OneRec-V2: Industrial Generative Recommender

• The paper introduces OneRec-V2 as an industrial-scale generative recommender that formulates recommendation as conditional sequence generation over semantic tokens.
• It leverages a lazy decoder-only architecture with shared static key-values and grouped query attention, reallocating compute from context encoding to target generation with up to 94% reduction in operations.
• The system aligns user preferences using duration-aware reward shaping and gradient-bounded policy optimization, leading to measurable improvements in engagement metrics during deployment.

Searching arXiv for the specified papers to ground the article. OneRec-V2 is an industrial-scale generative recommender that unifies retrieval and ranking as conditional sequence generation, while reworking the original OneRec design around a lazy decoder-only architecture and direct preference alignment from real-world user feedback. In the reported Kuaishou deployments, it is presented as both a modeling advance over OneRec-V1—primarily by reallocating compute from context encoding to the generative decision path—and a systems target for large-scale optimized serving, including FP8 post-training quantization in production (Zhou et al., 28 Aug 2025, Su et al., 12 Mar 2026).

1. Lineage, definition, and problem formulation

OneRec established the underlying generative recommendation paradigm by replacing the classical retrieve–pre-rank–rank cascade with an end-to-end model that directly generates item semantic IDs from user context, thereby unifying retrieval and ranking in a single objective (Deng et al., 26 Feb 2025). OneRec-V2 retains that end-to-end framing, but the technical report identifies two concrete limitations in V1: first, an encoder–decoder allocation problem in which most FLOPs were spent on processing long user context rather than on the short target sequence; second, reinforcement learning that relied solely on reward models, with expensive on-policy rollouts for only a small slice of traffic and exposure to reward hacking against proxy objectives (Zhou et al., 28 Aug 2025).

In OneRec-V2, recommendation is formulated as autoregressive sequence generation over semantic item tokens. Each candidate item is represented by three semantic IDs, $s^1, s^2, s^3$, produced by a semantic tokenizer. Training uses teacher forcing with a “new-impression-only” loss placement: the newest impressed item contributes to the loss, while earlier items serve only as conditioning context. The reported generative objective is

$$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$

The “Context” is a heterogeneous user history composed of static profile and behavioral features organized chronologically. This formulation keeps the optimization target aligned with the next recommendation decision rather than with reconstruction of the entire history (Zhou et al., 28 Aug 2025).

A concise way to understand the architectural transition is that V1 demonstrated that a sufficiently scaled generative recommender could replace the cascaded stack, whereas V2 reorganizes that generative formulation so that almost all compute is spent on the actual decision path rather than on repeated context encoding (Deng et al., 26 Feb 2025, Zhou et al., 28 Aug 2025).

2. Lazy decoder-only architecture and compute reallocation

The central architectural contribution of OneRec-V2 is the lazy decoder-only transformer. Instead of an encoder–decoder stack, V2 treats context as static conditioning accessed through cross-attention. A Context Processor ingests heterogeneous profile and behavior features, concatenates them into a unified context tensor, and emits layer-shared key–value pairs without per-layer K/V projections. These KV sets are shared across blocks, normalized via RMSNorm, and combined with Grouped Query Attention (GQA), in which the number of KV head groups is smaller than the number of query heads (Zhou et al., 28 Aug 2025).

Each lazy decoder block applies three operations in sequence: lazy cross-attention with KV sharing, causal self-attention over the generated tokens, and a feed-forward network; in deeper layers, the FFN may be replaced by Mixture-of-Experts. The target sequence is intentionally short—reported as $[\text{BOS}, s^1, s^2]$—and the model predicts the three semantic IDs for the recommendation. KV sharing maps decoder layer $l$ to a KV index $$l_{\text{kv}} = \lfloor l \cdot L_{\text{kv}} / N_{\text{layer}} \rfloor$$, and the report notes that values can be tied to keys to reduce memory further (Zhou et al., 28 Aug 2025).

