicks, ratings) and opaque ID-based embeddings, struggling with cold-start users, noisy implicit feedback, and lack of explainability. LLMs offer rich world knowledge, semantic understanding, and reasoning capabilities that can address these limitations when properly integrated.

Baseline: Conventional approaches use collaborative filtering with ID-based embeddings (e.g., LightGCN, SASRec) or content-based methods with shallow text encoders (e.g., BERT), which lack access to open-world knowledge and struggle to distinguish meaningful interactions from noise.

• Bridging the modality gap between LLM-generated textual representations and collaborative filtering embeddings without dimensional collapse
• Mitigating LLM hallucinations and propensity biases that propagate through feedback loops and corrupt recommendation quality over time
• Deploying LLM-enhanced pipelines at industrial scale while meeting strict online latency requirements (typically <100ms)
• Handling noisy implicit feedback where LLMs must distinguish genuinely informative 'hard' samples from spurious 'noisy' interactions

What: This topic covers methods that transfer the semantic understanding and reasoning capabilities of large language models (LLMs) into smaller, efficient recommendation models through teacher-student frameworks, embedding alignment, and structured knowledge extraction.

Why: LLMs achieve superior recommendation quality by capturing deep semantic nuances, but their prohibitive latency (often 100x–1000x slower) and infrastructure costs make direct deployment infeasible for real-time, industrial-scale recommendation systems.

Baseline: The conventional approach either deploys a full LLM as the recommender (accurate but slow and expensive) or trains a small model independently on interaction data alone (fast but lacking semantic understanding, especially for cold-start and long-tail items).

• Representation gap: LLM semantic embeddings and collaborative filtering ID-based embeddings live in fundamentally different spaces, making naive alignment ineffective
• Distillation noise: LLM predictions are unreliable across large portions of the user-item space, so blindly imitating them can hurt student model performance
• Efficiency-quality tradeoff: Achieving near-teacher quality while maintaining sub-millisecond latency and orders-of-magnitude smaller memory footprint
• Cold-start and sparsity: Distilled knowledge must generalize to new users, new items, and sparse interaction regimes where the student has little training signal

What: This topic covers methods that use Large Language Models to generate synthetic training data, augment sparse interaction datasets, and create enriched feature representations for recommendation systems.

Why: Recommendation systems chronically suffer from data sparsity, cold-start problems, and noisy interaction logs; LLMs offer a way to generate high-quality synthetic signals offline without requiring the LLM at serving time.

Baseline: Conventional approaches rely on raw user-item interaction logs for training collaborative filtering or sequential models, using only available metadata (titles, categories) as features, which leaves cold-start items and long-tail users underserved.

• Ensuring synthetic data faithfully reflects real user preference distributions without introducing hallucinated or biased patterns
• Maintaining diversity in generated data while avoiding mode collapse toward popular items or generic interactions
• Aligning the statistical distribution of synthetic data with real-world data so downstream models generalize rather than overfit to synthetic artifacts
• Scaling LLM-based generation cost-effectively for industrial catalogs with millions of items and users

What: This topic covers methods that leverage Large Language Models (LLMs) and pre-trained foundation models to generate, enhance, or replace the embedding representations of users and items in recommender systems, moving beyond traditional ID-based or shallow feature encodings.

Why: Traditional recommender systems rely on learned ID embeddings that lack semantic understanding, generalize poorly to new or long-tail items, and are opaque to users. LLM-derived representations offer rich world knowledge and semantic reasoning that can dramatically improve cold-start performance, interpretability, and cross-domain transferability.

Baseline: The conventional approach learns a unique embedding vector for each user and item ID from interaction history (e.g., matrix factorization or deep collaborative filtering), optionally augmented with shallow content features like TF-IDF or pre-trained word2vec encodings.

• Representation misalignment: LLM semantic spaces and collaborative filtering embedding spaces encode fundamentally different types of information, making naive alignment suboptimal.
• Cold-start and long-tail: Items with few or no interactions lack sufficient signal for ID-based embeddings, yet these constitute the majority of real-world item catalogs.
• Computational cost: Running LLMs at inference time for every recommendation is prohibitively expensive, requiring efficient offline extraction or distillation strategies.
• Noise and hallucination: LLMs can generate plausible but incorrect information, and direct LLM re-ranking has been shown to underperform traditional methods due to hallucinated suggestions.

