LLM-Based Recommender Systems

• LRSs are recommender systems that employ large language models to represent, reason about, and interact with users and items using natural language.
• They integrate embedding-based scoring, autoregressive generation, and hybrid retrieval methods to overcome limitations of classical ID-centric approaches.
• Recent advances focus on aligning textual semantics with collaborative filtering, enhancing efficiency, fairness, and robustness across diverse application domains.

LLM-based Recommender Systems (LRSs) are recommender systems in which a LLM plays a central role in representing, reasoning about, and interacting with users and items in natural language. In this paradigm, natural language becomes a unifying substrate for item descriptions, user profiles, interaction histories, queries, and dialogue turns; the LLM may be used for embedding-based scoring, autoregressive generation of recommendations, ratings, and explanations, or as a controller inside multi-module retrieval and recommendation pipelines (Korikov et al., 2024). Contemporary LRS research spans direct generative recommenders, hybrid LLM–retriever systems, conversational recommenders, industrial universal retrievers, and post-hoc alignment methods for fairness, stability, and long-tail robustness.

1. Conceptual foundations and representational shift

A standard formalization of LRSs is sequential recommendation. Let users be $\mathcal{U} = \{u_1,\dots,u_M\}$, items be $\mathcal{I} = \{i_1,\dots,i_N\}$, and the interaction sequence of user $u$ be $\mathcal{S}^u_{1:n} = [i_1^u,\dots,i_n^u]$. The basic objective is next-item prediction,

$$\hat{i}_{n+1}^u = \arg\max_{i \in \mathcal{I}} P(i \mid \mathcal{S}^u_{1:n}; \theta),$$

but in LRSs the conditioning context is frequently textualized into prompts, summaries, or structured natural-language profiles rather than remaining purely ID-based (Li et al., 24 Jan 2026).

This marks a shift from classical recommender systems centered on standardized non-verbal feedback such as clicks, views, purchases, ratings, or ID sequences toward language-driven recommendation. Items may be represented by titles, descriptions, metadata verbalized into text, reviews, or full content; interactions may be verbalized from logs or expressed directly as queries and dialogue utterances; users may be represented by editable natural-language profiles summarizing long-term preferences, constraints, and temporary context (Korikov et al., 2024). This shift is motivated by the ability of LLMs to process nuanced, soft attributes and open-ended preference statements that are hard to encode in sparse ID spaces.

A central technical issue is the mismatch between linguistic semantics and collaborative semantics. Pre-trained LLMs encode textual similarity and world knowledge, whereas recommenders must model user–item interaction structure, substitutes, complements, sequential transitions, and exposure effects. Multiple papers frame this as a semantic gap: pure text-centric approaches underutilize collaborative structure, while pure ID-centric approaches underutilize rich content and natural-language interaction (Hong et al., 20 Feb 2025, Zhu et al., 2023). This suggests that LRS design is fundamentally about building interfaces between language space and recommendation space rather than simply attaching prompts to a generic LLM.

Text richness also creates a practical bottleneck. In text-rich sequential recommendation, long item descriptions multiplied by long user histories produce prompts that can exceed context limits and incur substantial quadratic attention cost. LLM-TRSR addresses this by segmenting user histories into blocks, summarizing those blocks with a frozen LLM, and then combining a long-term preference summary with a short window of recent interactions in the downstream recommendation prompt (Zheng et al., 2024).

2. Core modeling paradigms

One major family of LRSs uses encoder-style representations. In dense retrieval, user preference text or query text $q$ and item text $T(i)$ are encoded into embeddings $\mathbf{e}_q = f_{\text{enc}}(q)$ and $\mathbf{e}_i = f_{\text{enc}}(T(i))$, and ranking is performed by similarity, often $s(q,i) = \mathbf{e}_q^\top \mathbf{e}_i$ or cosine similarity. Cross-encoders instead jointly encode query and item text with cross-attention and predict a relevance score, typically with higher accuracy but much higher inference cost (Korikov et al., 2024).

