REG4Rec: Reasoning-Enhanced Generative Rec

• REG4Rec is a generative model for sequential recommendation that reformulates next-item prediction as generation over unordered semantic tokens.
• It employs a MoE-based parallel quantization approach to encode multimodal features, enabling diverse and robust reasoning pathways.
• The model integrates reinforcement learning, multi-step reward augmentation, and self-reflection to improve recommendation accuracy and scalability.

REG4Rec is a reasoning-enhanced generative model for large-scale sequential recommendation. It formulates next-item prediction as generation over semantic tokens rather than direct scoring over a giant item-ID softmax, and augments that generative process with dynamic multi-step reasoning, preference-aligned reinforcement learning, and inference-time self-reflection (Xing et al., 21 Aug 2025). The model is designed for recommendation settings with tens of millions of items and billions of interactions, where traditional ID-based sequential recommendation and earlier generative recommendation methods are constrained by sparse supervision, rigid item representations, and limited mechanisms for reasoning-path reliability (Xing et al., 21 Aug 2025).

1. Problem setting and conceptual position

REG4Rec addresses sequential recommendation, where a user $u$ is associated with a chronological interaction sequence

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$

and the task is to predict the next item $i_{T+1}$, or more generally a distribution $p(i \mid \mathcal{S}_u)$ from which top-$N$ candidates are retrieved (Xing et al., 21 Aug 2025). In conventional sequential recommendation, exemplified by SASRec and BERT4Rec, models embed item IDs, process the sequence with a Transformer or RNN, and score all items by a large softmax. In very large catalogs, this creates difficulties associated with huge output layers, sparse training for long-tail items, weak use of multimodal semantics, and limited interpretability.

Generative recommendation replaces numeric IDs with compact semantic tokens and predicts the token sequence of the next item. Earlier systems such as TIGER and COBRA use residual quantization VAE-style tokenization, mapping each item to a single fixed ordered token sequence (Xing et al., 21 Aug 2025). Reasoning-enhanced generative recommenders such as ReaRec and STREAM introduce intermediate reasoning steps, but the paper characterizes them as still limited by the singularity of item semantic representations, narrow reasoning diversity, and insufficient reliability calibration (Xing et al., 21 Aug 2025).

REG4Rec is proposed around two explicit research questions: how to design flexible, diverse reasoning paths that capture heterogeneous user intents, and how to ensure that these paths are reliable, semantically coherent, and robust to noise (Xing et al., 21 Aug 2025). Its answer is a coupled architecture in which items receive multiple unordered semantic tokens, the model dynamically selects reasoning steps across those token spaces, reinforcement learning aligns reasoning traces with recommendation objectives, and a self-reflection module prunes inconsistent inference paths.

2. Item representation and reasoning space

The central representational device in REG4Rec is the MoE-based parallel quantization codebook, or MPQ (Xing et al., 21 Aug 2025). Each item is encoded from a multimodal feature vector $v_i \in \mathbb{R}^d$, built by concatenating text, image, and behavioral features extracted by pretrained models such as Qwen3-Embedding and Pailitao. Instead of assigning one deterministic ordered token sequence to an item, MPQ assigns $M$ semantic tokens, one from each expert or codebook, and these tokens are explicitly unordered.

For expert $r$, with codebook

$\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$

the expert embedding is

$e^r_i = E_r(v_i),$

and quantization selects the nearest codeword: $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$0 A gating network then forms the reconstructed encoding

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$1

Two training losses shape MPQ. The reconstruction term is

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$2

and the orthogonality term is

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$3

giving the combined objective

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$4

The orthogonality term encourages different experts to capture complementary semantic dimensions.

This design changes the structure of the reasoning space. Because each item is represented by multiple unordered semantic tokens rather than one fixed ordered path, the model can choose different subsets, different orderings, and different semantic facets during generation. The paper interprets this as a larger-scale diverse reasoning space, permitting recommendation to proceed through brand, category, style, price range, or other latent facets depending on the user context (Xing et al., 21 Aug 2025). A plausible implication is that REG4Rec moves generative recommendation from token decoding over a static code to path selection over a combinatorial semantic state space.

3. Core architecture and reasoning mechanisms

REG4Rec organizes its pipeline into codebook construction, training, and inference, with six named components that define the model’s operational behavior (Xing et al., 21 Aug 2025).