What: This topic covers methods that enhance traditional collaborative filtering (CF) by integrating signals from Large Language Models, including semantic embeddings, LLM-generated features, retrieval-augmented reasoning, and hybrid architectures that bridge the gap between behavioral interaction patterns and rich semantic understanding.

Why: Traditional CF relies solely on user-item interaction matrices, which suffer from data sparsity, cold-start problems, and inability to capture semantic nuances. LLMs offer rich world knowledge and reasoning capabilities that can complement CF's behavioral signals, but naively combining them leads to modality misalignment and computational inefficiency.

Baseline: Standard collaborative filtering methods such as Matrix Factorization (MF), BPR, LightGCN, and SASRec learn user and item representations purely from interaction history. These models excel when interaction data is dense but degrade significantly for new users/items or sparse domains.

• Modality gap: LLM semantic representations and CF collaborative embeddings exist in fundamentally different spaces, making naive alignment noisy and sub-optimal
• Scalability: LLMs are computationally expensive for real-time inference, requiring offline pre-computation or lightweight distillation strategies
• Cold-start persistence: While LLMs provide semantic priors, effectively transferring this knowledge to improve recommendations for users/items with minimal interactions remains difficult
• Signal contamination: Forcing collaborative and semantic signals into a shared space can dilute modality-specific information that is uniquely valuable for recommendation

What: This topic covers methods that ensure recommendation results are diverse, fair, and capable of surfacing surprising or long-tail items, balancing traditional accuracy-focused optimization with novelty, coverage, and multi-stakeholder objectives.

Why: Purely accuracy-optimized recommender systems tend to create filter bubbles, reinforce popularity bias, and exclude minority perspectives, ultimately degrading long-term user satisfaction and platform health. Addressing diversity and serendipity is essential for user trust, creator incentives, and societal fairness.

Baseline: Conventional recommender systems optimize a single objective (e.g., click-through rate or rating prediction accuracy) using collaborative filtering or neural ranking models, treating items as passive entities and ignoring stakeholder trade-offs, governance constraints, and fairness requirements.

• Balancing multiple competing objectives (accuracy vs. diversity vs. fairness vs. safety) without a clear Pareto-optimal solution
• Ensuring hard governance constraints (e.g., diversity quotas, safety policies) are reliably satisfied rather than approximately followed
• Detecting and mitigating systemic biases (geographic, demographic, popularity) embedded in training data and LLM world knowledge
• Avoiding filter bubbles in interactive settings where short-term feedback loops reinforce narrow user preferences over time

What: Result diversification encompasses algorithmic methods that re-rank, restructure, or augment recommendation lists to reduce redundancy, improve item coverage, and expose users to a broader range of content beyond what pure relevance optimization would surface.

Why: Relevance-only optimization creates echo chambers and filter bubbles, limiting user discovery and concentrating exposure on popular items. Diversification is essential for user satisfaction, platform fairness, and long-term engagement.

Baseline: The conventional approach ranks items solely by predicted relevance (e.g., collaborative filtering scores), producing homogeneous lists dominated by a narrow set of popular or topically similar items. Post-hoc methods like Maximal Marginal Relevance (MMR) apply simple pairwise dissimilarity penalties but use coarse category-level features.

• Balancing the accuracy-diversity trade-off: increasing diversity typically reduces short-term relevance metrics like nDCG
• Defining diversity at the right granularity—coarse category-level metrics miss fine-grained semantic redundancy while overly detailed metrics are expensive to compute
• Satisfying hard business constraints (seller coverage, fairness quotas) while jointly optimizing multiple objectives
• Bridging the exposure-consumption gap: users may ignore diverse items even when shown, requiring presentation-level interventions beyond re-ranking

What: This topic covers methods for recommending items that are both surprising and relevant to users, as well as strategies for balancing exploration of new content with exploitation of known user preferences in recommender systems.

Why: Standard recommendation algorithms create filter bubbles by reinforcing existing preferences, leading to user fatigue and reduced long-term satisfaction. Introducing serendipity and principled exploration helps users discover content they would not have found on their own but genuinely enjoy.

Baseline: Conventional approaches rely on collaborative filtering or content-based methods that optimize for predicted relevance (e.g., click-through rate), occasionally augmented with simple diversity heuristics such as random injection or popularity-based re-ranking.