A second family treats recommendation as autoregressive generation. BIGRec, as instantiated in fairness work, converts $\mathcal{I} = \{i_1,\dots,i_N\}$0 into a natural-language prompt $\mathcal{I} = \{i_1,\dots,i_N\}$1, fine-tunes an LLM on the resulting training set, and at inference generates a next-item token that is then grounded back into the catalog by ranking item embeddings by distance to the generated token embedding. This turns the LLM into a generative user model while preserving compatibility with a discrete candidate set (Li et al., 24 Jan 2026). Prompt-only zero-shot and few-shot variants also exist, but evidence summarized in the literature indicates that zero-shot and few-shot LLMs are usually weaker than supervised collaborative models when sufficient recommendation data are available, although they can be competitive in near cold-start regimes (Korikov et al., 2024).

A third family explicitly fuses the ID paradigm with the LLM paradigm. CLLM4Rec extends the vocabulary with indivisible user and item ID tokens, learns separate collaborative and content embeddings for those tokens, and uses a soft+hard prompting strategy in which prompts contain heterogeneous soft and hard tokens while the main text remains homogeneous—item IDs for collaborative language modeling or vocabulary tokens for content modeling. Recommendation is then performed with a multinomial item prediction head tied to item embeddings, avoiding token-by-token hallucinated item generation (Zhu et al., 2023). This suggests that explicit item-token design is not a superficial engineering detail but a structural choice about how collaborative semantics enter the LLM.

A fourth family redesigns item indexing itself. EAGER-LLM introduces dual-source knowledge-rich item indices that separately discretize semantic and behavioral embeddings through hierarchical K-Means, so that each item is represented by structured multi-token codes. It further adds multiscale alignment reconstruction tasks and an annealing adapter tuning stage to improve how the LLM interprets compressed behavioral and semantic codes (Hong et al., 20 Feb 2025). This line of work treats item-index geometry, auxiliary alignment losses, and lightweight adaptation as first-class components of LRS architecture.

3. Hybrid, tool-using, and industrial architectures

A large portion of effective LRSs are hybrid rather than purely generative. Retrieval-augmented recommendation follows a two-stage pattern: a retriever or traditional recommender produces candidates, and an LLM reranks, explains, or converses over that candidate set. This architecture reduces hallucination, improves controllability, and allows knowledge updates without retraining the base LLM (Korikov et al., 2024).

DeepRec pushes this hybrid pattern into an agentic regime. It couples an LLM with a preference-aware traditional recommendation model (TRM) and lets the LLM iteratively generate preference descriptions, call the TRM, inspect retrieved items, and refine its next query over multiple turns. The interaction is optimized by reinforcement learning with process-level and outcome-level rewards, and the final recommendation is obtained after multi-turn exploration of the item space rather than a single retrieval–rerank pass (Zheng et al., 22 May 2025). This suggests that one role of an LLM in recommender systems is not only ranking or explanation but also adaptive control of retrieval trajectories.

Other hybrids place the LLM on the feature side rather than the control side. A GAT-based framework for end-to-end personalization uses LLMs to generate structured user and item profiles from titles, genres, and overviews, embeds those profiles with a SentenceTransformer, and uses the resulting vectors as initial node features in a bipartite user–item graph processed by a graph attention network. The same LLM can then rerank the top-20 graph candidates and generate explanations (Ebrat et al., 2 Aug 2025). This reflects a broader design pattern in which LLMs provide semantically enriched features while a graph or sequential backbone performs the core collaborative ranking.

Industrial work extends this pattern to extreme-scale retrieval. The Universal Retrieval Model (URM) treats an LLM as a universal retriever: prompt templates define different recommendation or search objectives, multiple special [UM] tokens produce multiple user embedding heads in one forward pass, and recommendation becomes nearest-neighbor search over precomputed multimodal item embeddings. The item side fuses ID, text, and image features, while training jointly optimizes language modeling and retrieval losses. URM is reported to retrieve from tens of millions of candidates with latency within tens of milliseconds and to support zero-shot prompt control over scenario, long-tail, serendipity, and search behavior (Jiang et al., 5 Feb 2025).

