OneRec Series: Unified Recommender Systems

• OneRec Series is a unified recommender system that integrates candidate retrieval, ranking, and explanation generation into a single generative model.
• It employs advanced Transformer-based architectures including encoder-decoder and decoder-only models to achieve large-scale, industrial-grade performance.
• The series leverages MoE layers, reinforcement learning, and multimodal tokenization to significantly reduce computational overhead and enhance recommendation accuracy.

The OneRec Series designates a lineage of recommender system architectures that unify candidate retrieval, ranking, and, in advanced iterations, reasoning and explanation generation within a single generative modeling framework. Emerging from foundational work on unified embedding and recall-ranking models—exemplified by UniRec—the OneRec paradigm has rapidly advanced to large-scale, highly efficient Transformer-based systems deployed in major industrial contexts. Notable milestones include encoder–decoder designs, pure decoder-only "lazy" architectures for maximal compute/utilization, multimodal tokenization, and instruction-following capabilities. Throughout its evolution, the series has demonstrated robust empirical scaling laws, substantial reductions in resource requirements, and state-of-the-art performance on both public and private benchmarks.

1. Origins and Motivation

The OneRec paradigm grew out of limitations in conventional, multi-stage recommendation pipelines, which separate candidate retrieval and ranking using distinct models and infrastructure. This fragmentation increased computational, storage, and engineering overhead, and precluded globally optimal end-to-end learning. The foundational UniRec (Wu et al., 2021) introduced the "one-model" concept, showing that a unified architecture can generate both retrieval and ranking user embeddings in a single forward pass, using a basis-attention mechanism on user representations. This not only reduced computational cost and latency but also set the conceptual groundwork for generative, end-to-end modeling.

Subsequent OneRec series models generalized this approach, employing Transformer-based architectures and revisiting recommendation as an autoregressive token generation task, mapping diverse recommendation objectives (next-item prediction, session modeling, explanation, instruction following) to a standardized, flexible generative format (Zhou et al., 16 Jun 2025, Deng et al., 26 Feb 2025, Zhou et al., 31 Dec 2025).

2. Unified Generative Architectures

Core to the OneRec series is the formulation of recommendation as conditional language modeling over a vocabulary augmented with compact "itemic" or "semantic ID" tokens, derived from residual quantization of multimodal item embeddings (Zhou et al., 16 Jun 2025, Kong et al., 28 Oct 2025). This enables the entire recommendation workflow—from user history encoding to candidate generation and scoring—to be implemented as a single-stage, autoregressive generation process, typically realized atop decoder-only or encoder–decoder Transformer backbones.

Architectural innovations include:

• Itemic Tokens and Tokenizer: Item representations are compressed into discrete code sequences using hierarchical quantization (e.g., RQ-Kmeans or Residual Quantized VAE), mapped to the generative model's vocabulary (Zhou et al., 16 Jun 2025, Kong et al., 28 Oct 2025).
• Encoder–Decoder and Decoder-Only Regimes: Earlier iterations apply encoder–decoder structures to model rich user histories; OneRec-V2 shifts to a "lazy" decoder-only format, which leverages shared context key/value caching for efficiency, scaling, and maximal FLOPs utilization (Zhou et al., 28 Aug 2025). This design enables 94% reduction in training computation and supports parameter-efficient scaling to 8B+ models.
• Mixture-of-Experts (MoE): Many variants deploy sparse MoE layers in feedforward subnetworks, increasing model capacity linearly with expert count while maintaining per-token compute density (Deng et al., 26 Feb 2025, Zhou et al., 16 Jun 2025).

3. Training Paradigms and Reinforcement Learning Integration

Training is typically staged: initial pretraining using next-token cross-entropy over itemic-labeled sequences and multimodal captions, followed by supervised fine-tuning (SFT) on recommendation and instruction datasets, and culminating in reinforcement learning (RL) or preference optimization phases.

Notable RL advancements within the series include:

• Preference Alignment via Direct Preference Optimization (DPO): OneRec integrates session-level reward modeling and DPO loss, sampling candidate sessions via beam search and using learned reward functions or real-world feedback to construct hard-positive/negative pairs for iterative optimization (Deng et al., 26 Feb 2025).
• Group-Relative Policy Optimization (GRPO) and Early-Clipped GRPO (ECPO): These variants enable stable, sample-efficient policy updates, integrating soft ranking within candidate sets and custom regularization to prevent reward hacking or sample degeneracy (Zhou et al., 16 Jun 2025, Kong et al., 28 Oct 2025, Zhou et al., 31 Dec 2025).
• Real-World Feedback and Duration-Aware Reward Shaping: OneRec-V2 specifically aligns with real user feedback using quantile-based reward construction and adaptive ratio clipping (Gradient-Bounded Policy Optimization), promoting robustness against gradient explosion on negative samples and improving reward signal fidelity (Zhou et al., 28 Aug 2025).