• Serendipity is inherently subjective and emotional, making it extremely difficult to measure without costly user studies
• Balancing novelty with relevance is a tension: too much surprise leads to irrelevant recommendations, too little perpetuates filter bubbles
• Feedback loops in production systems systematically suppress novel content because models are trained on biased historical engagement data
• Scaling exploration strategies to billions of users while maintaining low latency and acceptable conversion rates is operationally challenging

What: Conversational recommendation enables users to discover and refine item suggestions through multi-turn natural language dialogue, combining preference elicitation, clarification, and interactive exploration.

Why: As user preferences become more nuanced and context-dependent, static recommendation lists fail to capture evolving intent; conversational interfaces allow systems to iteratively understand and satisfy complex, multi-faceted needs.

Baseline: Traditional approaches use collaborative filtering on historical interaction data or slot-filling dialogue systems that map user utterances to rigid metadata attributes, often failing with sparse data or complex natural language expressions.

• Bridging the semantic gap between free-form natural language preferences and structured item metadata
• Handling cold-start scenarios where users have little or no interaction history
• Scaling to long contexts with many candidate items without LLM degradation or token overflow
• Aligning general-purpose LLM outputs with task-specific ranking objectives like click-through rate

What: Dialogue-based conversational recommender systems (CRS) use multi-turn natural language interactions to elicit user preferences, provide personalized item recommendations, and explain reasoning through conversational exchanges.

Why: Traditional recommender systems passively infer preferences from behavioral signals, but many real-world decisions require iterative preference refinement, clarification, and explanation that only interactive dialogue can provide. CRS enables richer user engagement and handles cold-start and ambiguous preference scenarios more naturally.

Baseline: Conventional CRS approaches use separate modules for natural language understanding and recommendation, typically matching dialogue context against a fixed item catalog using learned latent vectors, then generating template-like responses with limited personalization or explanation.

• Bridging the modality gap between unstructured dialogue text and structured item catalogs or knowledge graphs, which requires aligning language understanding with domain-specific recommendation knowledge
• Balancing exploration (asking clarifying questions to narrow preferences) and exploitation (making recommendations) across multi-turn dialogues without exhausting user patience
• Evaluating CRS holistically, since standard metrics assess recommendation accuracy and dialogue quality separately and fail to capture the dynamic, interactive nature of real conversations
• Ensuring LLMs recommend only valid catalog items without hallucinating non-existent products, while maintaining fluent and contextually appropriate dialogue

What: This topic covers the use of Large Language Models (LLMs) to simulate realistic user behavior for training, evaluating, and improving recommender systems, particularly conversational and interactive ones.

Why: Real-user evaluation is expensive, risky, and slow, while static offline metrics often fail to capture the interactive, multi-turn nature of modern recommender systems. User simulation enables safe, scalable, and repeatable experimentation.

Baseline: Traditional approaches rely on static offline evaluation against ground-truth items in logged datasets, or use simple rule-based/agenda-based user models with fixed response templates and hand-crafted preference logic.

• Faithfulness gap: LLM simulators possess far more world knowledge than real users ('cognitive superman' bias), leading to unrealistically informed responses
• Data leakage: simulators may inadvertently reveal target items in conversation history, inflating evaluation metrics
• Behavioral realism: capturing nuanced human behaviors like preference drift, fatigue, spontaneous interests, and disengagement is difficult to encode
• Scalability vs. fidelity tradeoff: powerful LLMs produce realistic behavior but are too expensive for large-scale simulation, while smaller models sacrifice quality

What: This topic covers methods that enhance recommender systems by incorporating external knowledge sources—such as large language models, knowledge graphs, retrieval-augmented generation, and structured domain expertise—to improve recommendation quality beyond what collaborative filtering alone can achieve.

Why: Traditional recommender systems rely heavily on user-item interaction data, which suffers from cold-start problems, data sparsity, and inability to reason about item semantics or cross-domain preferences. External knowledge sources bridge these gaps by providing world knowledge, domain expertise, and semantic understanding.

Baseline: The conventional approach uses collaborative filtering (e.g., matrix factorization, LightGCN) or content-based methods with BERT-style encoders that learn from interaction histories alone, without leveraging external knowledge or reasoning capabilities.