Hybridization can also be segment-aware. CoReLLa jointly trains a conventional recommender model and an LLM, aligns their hidden representations with an explicit alignment loss, and routes only low-confidence samples from the efficient CRM to the slower LLM. Empirically, the LLM excels on data segments where the CRM has low confidence or low precision, while the CRM remains stronger on simpler and moderate cases (Xi et al., 2024). This directly challenges the misconception that effective LRSs necessarily replace classical recommenders end-to-end.

4. Training, alignment, and efficiency

The training spectrum of LRSs spans zero-shot prompting, few-shot prompting, full fine-tuning, prompt tuning, adapters, and LoRA-based parameter-efficient fine-tuning. The literature repeatedly emphasizes that recommendation-specific supervision matters: fine-tuning substantially improves ranking quality over prompt-only use of an off-the-shelf model, while LoRA and related PEFT methods make this feasible for multi-billion-parameter backbones (Korikov et al., 2024, Chen, 2023, Zhu et al., 2023).

Sample efficiency is one of the prominent claims about LLM-enhanced recommendation. Laser studies click-through-rate prediction and evaluates two modes: Laser$\mathcal{I} = \{i_1,\dots,i_N\}$2, where the LLM directly predicts “Yes” or “No” from a textualized user–item context, and Laser$\mathcal{I} = \{i_1,\dots,i_N\}$3, where the LLM generates user and item knowledge vectors that are adapted via Mixture-of-Experts modules and injected into a conventional CTR model. On BookCrossing and MovieLens-1M, Laser$\mathcal{I} = \{i_1,\dots,i_N\}$4 trained on 10% of training data surpasses the best full-data CRM in AUC, while Laser$\mathcal{I} = \{i_1,\dots,i_N\}$5 trained on 50% of the data matches or slightly surpasses strong CRMs with near-identical online latency to the base CRM (Lin et al., 2024). A plausible implication is that semantic priors from LLMs can partially substitute for dense interaction supervision, particularly in sparse settings.

Ranking alignment is another recurring issue. ALRO argues that standard next-token cross-entropy is misaligned with listwise ranking objectives and proposes a soft lambda loss that constructs differentiable item scores from generative probabilities and plugs them into a LambdaLoss-style objective. It also adds permutation-sensitive learning to reduce position bias without extra inference cost. This combination is reported to improve listwise LLM ranking and to match or exceed bootstrapping-style debiasing at a fraction of the latency (Chao et al., 2024).

Efficiency constraints remain decisive in practice. Direct autoregressive generation of item identifiers or long ranked lists is slow, and many approaches therefore move recommendation into embedding space, candidate reranking, or selective routing. LLM-TRSR reduces prompt length through summarization (Zheng et al., 2024), CLLM4Rec avoids hallucinated candidate generation by using a multinomial item prediction head (Zhu et al., 2023), and URM avoids token-by-token generation entirely by using sequence-in–set-out retrieval (Jiang et al., 5 Feb 2025). Across these systems, the general pattern is to preserve the LLM where its semantic reasoning matters and to offload large-scale retrieval or exhaustive ranking to more structured mechanisms.

5. Fairness, stability, and long-tail pathologies

LRSs inherit and sometimes amplify multiple pathologies from both their pre-training corpora and recommendation fine-tuning data. On the item side, popularity bias and broader provider-side inequities appear when historically frequent items receive disproportionate exposure. On the user side, demographic or attribute bias can manifest as unequal hit rates or recommendation probabilities across groups (Li et al., 24 Jan 2026).

FUDLR addresses fairness without costly retraining by reformulating debiasing as machine unlearning. It first learns a bias-agnostic mask over candidate training samples using a fairness term, an accuracy-preservation term, and a sparsity term; it then applies influence-based unlearning to approximate the parameter update that would result from removing the selected bias-inducing samples. The framework supports popularity bias, demographic parity-style attribute bias, and weighted combinations thereof, and is reported to improve fairness while preserving recommendation accuracy with roughly 96% runtime reduction relative to retraining-based debiasing on the reported setup (Li et al., 24 Jan 2026).

