GenRec: Generative Recommendation and Related Systems

Updated 20 August 2026

GenRec is an ambiguous research term most commonly describing generative recommender systems that predict items or semantic identifiers from user histories instead of independently scoring every catalog item.
Generative recommendation methods include LLaMA-based text generation, BART masked-item prediction, semantic-ID retrieval, memory compression, and catalog-aware LLM ranking for scalable personalization.
Key challenges include validating generated items, preventing hallucinations and popularity bias, handling exposure bias, aligning rewards with business goals, and comparing systems across inconsistent evaluation settings.

GenRec is an ambiguous name used for multiple research systems and frameworks spanning generative recommendation, algorithmic recourse, video understanding, novel-view synthesis, genetic pedigree reconstruction, synthetic preference-data generation, and LLM-backed ranking. In recommendation research, “GenRec” most commonly denotes a generative recommender that predicts items or item identifiers directly from user histories rather than assigning scores independently to every candidate; however, the name is also used by distinct papers with substantially different architectures and objectives. The principal recommendation systems include LLaMA-based textual generation (Ji et al., 2023), prompt-free masked sequential generation with BART (Cao et al., 2024), large-scale industrial generative retrieval (Zou et al., 16 Apr 2026), and Netflix’s LLM-backed catalog ranker (Li et al., 10 Aug 2026).

1. Terminological scope and research lineage

The term GenRec has no single canonical referent. Several papers use it independently:

Research area	System or framework	Central task
Recommendation	GenRec, Rec2PM, GenPlugin, S-GRec, AsymRec, HD-Rec	Generative or semantic-ID recommendation
Algorithmic recourse	GenRe	Generate plausible, valid, low-cost recourse
Video	GenRec	Joint video generation and recognition
Novel-view synthesis	GenRec	Separate reconstruction from generation
Genetics	Rec-Gen	Reconstruct stochastic pedigrees
Preference simulation	GenRec	Generate synthetic recommendation interactions
Ranking	Netflix GenRec	LLM-backed catalog ranking

The earliest supplied use associated with recommendation is “GenRec: LLM for Generative Recommendation,” which fine-tunes LLaMA on textual user–item interactions and directly generates a target item rather than computing a ranking score for every candidate (Ji et al., 2023). A later system formulates sequential recommendation as masked item generation using a BART encoder–decoder and explicit user- and item-ID embeddings (Cao et al., 2024). Industrial work subsequently extends the paradigm through semantic identifiers, reinforcement learning, page-level supervision, retrieval augmentation, memory compression, debiasing, and domain-adaptive quantization (Yang et al., 4 Jul 2025, Chen et al., 12 Feb 2026, Zou et al., 16 Apr 2026, He et al., 7 Aug 2026).

The recommendation meaning of GenRec should therefore be disambiguated by architecture and publication year. It may refer to open-ended text generation, semantic-ID generation, catalog-aware ranking, or a broader family of generative-retrieval systems.

2. Generative recommendation as sequence prediction

Generative recommendation reframes next-item prediction as conditional sequence generation. Given a user $u$ and chronologically ordered history $t^u_{1:n}$ , the system predicts a subsequent item using a conditional model such as

$\hat{t}_{n+1}^u=\arg\max_t P(t\mid u,t^u_{1:n}).$

This differs from conventional discriminative recommendation, in which a model computes a relevance score for each candidate item and ranks the resulting scores. A generative recommender instead produces an item, item identifier, or semantic-ID sequence autoregressively.

The first supplied GenRec recommendation system uses item names as textual representations. A prompt contains an instruction and a user’s previous interactions, and the fine-tuned LLaMA backbone generates the target item name (Ji et al., 2023). In the MovieLens example, movie titles and years such as Pinocchio (1940) and In the Line of Fire (1993) constitute the principal textual signal. User preferences are represented implicitly through chronological interaction histories rather than explicit preference vectors.