• Integrating heterogeneous knowledge sources (unstructured text, knowledge graphs, LLM reasoning) with structured interaction data without introducing noise or hallucinations
• Maintaining computational efficiency when augmenting recommendations with expensive LLM inference or large-scale knowledge retrieval
• Preserving user privacy while leveraging external knowledge in federated or distributed settings
• Ensuring factual consistency and explainability when LLMs generate recommendation explanations or cross-domain inferences

What: Knowledge Graph-enhanced Recommendation leverages structured graphs of entities and relations (e.g., movies, actors, genres connected by typed edges) to enrich user and item representations, enable multi-hop reasoning paths, and improve both accuracy and explainability of recommender systems.

Why: Traditional collaborative filtering suffers from data sparsity and cold-start problems, and provides opaque predictions. Knowledge graphs supply rich side information and relational structure that help systems reason about why a user might like an item, enabling both better predictions and human-understandable explanations.

Baseline: The conventional approach uses collaborative filtering (matrix factorization or basic GNNs on user-item bipartite graphs) that relies solely on interaction signals. These baselines lack external knowledge about item attributes and relationships, struggle with new users/items, and cannot explain their recommendations.

• KG noise and incompleteness: Knowledge graphs contain irrelevant triples and missing links that can mislead recommendation models if not properly filtered
• Scalability of reasoning: Multi-hop path reasoning over large KGs is computationally expensive, with action spaces growing exponentially at each hop
• Alignment between KG structure and user preferences: Not all KG relations are equally relevant for recommendation; learning which paths matter for which users remains difficult
• Generating faithful explanations: Many models produce attention-based or generated explanations that do not reflect the actual reasoning paths in the KG, undermining user trust

What: Retrieval-augmented recommendation combines retrieval-augmented generation (RAG) techniques with recommendation systems, fetching relevant external context—such as knowledge graph entries, historical trajectories, or domain-specific documents—to ground LLM-based recommendations in factual, up-to-date information.

Why: Standard LLM-based recommenders often hallucinate or produce generic suggestions because they lack access to domain-specific knowledge, real-time data, and structured user–item relationships. RAG bridges this gap by dynamically retrieving relevant context before generating recommendations.

Baseline: Conventional approaches rely on collaborative filtering (using user–item interaction matrices) or direct LLM prompting without external retrieval. These baselines struggle with cold-start users, domain-specific reasoning, and spatial or temporal relevance.

• Cold-start problem: new users or items lack sufficient interaction history for collaborative filtering, and naive LLM prompts produce generic results
• Domain knowledge integration: specialized domains (medical, legal, geospatial) require structured knowledge that general-purpose LLMs do not possess
• Retrieval quality: fetching irrelevant or noisy context can mislead the LLM, degrading recommendation quality rather than improving it
• Scalability of retrieval: hierarchical or graph-based retrieval adds computational overhead that must be balanced against recommendation latency requirements

What: This topic covers recommendation systems research that spans diverse themes—including LLM-based ranking, bias and fairness, explainability, clinical medication recommendation, scalable architectures, cold-start solutions, evaluation methodology, and domain-specific applications—that do not fit neatly into the main taxonomy categories.

Why: These cross-cutting research directions address foundational challenges (scalability, trust, safety, evaluation) that underpin every recommendation system, and their advances often transfer across the core taxonomy categories.

Baseline: Conventional approaches rely on collaborative filtering with ID-based embeddings, handcrafted feature-crossing modules, and historical train-test splits for evaluation—methods that struggle with cold-start users, lack transparency, and exhibit systematic biases.

• LLMs introduce new bias vectors (brand, product, demographic stereotyping) that traditional debiasing methods cannot mitigate
• Industrial ranking models must scale to billions of parameters under strict latency and throughput constraints on modern GPU hardware
• Generating faithful, sentiment-aligned explanations that are robust to noisy user interaction histories
• Evaluating recommendation quality reliably when offline metrics based on historical splits suffer from exposure bias and label sparsity

What: This topic covers methods for recommending items to new users or new items with limited interaction history, including cross-domain transfer, semantic augmentation, and zero-shot generalization approaches.

Why: Cold-start and data sparsity are fundamental bottlenecks in production recommender systems—new users and items arrive continuously, yet most models require dense interaction histories to generate quality predictions, directly impacting revenue and user retention.

Baseline: Conventional collaborative filtering (e.g., matrix factorization, LightGCN) relies on learned ID-based embeddings from historical interactions, which produce near-random recommendations for entities with few or no interactions.