The motivation is quantitative rather than stylistic. In OneRec-V1, with context length $N=512$, 97.66% of FLOPs were spent on sequence encoding and only 2.34% on target generation; at $N=3000$, the target proportion dropped to 0.41%. Under the V2 lazy decoder-only design, the report states that the target proportion rises to approximately 100%. For a controlled 1B comparison, encoder–decoder models consumed 204–296 GFLOPs per sample with 12.2–17.6B activated elements, whereas the lazy decoder-only used 18.9 GFLOPs and 1.24B activations, with convergence loss around 3.27. The reported headline effect is a 94% reduction in total computation and a 90% reduction in training resources, enabling successful scaling to 8B parameters (Zhou et al., 28 Aug 2025).

The complexity analysis in the report makes the same point in asymptotic terms. Because context length dominates in industrial settings, conventional architectures incur $O(N)$ FFN and projection cost on the context path. Lazy decoding collapses the compute-heavy path to the short target length $T \approx 3$, while static KVs are shared rather than recomputed. This suggests that the architectural change is less a minor decoder simplification than a redistribution of the computational budget toward the exact portion of the network that determines the next recommendation (Zhou et al., 28 Aug 2025).

3. Preference alignment from real-world user feedback

OneRec-V2 also redefines preference alignment. The V1 line incorporated reward-model-based optimization, including Iterative Preference Alignment with DPO in the original OneRec formulation, but the V2 report argues that relying solely on reward models constrains scalability and may produce reward hacking (Deng et al., 26 Feb 2025, Zhou et al., 28 Aug 2025). V2 therefore introduces preference alignment directly from logged user interactions through Duration-Aware Reward Shaping and Gradient-Bounded Policy Optimization (GBPO).

Duration-Aware Reward Shaping first normalizes watch-time signals within duration buckets formed per user using a logarithmic partition,

$$\mathcal{F}(d) = \left\lfloor \log_{\beta}(d+\epsilon) \right\rfloor,$$

and computes a percentile score $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$0 for an item within the user’s historical set for that bucket. Positives are defined as the top quartile by $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$1 within the batch and bucket, while explicit dislikes are negatives. Advantages are then assigned as

$$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$2

This produces dense, duration-normalized reward signals intended to correlate with App Stay Time while filtering out raw duration bias and low-quality samples (Zhou et al., 28 Aug 2025).

GBPO replaces explicit ratio clipping with a gradient-bounded construction derived from the BCE loss structure. The logging policy is handled differently depending on exposure source. For OneRec-generated samples, $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$3 is the logged generation probability at exposure time. For traditional pipeline samples, $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$4 is approximated as the stop-gradient of the current model probability, which makes the policy ratio equal to 1 for those samples. The report emphasizes that gradient bounding remains necessary, especially on negative samples, to avoid instability (Zhou et al., 28 Aug 2025).

The alignment regime is not purely single-objective. V2 combines user-feedback signals—especially duration-normalized watch-time percentile and explicit dislikes—with a reward model that fuses multiple interaction objectives such as likes, follows, and comments. The report describes the resulting training as hybrid rather than strictly additive, with the intent of balancing stay-time and interaction metrics without seesaw effects. In ablations on Kuaishou, reward-model RL yielded +0.269% App Stay Time, user-feedback RL yielded +0.299%, and the hybrid regime yielded +0.283%; the interaction profile differed by reward source, with reward-model RL tending to favor interactions more and user-feedback RL favoring stay-time (Zhou et al., 28 Aug 2025).

A common misconception is that V2 eliminates reward modeling entirely. The reported design instead demotes reward models from being the sole source of alignment to being one component in a hybrid preference signal grounded in real-world user feedback (Zhou et al., 28 Aug 2025).