The BART-based GenRec system uses a different representation. Users and items are serialized as text through the pretrained BART tokenizer, but explicit user-ID and item-ID embeddings are added to token, positional, and identity representations:

$X_j=\mathrm{Emb}(S_j)+\mathrm{PosEmb}(j)+\mathrm{IDEmb}(\mathrm{ID}(j)).$

The model uses a six-layer encoder and six-layer decoder, 12 attention heads, hidden dimension 768, approximately 184 million parameters, and pretrained BART-base weights (Cao et al., 2024). Unlike prompt-based systems, it does not require manually authored natural-language instructions. Its input is a serialized interaction sequence with a mask indicating the prediction position.

A further industrial formulation uses semantic IDs rather than item names. Each item is represented by a short sequence of discrete tokens, for example

$\mathrm{SID}(v_i)=\{s_i^1,s_i^2,s_i^3\}.$

A decoder-only model receives historical semantic IDs and generates the semantic ID of a target item (Zou et al., 16 Apr 2026). This representation reduces the output vocabulary relative to atomic item identifiers, permits semantic sharing across items, and supports catalog-constrained decoding.

3. Major architectural variants

Textual-item generation

The LLaMA-based GenRec system represents interactions as natural-language item names and fine-tunes the LLM with specialized recommendation prompts (Ji et al., 2023). Its reported implementation uses LLaMA-LoRA, AdamW, five epochs, a batch size of 128, a peak learning rate of $3\times10^{-4}$ , a maximum input length of 256 tokens, and four NVIDIA RTX A5000 GPUs. It evaluates MovieLens 25M and Amazon Toys using HR@5, NDCG@5, HR@10, and NDCG@10.

On MovieLens 25M, GenRec reports HR@5 of 0.1034, NDCG@5 of 0.0716, HR@10 of 0.1311, and NDCG@10 of 0.0837, exceeding the P5 baseline on all reported metrics. On Amazon Toys, P5 performs better on all four metrics. The paper does not specify constrained decoding, exact catalog matching, or a complete procedure for converting generated text into catalog item identities.

Prompt-free masked generation

The BART-based GenRec uses single-item masked prediction during both pretraining and fine-tuning (Cao et al., 2024). During pretraining, one item in a sequence is replaced by [MASK], while the decoder generates the original item. During next-item fine-tuning, a mask is appended to the end of the historical sequence. The model can exploit bidirectional context during masked pretraining, while inference generates the next item autoregressively.

Beam search with beam size 20 produces the top 20 predictions in the all-item evaluation setting. The model is evaluated on Amazon Sports, Amazon Beauty, and Yelp. It reports HR@5, NDCG@5, HR@10, and NDCG@10. GenRec obtains HR@10 values of 0.0462 on Sports, 0.0641 on Beauty, and 0.0724 on Yelp. An ablation removing masked-item pretraining reduces performance across all reported metrics.

Semantic-ID generative retrieval

Semantic-ID systems replace item names or atomic IDs with structured discrete codes. TIGER, LETTER, MQL4Rec, MiniOneRec, and related models generate such codes directly. GenPlugin improves these systems using a dual-encoder, shared-decoder architecture (Yang et al., 4 Jul 2025). A language-semantics encoder processes item titles and descriptions, while an ID-semantics encoder processes generative item tokens. Contrastive alignment is applied at item and user-preference levels.

GenPlugin also introduces semantic substitution: predicted tokens from the language branch probabilistically replace ground-truth ID tokens during training. This exposes the decoder to imperfect prefixes and mitigates generation exposure bias. Retrieval augmentation further supplies relevant users through BM25 content retrieval and collaborative retrieval based on SASRec representations. GenPlugin improves TIGER, LETTER, and MQL4Rec across five Amazon domains, with particularly strong gains for combined medium- and long-tail items.