• New users and items lack the interaction histories that collaborative filtering requires, creating a chicken-and-egg problem where the system cannot learn preferences without data
• Content-based fallbacks (using item metadata) suffer from a semantic gap—textual descriptions do not directly translate into behavioral compatibility signals
• Cross-domain knowledge transfer is hindered by disjoint ID spaces and distribution shifts between source and target domains
• Deploying LLMs at inference time for cold-start is computationally prohibitive at industrial scale (billions of items), requiring efficient distillation or offline augmentation strategies

What: This topic covers methods for making recommendation decisions transparent and interpretable, including generating natural language explanations, reasoning chains, and human-readable user profiles that reveal why specific items are recommended.

Why: Users increasingly demand transparency from recommendation systems to build trust, enable informed decision-making, and allow them to correct or steer their preferences. Regulatory pressures and platform accountability further drive the need for explainable recommendations.

Baseline: Traditional recommender systems rely on opaque embedding-based collaborative filtering (e.g., matrix factorization, graph neural networks) that produce accurate rankings but offer no human-understandable rationale for their decisions.

• Bridging the modality gap between latent collaborative filtering embeddings and natural language that LLMs can reason over
• Generating explanations that are both faithful to the model's actual decision process and factually consistent with user preferences
• Maintaining recommendation accuracy while adding interpretability, as jointly optimizing ranking and explanation often creates conflicting objectives
• Deploying reasoning-enhanced recommenders at scale, since LLM inference is too slow and expensive for real-time industrial systems

What: Multimodal recommendation encompasses approaches that leverage multiple data modalities—text, images, video, audio, and structured knowledge—alongside collaborative filtering signals to build richer representations of users and items for more accurate and explainable recommendations.

Why: Users interact with items through diverse signals (visual appeal, textual descriptions, audio qualities), yet traditional recommenders rely on sparse interaction IDs alone. Integrating multiple modalities enables better cold-start handling, richer preference modeling, and more human-interpretable explanations.

Baseline: The conventional approach uses collaborative filtering with item ID embeddings learned from user-item interaction matrices, optionally augmented by shallow feature extraction from pre-trained encoders (e.g., ResNet for images, BERT for text) that are treated as static, frozen side information.

• Semantic gap between collaborative filtering signals (behavioral patterns) and rich semantic representations from language/vision models, requiring non-trivial alignment strategies
• Modality fusion conflicts: jointly training on heterogeneous modalities (text, image, graph) causes gradient interference, where one modality's updates degrade another's learned representations
• Scalability of multimodal processing: encoding images, text, and video for millions of items in real-time is computationally prohibitive, especially when using large foundation models
• Cold-start and long-tail items lack sufficient interaction history, forcing systems to rely entirely on content features whose representation quality varies across modalities

What: Agentic recommender systems leverage LLM-based agents that autonomously plan, reason, use tools, and collaborate to deliver personalized recommendations, moving beyond traditional passive retrieval-and-rank pipelines.

Why: Traditional recommender systems treat users as passive recipients of ranked lists, leaving the cognitive burden of exploration, comparison, and synthesis entirely on the user. Agentic approaches enable proactive, context-aware, and multi-stakeholder recommendation through autonomous reasoning and tool use.

Baseline: Conventional recommenders rely on collaborative filtering (matrix factorization, sequential models like SASRec) or single-pass LLM prompting that generates a static ranked list without iterative reasoning, tool use, or multi-agent coordination.

• Bridging the gap between LLM semantic reasoning and behavioral collaborative filtering signals hidden in user-item interaction graphs
• Coordinating multiple agents with competing objectives (user relevance vs. fairness vs. diversity) while maintaining hard constraint satisfaction
• Reducing inference latency of multi-turn agent reasoning to meet real-time production requirements
• Preventing hallucinations and ensuring auditability when LLM agents autonomously generate recommendations

What: This topic covers research on making recommendation systems computationally efficient and scalable, including inference acceleration, model compression, scalable architectures, efficient training paradigms, on-device deployment, and privacy-preserving federated approaches.

Why: As recommendation systems increasingly adopt Large Language Models (LLMs) for superior semantic understanding, the gap between model quality and deployment constraints—latency budgets, throughput requirements, memory limits, and cost—has become the central bottleneck preventing industrial adoption.