Long-tail effects in LRSs have also been analyzed as a combination of pre-training priors and recommendation-data skew. EISAM distinguishes prior long-tail, inherited from the base LLM, from data long-tail, induced by skewed recommendation datasets. Its analysis shows that both contribute to performance disparity between head and tail items, but data long-tail dominates the overall tail behavior. EISAM then introduces item-wise sharpness-aware minimization, weighting item-specific sharpness penalties by item frequency so that flatter minima are explicitly encouraged for underrepresented items. The method is presented as the first systematic solution to the long-tail problem in LRSs and is reported to improve tail-item performance while preserving overall quality (Zhang et al., 13 Mar 2026).

A distinct robustness issue is positional instability. STELLA shows that zero-shot LLM recommenders are highly sensitive to the position of candidates in the prompt, with accuracy varying substantially across permutations of the same candidate set. It learns a transition matrix that captures how the LLM maps true positions to predicted top-ranked positions during a probing stage, then uses Bayesian updating and an entropy indicator to calibrate outputs at recommendation time (Ma et al., 2023). This makes explicit that instability in LRSs can arise from prompt-order effects rather than preference modeling alone.

Recent fairness work further argues that item-side unfairness is not generated solely during SFT. UFO decomposes unfairness into pre-training and SFT components and formulates mitigation as a self-play fine-tuning game between a judger, which identifies unfairness from both sources, and a corrector, which updates the LRS to address identified unfairness while preserving recommendation performance (Zhang et al., 23 Nov 2025). This directly contests the view that reweighting or reranking at the end of the pipeline is sufficient to address root causes.

6. Evaluation, applications, and research agenda

LRS evaluation inherits standard recommender metrics but adds generative, conversational, and fairness dimensions. Top-$\mathcal{I} = \{i_1,\dots,i_N\}$6 ranking remains central: HR@K, Recall@K, NDCG@K, and MRR are standard in sequential recommendation; CTR-style tasks use AUC and LogLoss; explanation-oriented work uses BLEU, ROUGE, and diversity metrics such as USR and FCR; conversational systems additionally require dialogue quality, satisfaction, or human evaluation (Korikov et al., 2024, Lin et al., 2024, Chen, 2023). Because many LRSs produce text as well as rankings, offline ranking metrics alone do not fully characterize system quality.

Application domains are broad. The literature explicitly covers movies and books, news recommendation, e-commerce, product search, restaurant conversational recommendation, social-media-style textual interaction, and educational technologies (Korikov et al., 2024, Yu et al., 4 Jan 2025). In these settings, LRSs are used not only to rank items but to interpret unstructured reviews, generate explanations, update preferences from dialogue, and support cold-start recommendation from sparse interaction histories plus rich content.

Several misconceptions recur. One is that LLMs simply replace recommender systems. The evidence instead supports multiple hybrid regimes: LLMs as feature generators, rerankers, controllers, prompt-conditioned retrievers, and specialized experts for hard cases (Korikov et al., 2024, Xi et al., 2024, Jiang et al., 5 Feb 2025). Another is that natural-language interaction alone solves classical recommender issues. The literature documents persistent hallucination, limited control, latency constraints, privacy concerns, prompt sensitivity, position bias, unfairness, and long-tail failure modes (Korikov et al., 2024, Ma et al., 2023, Li et al., 24 Jan 2026, Zhang et al., 13 Mar 2026).

The current research agenda therefore centers on structured integration rather than uncritical substitution. Prominent directions include retrieval-augmented and tool-using architectures, editable natural-language user profiles, multimodal item indexing, efficient PEFT and adapter strategies, fairness-aware and long-tail-aware optimization, calibrated listwise ranking, and selective deployment patterns in which LLMs are invoked only where their semantic reasoning yields a clear advantage (Hong et al., 20 Feb 2025, Zheng et al., 22 May 2025, Chao et al., 2024). A plausible implication is that the mature form of LRSs will be neither purely ID-centric nor purely language-centric, but a class of systems in which LLMs, retrieval modules, graph or sequential recommenders, and post-hoc calibration components are jointly engineered around the structure of recommendation data and objectives.