Industrial generative retrieval and ranking

The JD App GenRec framework uses a decoder-only Qwen2.5 backbone, three-token semantic IDs, Page-wise Next-Token Prediction, an asymmetric Token Merger, and GRPO-SR preference alignment (Zou et al., 16 Apr 2026). Page-wise NTP places multiple positive interactions into one ordered target sequence rather than creating separate point-wise examples with identical histories. This addresses one-to-many ambiguity in pagination and provides denser supervision.

The Token Merger compresses historical multi-token semantic IDs into one latent vector during prefilling while preserving full-resolution semantic-ID decoding at the output. The paper reports approximately a twofold prompt-length reduction with negligible accuracy loss. GRPO-SR combines group-relative policy optimization with positive-trajectory NLL regularization, a dense preference model, a relevance gate, and reward calibration for observed positive items. The reported Qwen2.5-3B model achieves HR@1 of 0.1189, HR@10 of 0.4456, HR@50 of 0.7192, NDCG@10 of 0.2635, NDCG@50 of 0.3247, and hallucination rate of 4.96%. GRPO-SR reduces hallucination rate to 2.68% and raises HR@50 to 0.7438. In a month-long online experiment, the paper reports a 9.5% improvement in click count and an 8.7% improvement in transaction count.

Netflix GenRec uses a different meaning of “generative.” It does not autoregressively generate item names or semantic IDs. Instead, a decoder-only LLM encodes a verbalized member history and request context, and a catalog-aware ranking head scores known catalog items (Li et al., 10 Aug 2026). The model operates in prefill-only mode, avoiding step-by-step decoding and beam search. Netflix reports approximately 40 times fewer Phase-2 labeled examples than its production ranker, an approximately 1.6% relative offline MRR gain, and a statistically significant online improvement of 0.006% relative in a four-week test using approximately 10% of traffic.

4. Training objectives and alignment mechanisms

Generative recommendation systems differ substantially in how they supervise item generation.

The LLaMA-based system uses prompt-based fine-tuning on user–item histories (Ji et al., 2023). The BART-based system uses masked item prediction, maintaining the same sequence-to-sequence objective during pretraining and next-item fine-tuning (Cao et al., 2024). Semantic-ID systems generally use autoregressive token-level cross-entropy over item codes.

Reward-based post-training addresses the mismatch between behavioral logs and desired recommendation quality. The JD GenRec system uses GRPO-SR with dense preference rewards, a relevance gate, and NLL regularization (Zou et al., 16 Apr 2026). S-GRec uses an offline Personalized Semantic Judge and Asymmetric Advantage Policy Optimization (Jiang et al., 11 Feb 2026). Its judge decomposes relevance into User Profile Relevance, Future Interest Relevance, Novelty, and Contextual Relevance, then learns user-conditional aggregation weights. Semantic advantages are added only when their signs agree with business advantages, and their magnitude is bounded by the business advantage.

S-GRec uses Qwen3-4B for the offline semantic judge and Qwen2.5-1.5B with RQ-VAE tokenization for the online generator (Jiang et al., 11 Feb 2026). The reported public-data results are HR@10 of 0.1632 and NDCG@10 of 0.1202 on Industrial, and HR@10 of 0.1689 and NDCG@10 of 0.1308 on Office. In Tencent WeChat Channels advertising, the paper reports a 1.19% GMV lift, a 1.16% CTR lift, and a 2.02% reduction in dislike rate without real-time LLM inference.

HCGRec addresses a different failure mode in reward-based semantic-ID optimization (Zhang et al., 12 Aug 2026). If all finite rollouts enter the wrong early semantic branch, they may receive identical zero rewards, producing zero group-relative advantages. HCGRec diagnoses reachability using checkpoint rollouts and supplies the shortest target prefix that enables at least one successful completion. It applies supervised learning to hinted prefix tokens and GRPO only to sampled suffix tokens. The paper reports that inactive groups fall from approximately 55–63% to approximately 12–17% in displayed experiments, while the abstract summarizes a reduction from above 70% to below 20%.