4. Scaling behavior, serving configuration, and operational profile

The V2 report presents OneRec-V2 as a model that scales smoothly in both dense and sparse configurations. Dense lazy decoder models are reported at 0.1B, 0.5B, 1B, 2B, 4B, and 8B parameters, with convergence loss decreasing from 3.57 to 3.19 as model size increases. A sparse MoE variant is also reported: 4B total parameters, 0.5B active, with 53 routed experts plus 1 shared expert and top-3 routing, reaching convergence loss 3.22 at roughly 0.5B compute (Zhou et al., 28 Aug 2025).

The serving configuration described for the principal online experiment uses a 1B model, context length of approximately 3000, beam size 512, and L20 GPUs. In this setting, the model achieved 36 ms latency and 62% MFU in online serving. MFU is defined as

$$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$5

and the report uses it to argue that V2’s design translates architectural efficiency into high practical utilization rather than merely lower nominal FLOPs (Zhou et al., 28 Aug 2025).

Several serving optimizations are coupled to the lazy decoder design. Because context is static during a recommendation window and often reused across multiple impressed items, common context compression reduces effective context length from $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$6 to $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$7, where $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$8 is commonly 5. The shared KV representation can then be reused across layers and items, functioning analogously to a cross-attention KV cache. Combined with the removal of per-layer K/V projections, this reduces context-side compute and memory traffic while preserving long-context personalization (Zhou et al., 28 Aug 2025).

Relative to the broader OneRec line, this marks a shift from demonstrating that end-to-end generation can replace the classical pipeline to demonstrating that such a generative recommender can also be compute-efficient at production scale. The earlier OneRec report emphasized 23.7% MFU in training and 28.8% in inference for V1-style deployment, whereas the V2 report presents a serving configuration with 62% MFU at 36 ms latency, reflecting a more tightly compute-centric design (Zhou et al., 16 Jun 2025, Zhou et al., 28 Aug 2025).

5. FP8 quantized inference and systems co-design

A separate systems study analyzes quantized inference for OneRec-V2 and frames it as a recommender whose numerical behavior is closer to LLMs than to traditional industrial recommendation stacks. At the operator level, the production-scale serving model is described as Transformer-style blocks plus a fat-MoE: attention with linear $$\mathcal{L}_{\text{Gen}} = -\frac{1}{3}\sum_{i=1}^{3}\log p\big(s^i \mid \text{BOS}, s^{<i}, \text{Context}\big).$$9 projections, dense FFNs, sparse expert selection executed as grouped GEMMs, learned embeddings, residual pathways, and normalization around attention and FFN blocks (Su et al., 12 Mar 2026).

The reported reason FP8 post-training quantization is viable is that OneRec-V2’s weights and activations are “significantly more controlled” than those of traditional recommenders. The study reports that a traditional recommendation model exhibits mean weight variance of approximately $[\text{BOS}, s^1, s^2]$0, mean AbsMax greater than $[\text{BOS}, s^1, s^2]$1, and activation mean variance of approximately $[\text{BOS}, s^1, s^2]$2. By contrast, Qwen3-8B has mean weight variance around $[\text{BOS}, s^1, s^2]$3 and extreme values around $[\text{BOS}, s^1, s^2]$4, while OneRec-V2 has mean weight variance below $[\text{BOS}, s^1, s^2]$5 and activation statistics orders of magnitude smaller than traditional recommender baselines. The paper attributes this to Transformer-style normalization, residual pathways, and gating, noting that in the fat-MoE configuration roughly 0.5B parameters are activated per token out of a 4B backbone (Su et al., 12 Mar 2026).

The FP8 PTQ framework is selective rather than global. Quantization is applied to the computation-dominant operators only: attention projection linears, FFN matmuls, and grouped GEMMs in sparse MoE. Embeddings, normalization, softmax, gating, and routing remain in higher precision. Weight tensors use offline per-channel scaling for ordinary linear layers and 128×128 block-wise scaling for MoE weights; activations use runtime per-token row scaling for ordinary linears and 1×128 block-wise scaling for MoE activations. The system stores pre-quantized FP8 weights with FP32 scales on GPU and uses FP8 TensorCore multiplication with FP32 accumulation, after which outputs are cast back to FP16 for downstream layers (Su et al., 12 Mar 2026).