4. Reasoning, Interpretability, and Foundation Capabilities

Whereas initial OneRec variants functioned as implicit, black-box predictors, newer frameworks such as OneRec-Think explicitly scaffold in-text reasoning through a multi-stage process (Liu et al., 13 Oct 2025):

• Itemic Alignment: Task curriculum aligns itemic code sequences with their natural language semantics for grounded token-text translation.
• Reasoning Activation: Chain-of-Thought (CoT) style prompting and rationale bootstrapping enable supervised fine-tuning on both rationale generation and item prediction, promoting human-interpretable, step-wise rationales.
• Reasoning Enhancement: RL with multi-valid preference-aware rewards, such as rollout-beam rewards, reinforces both rationale quality and accurate, diverse recommendation generation.
• Think-Ahead Architecture: Decouples CoT reasoning and heavy decoding to offline stages while constraining the online component to prefix-conditioned finalization, ensuring production-grade latency even in explicit reasoning deployments.

Foundation-scale models are benchmarked on RecIF-Bench, evaluating capabilities across alignment, recommendation, instruction following, and explanation (Zhou et al., 31 Dec 2025).

5. Empirical Scaling Laws and Public Reference Frameworks

Empirical studies across industrial and public benchmarks confirm that generative models with itemic SID vocabularies and deep transformer backbones exhibit consistent downward trends in loss as model size increases, contrasting with saturation in classical embedding-heavy recommenders (Zhou et al., 16 Jun 2025, Kong et al., 28 Oct 2025). The scaling optimality frontier is characterized by data-hungry regimes—adding training sequences often outperforms further parameter scaling.

Open-source frameworks such as MiniOneRec have demonstrated, on public Amazon datasets, that full-process SID alignment and GRPO with hybrid rule+rank rewards suffice for closing the accuracy gap with proprietary systems (Kong et al., 28 Oct 2025). Key characteristics include compact SID vocabularies, low-latency constrained decoding (beam search), and parameter-efficient transferability across domains.

6. Industrial Deployment and Real-World Impact

OneRec and its descendants have seen deployment at scale in highly-trafficked platforms such as Kuaishou and Kuaishou Lite (Deng et al., 26 Feb 2025, Zhou et al., 16 Jun 2025, Zhou et al., 28 Aug 2025, Liu et al., 13 Oct 2025):

• Latency and OPEX: Single-stage generative architectures have increased model FLOPs utilization (MFU) from <5% to ~25–30%, with infrastructure (communication, storage, OPEX) reduced to a fraction (e.g., 10.6%) of that in classical cascaded pipelines.
• Business metrics: Deployed models consistently achieve percentage-level lifts in core metrics such as App Stay Time (+0.54% to +1.24% for OneRec; +0.467% to +0.741% for OneRec-V2), watch time, and 7-day lifetime, with further gains seen in advanced variants (e.g., +0.159% with explicit reasoning in OneRec-Think).
• Practical lessons: Tokenization accuracy, codebook entropy, MoE balancing, RL reward shaping, and decoupled sample generation are highlighted as key operational levers. Pitfalls include reward hacking, over-large search spaces, and negative-advantage gradient squeezing.

7. Challenges and Future Directions

Despite its successes, the OneRec series faces ongoing challenges (Zhou et al., 31 Dec 2025):

• Tokenizer/quantizer transferability: Hierarchical itemic codes trained in a single domain may not extend optimally to novel or cross-domain contexts.
• Data-mixing strategies: Rebalancing recommendation-domain and general-domain corpora during pretraining to optimize world knowledge retention remains an open area.
• Consistent reasoning benefits: While CoT scaffolding improves rationales, its impact is not uniform across tasks; further innovation in self-consistent decoding and prompt design is warranted.
• External memory integration: Handling dynamic catalogs with external retrieval or memory components, while adhering to the generative paradigm, is under-explored.
• Fairness, robustness, interpretability: As recommendation systems become more autonomous and opaque, stringent guarantees on bias, robustness, and transparency are critical.

Research continues with open benchmarks (RecIF-Bench), large-scale public datasets, and comprehensive code releases, providing a foundation for further progress toward instruction-guided, intelligent, and interpretable recommendation systems (Zhou et al., 31 Dec 2025).