5. Representation, scalability, and bias mitigation

Representation design is a central research problem in GenRec. Rec2PM compresses long user histories into persistent Preference Memory tokens (Chen et al., 12 Feb 2026). It separates recent Working Memory from compact Preference Memory and uses self-referential teacher forcing: global raw-history computations generate reference memories that supervise parallel segment-level recurrent updates. At inference, memory is updated iteratively.

With overwriting, the memory remains fixed at $C$ tokens. The reported MerRec configuration uses $C=4$ and embedding dimension 64. Rec2PM-O improves HSTU Full from H@10 of 42.77 to 44.20 and reduces per-user storage to 1 KB in the reported HSTU timing comparison, compared with 32 KB for KV-Mask-O. An industrial experiment uses 20 memory tokens and reports H@50 of 6.64 and H@1000 of 33.06, with approximately 8% of full-model latency.

AsymRec addresses a different representation bottleneck by decoupling continuous input representations from discrete output targets (Huang et al., 14 May 2026). Multi-expert Semantic Projection maps continuous item embeddings into the Transformer hidden space, while Multi-faceted Hierarchical Quantization produces structured discrete targets. On Beauty, AsymRec achieves NDCG@10 of 0.0516, compared with 0.0491 when discrete codes are used as inputs and 0.0406 when continuous embeddings are used as outputs. Across four Amazon datasets, the paper reports an average 15.8% improvement over the strongest baseline on NDCG@10.

HD-Rec extends semantic-ID generation to cross-domain recommendation (He et al., 7 Aug 2026). Its Hierarchical Domain-Aware Quantizer uses globally shared coarse codebooks and adaptively routed fine-level codebooks. Its Domain-Adaptive Sparse Mixture-of-Experts combines an always-active shared expert with one selected specialized expert. Cross-Granularity Routing Consistency Learning aligns token-level and item-level routing distributions. HD-Rec improves HR@10 over the strongest baseline on Clothing, Sports, Electronics, Phones, Books, and Movies by 4.9%, 17.6%, 16.3%, 9.9%, 3.4%, and 0.4%, respectively.

Popularity bias can arise within semantic tokenization itself. CRAB splits over-popular codebook tokens while preserving their hierarchical child structure and applies a tree-structured regularizer (Fan et al., 6 Apr 2026). On Industrial, CRAB matches MOR’s HR@10 of 0.152, raises NDCG@10 from 0.116 to 0.117, lowers DGU@10 from 0.418 to 0.356, and lowers MGU@10 from 0.109 to 0.091. SPRINT addresses popularity amplification in Transformer attention and feed-forward layers by regularizing attention column sums and feed-forward spectral norms (Yang et al., 20 Jun 2026). On evaluated TIGER and LETTER systems, SPRINT improves both accuracy and long-tail exposure.

Algorithmic recourse

GenRe is a conditional generative model for algorithmic recourse, not a recommender system (Garg et al., 12 May 2025). It learns a distribution

$R(x^+\mid x^-)\propto \exp[-\lambda C(x^-,x^+)]P(x^+\mid y^+)V(x^+),$

where $x^-$ is an unfavorable instance and $t^u_{1:n}$ 0 is a plausible favorable alternative. It synthesizes negative–positive training pairs from labeled data, trains an autoregressive Transformer, and uses forward sampling rather than gradient-based search at inference. On Adult, COMPAS, and FICO HELOC, GenRe obtains combined scores of 1.93, 1.89, and 1.90, respectively.

Genetic pedigree reconstruction

Rec-Gen, written as $t^u_{1:n}$ 1, reconstructs a large induced subpedigree from extant genetic sequences under an idealized stochastic pedigree model (Kim et al., 2020). It identifies sibling triples through mutual block overlap, recursively infers parental couples, and reconstructs the coupled pedigree up to permutation of partners and block phasing. Its formal theorem gives a recovered fraction $t^u_{1:n}$ 2 satisfying $t^u_{1:n}$ 3 as the expected offspring parameter $t^u_{1:n}$ 4 grows, under a sequence-length regime $t^u_{1:n}$ 5 and depth $t^u_{1:n}$ 6. The model assumes no mutation, no recombination, haploid individuals, monogamy, and an effectively infinite alphabet.