This quantization regime is integrated with an optimized serving stack rather than evaluated in isolation. The study reports direct construction of the TensorRT execution graph, fused quantization-plus-GEMM kernels, RadixTopK with kernel fusion and zero-copy pathways, an attention kernel redesigned for large-batch short-context inference, and Hopper TMA-enabled grouped GEMM kernels for MoE. At batch size 32, the end-to-end result is a latency reduction from 139 ms to 70 ms, a 49% decrease, and throughput improvement from 205 to 394, a 92% increase. The gain breakdown is reported as +27% from infrastructure upgrades, +42% from enabling FP8 on Linear and MoE operators, and +23% from operator-level optimizations (Su et al., 12 Mar 2026).

Online A/B evaluation in that study reports no degradation in core metrics. For the Kuaishou app, App Stay Time is +0.047%, Watch Time +0.044%, Video View +0.073%, Like +0.085%, Follow +0.020%, Comment +0.805%, Collect +0.430%, and Forward +0.313%. For Kuaishou Lite, the paper reports small fluctuations, including Watch Time at −0.157% and Forward at +1.047%, and characterizes the overall result as “no degradation” (Su et al., 12 Mar 2026).

6. Empirical impact, ecosystem trade-offs, and limitations

The principal V2 online deployment reported in the technical report ran for one week on 5% traffic. Relative to OneRec-V1, OneRec-V2 improved App Stay Time by 0.467% on Kuaishou and 0.741% on Kuaishou Lite, while LT7 increased by 0.069% and 0.034%, respectively. Interaction metrics also rose broadly. On Kuaishou, the reported changes were Watch Time +1.367%, Video View +0.331%, Like +3.924%, Follow +4.730%, Comment +5.394%, Collect +2.112%, and Forward +3.183%. On Kuaishou Lite, the corresponding changes were Watch Time +0.762%, Video View +0.259%, Like +5.393%, Follow +5.627%, Comment +5.013%, Collect +3.202%, and Forward +7.958% (Zhou et al., 28 Aug 2025).

The same report also documents trade-offs that matter for ecosystem-aware evaluation. In a stricter 1% group with caching disabled, interaction metrics increased sharply—for example, Kuaishou Comment +29.249% and Forward +24.741%, and Kuaishou Lite Forward +30.957%—but cold-start video views declined by 44.704% and 36.730%, respectively, while content cluster density rose by 11.692% and 7.933%. This is an explicit reminder that maximizing short-term engagement in a fully generative recommender can alter exposure diversity and cold-start behavior (Zhou et al., 28 Aug 2025).

Several limitations are stated or directly implied in the V2 literature. The technical report notes potential bias toward longer content despite duration normalization, incomplete off-policy correction for traditional-pipeline samples because the effective ratio is 1 there, and scaling laws that do not cleanly follow canonical patterns beyond 2B. The quantized inference study is limited to FP8; INT8, FP6, and FP4 are not explored, and the serving stack requires substantial infrastructure, including direct TensorRT graph construction, fused kernels, and Hopper-specific features (Zhou et al., 28 Aug 2025, Su et al., 12 Mar 2026).

Three misconceptions are therefore worth rejecting. First, OneRec-V2 is not merely a faster version of V1; its core contribution is a reallocation of model computation around lazy decoding and real-feedback alignment. Second, it is not a context-free decoder-only recommender; context remains central, but it is accessed through shared static KVs rather than repeatedly encoded. Third, the strong latency results should not be read as a purely systems-side artifact, because the FP8 study explicitly links quantization robustness to V2’s LLM-like architectural regularization and dense compute profile (Zhou et al., 28 Aug 2025, Su et al., 12 Mar 2026).