Baseline: Conventional systems use massive embedding tables with shallow feature-crossing networks (e.g., DLRM, DCNv2) that plateau in capacity, while naive LLM-based recommenders process full-text prompts autoregressively, incurring prohibitive latency and cost for real-time serving.

• Autoregressive decoding in generative recommenders requires multiple sequential forward passes per recommendation, creating latency that violates real-time serving constraints (<100ms)
• LLM parameter counts (7B–100B+) far exceed the memory and compute budgets of production serving infrastructure and edge devices
• Traditional sparse scaling (larger embedding tables) exhibits diminishing returns, while dense scaling suffers from low GPU utilization on hardware designed for dense compute
• Transferring rich LLM knowledge to lightweight models without losing semantic understanding or collaborative filtering signals remains difficult due to capacity gaps and divergent representation spaces

What: This topic covers privacy-preserving recommendation approaches (federated learning, local processing, machine unlearning), security threats against recommender systems (adversarial attacks, membership inference, data poisoning), and fairness evaluation of LLM-based recommenders.

Why: As LLM-based recommender systems process increasingly sensitive user data—purchase histories, location check-ins, health records—the risks of data leakage, adversarial manipulation, and unfair treatment demand dedicated research into both attack surfaces and defense mechanisms.

Baseline: Traditional recommender systems centralize user interaction data for training, rely on ID-based collaborative filtering without explicit privacy guarantees, and assume benign inputs without adversarial robustness mechanisms.

• Balancing recommendation quality with privacy protection: privacy mechanisms (noise injection, data partitioning) often degrade the collaborative signals that make recommendations effective
• Defending against semantically sophisticated attacks: LLM-powered attackers can craft human-readable adversarial text that evades traditional detection heuristics
• Efficiently removing user data from large models: full retraining for each deletion request is computationally prohibitive, but approximate methods risk incomplete erasure
• Handling data heterogeneity in federated settings: users have vastly different interaction volumes and patterns, making uniform federated aggregation suboptimal

What: This topic covers the detection, measurement, and mitigation of biases and unfairness in recommendation systems that incorporate Large Language Models, spanning popularity bias, demographic stereotyping, exposure inequality, and feedback-loop amplification.

Why: As LLM-based recommenders are deployed at scale for high-stakes decisions (education, finance, hiring), inherited biases from pre-training data and fine-tuning procedures can systematically disadvantage specific user groups, erode trust, and concentrate exposure among already-popular items.

Baseline: Traditional fairness evaluation compares recommendation lists with and without sensitive attributes, treating any difference as bias. Conventional debiasing relies on re-weighting or re-ranking with known sensitive labels, and assumes collaborative-filtering models with separate user/item ID embeddings.

• Distinguishing valid personalization from harmful stereotyping when LLMs act on implicit identity cues in natural language
• Addressing multiple, interacting sources of bias (pre-training priors, fine-tuning amplification, decoding artifacts, and feedback loops) within a single system
• Ensuring fairness without access to sensitive user attributes, which are often unavailable due to privacy regulations
• Preventing feedback loops where biased recommendations generate biased interaction data that further entrenches the original bias over successive retraining cycles

What: This topic covers papers that conduct systematic experiments to evaluate the performance, fairness, robustness, and limitations of LLM-based and generative recommender systems, revealing gaps between current capabilities and real-world requirements.

Why: As LLMs rapidly enter recommendation pipelines, rigorous analysis is essential to distinguish genuine advances from artifacts of unfair comparisons, benchmark leakage, or inherited biases, and to identify the most promising directions for future research.

Baseline: Conventional evaluation relies on offline accuracy metrics (Hit Rate, NDCG) computed from historical interaction splits, using models trained with pointwise or pairwise loss functions on fixed item catalogs.

• Standard offline metrics suffer from exposure bias, popularity bias, and missing-not-at-random patterns, often failing to correlate with online A/B test results or real user satisfaction
• LLMs may memorize benchmark datasets during pre-training, inflating reported performance and undermining generalizability of research findings
• Fairness evaluation is complicated by the tension between valid personalization and harmful bias—differences in recommendations across demographic groups may reflect legitimate preferences rather than discrimination
• Evaluating generative outputs (explanations, dialogues, narratives) requires subjective quality judgments that traditional reference-based metrics like BLEU cannot capture