Synthetic preference-data generation

Another GenRec generates synthetic user–item interactions from latent user preference vectors, item attribute vectors, user populations, item categories, popularity scores, and user-specific interaction counts (Mungari et al., 2024). Binary interactions are sampled from utility-dependent Bernoulli probabilities, while long-tail distributions control item popularity and user engagement. The framework controls parameters such as $t^u_{1:n}$ 7, $t^u_{1:n}$ 8, $t^u_{1:n}$ 9, $\hat{t}_{n+1}^u=\arg\max_t P(t\mid u,t^u_{1:n}).$ 0, and $\hat{t}_{n+1}^u=\arg\max_t P(t\mid u,t^u_{1:n}).$ 1. It generates population-specific item-adoption patterns but does not model user–user social edges, diffusion, influence cascades, or evolving communities. The supplied paper does not describe the HYDRA model named in its abstract; its detailed method concerns GenRec.

Video generation and recognition

Video GenRec unifies video generation and recognition through a Stable Video Diffusion backbone with random-frame conditioning (Weng et al., 2024). A shared spatial-temporal UNet supports latent reconstruction and classification. The model reports 75.8% accuracy on Something-Something V2, 87.2% on Kinetics-400, 46.5 FVD on SSV2, and 49.3 FVD on Epic-Kitchens-100. Its random-frame conditioning supports first-frame generation, interpolation, sparse-frame completion, unconditional generation, and recognition from incomplete videos.

Novel-view synthesis

A separate GenRec distinguishes reconstructable pixels from genuinely unobserved pixels in sparse-input novel-view synthesis (Çelen et al., 18 Aug 2026). A depth-derived observation mask identifies regions with source-camera evidence. A flow-matching backbone jointly predicts RGB and scene-coordinate maps, while a pixel-space refinement branch restores high-frequency details only in observed regions. The final image is

$\hat{t}_{n+1}^u=\arg\max_t P(t\mid u,t^u_{1:n}).$ 2

The model is evaluated on RealEstate10K, DL3DV-10K, and Mip-NeRF 360 in single-view extrapolation and two-view interpolation settings.

7. Open problems and methodological limitations

Across recommendation-oriented GenRec systems, several recurring issues remain unresolved. Open-ended text generation requires catalog validation, disambiguation of duplicate titles, and handling of invalid outputs. Semantic-ID systems require stable codebooks, collision handling, valid-prefix decoding, and protection against popularity concentration in shared token branches. Reward-based post-training requires calibrated rewards, sufficient rollout diversity, and mechanisms for handling zero-variance groups, reward hacking, and conflict between semantic relevance and business objectives.

Comparability is also limited. The supplied studies use different datasets, catalog representations, evaluation protocols, baselines, and decoding procedures. Some compare against public baseline numbers rather than rerunning all systems under identical conditions. Several industrial papers omit exact prompt templates, reward formulas, model dimensions, negative-sampling procedures, latency distributions, and A/B-test confidence intervals.

The broader research direction is therefore not a single algorithm but a family of representation and optimization strategies:

$\hat{t}_{n+1}^u=\arg\max_t P(t\mid u,t^u_{1:n}).$ 3

Current work addresses different stages of this pipeline: masked-item pretraining, semantic-ID quantization, memory compression, exposure-bias mitigation, long-tail retrieval augmentation, codebook rebalancing, spectral regularization, domain-adaptive routing, page-level supervision, and offline LLM-based semantic judging. GenRec consequently denotes a broad and rapidly evolving methodological family rather than one uniquely defined model.