Component	Role	Mechanism
MPQ	Item tokenization	Multiple unordered semantic tokens via parallel codebooks
CRSS	Step selection	Chooses the most confident token across active codebooks
PARS	Reasoning alignment	RL reward over token hits, category hits, consistency, robustness
MSRA	Horizon extension	Time-decayed rewards over multiple future items
CORP	Self-reflection	JS-divergence pruning and rollback of inconsistent paths
LADQ	Scalability	Layer-adaptive fp32/bf16/fp8 training

The generative model is a Transformer encoder-decoder trained to predict the next item’s semantic tokens. During pretraining, the token-level objective is

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$5

and an auxiliary category prediction loss

$$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$6

provides a step-wise semantic signal that is later reused for consistency measurement (Xing et al., 21 Aug 2025).

A reasoning path is a sequence of generated tokens $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$7, together with the associated step-wise category distributions $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$8 and predicted categories. The dynamic order of those reasoning steps is determined by Confidence-based Reasoning Step Selection, or CRSS. If the active codebooks at step $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$9 are

$i_{T+1}$0

CRSS selects

$i_{T+1}$1

At each step, the model therefore takes the highest-confidence token across all active codebooks, then removes that codebook from further use. The resulting ordering is data-dependent rather than hard-coded.

The post-training stage is dominated by Preference Alignment for Reasoning, or PARS. It treats token-sequence generation as a policy $i_{T+1}$2, samples multiple paths from an older policy, and optimizes a GRPO-style clipped objective using normalized advantages (Xing et al., 21 Aug 2025). The reward has four components. The step-hit reward verifies whether generated tokens belong to the target item’s token set; the category-hit reward checks alignment of step-wise category predictions; the step-consistency reward penalizes abrupt semantic drift using Jensen-Shannon divergence between adjacent category distributions; and the global path reward tests whether partially corrupted reasoning paths still retrieve the target item in top-$i_{T+1}$3. These are combined as

$i_{T+1}$4

Multi-Step Reward Augmentation, or MSRA, extends this reward beyond the immediate next item to the next $i_{T+1}$5 items, using exponentially decayed weights $i_{T+1}$6. This is intended to stabilize learning under behavioral noise and to capture short-horizon long-term preference persistence (Xing et al., 21 Aug 2025).

4. Training, inference, and industrial scalability

REG4Rec has a two-stage optimization regime: pretraining and RL-based post-training (Xing et al., 21 Aug 2025). Pretraining learns next-item token generation and auxiliary category prediction. Post-training then applies PARS and MSRA to optimize reasoning paths under recommendation-specific rewards. The policy is regularized with a KL term relative to the older policy, following the GRPO-style objective described in the paper.

Inference introduces Consistency-Oriented Self-Reflection for Pruning, or CORP. At step $i_{T+1}$7, the model monitors the category predictor’s output distributions $i_{T+1}$8. For steps separated by $i_{T+1}$9, it computes $p(i \mid \mathcal{S}_u)$0; if the divergence exceeds a threshold $p(i \mid \mathcal{S}_u)$1, the current path is deemed inconsistent, terminated, rolled back, and regenerated (Xing et al., 21 Aug 2025). CORP is therefore an online pruning mechanism that prevents error propagation from semantically unstable reasoning trajectories.

The system also includes Layer-Adaptive Dynamic Quantization, or LADQ, to make RL post-training practical at industrial scale. LADQ periodically estimates layer-wise sensitivity and assigns precision from $p(i \mid \mathcal{S}_u)$2 according to the trade-off between accuracy sensitivity and latency contribution (Xing et al., 21 Aug 2025). On the industrial dataset reported in the paper, the fp32 baseline achieves 3.4 QPS with $p(i \mid \mathcal{S}_u)$3 and $p(i \mid \mathcal{S}_u)$4, whereas LADQ reaches 4.5 QPS, a $p(i \mid \mathcal{S}_u)$5 increase, with $p(i \mid \mathcal{S}_u)$6 and $p(i \mid \mathcal{S}_u)$7, corresponding to $p(i \mid \mathcal{S}_u)$8 and $p(i \mid \mathcal{S}_u)$9 changes respectively.

The implementation details underline the model’s engineering orientation. The paper reports 3 Transformer layers, specifically 1 encoder and 2 decoder layers, input dimension 128, hidden size 640, and MPQ with $N$0 codebooks and $N$1 codewords per codebook (Xing et al., 21 Aug 2025). Text and image features are 256-dimensional, derived from Qwen3-Embedding 7B and Pailitao v8. The industrial training setup uses approximately 35M users, 48M items, and 5.7B interactions, distributed across 10 parameter servers and 10 workers with 2 GPUs each.

5. Empirical performance and ablation results

The offline evaluation covers Amazon Product Reviews in Beauty, Sports and Outdoors, and Toys and Games, together with an industrial dataset comprising 35,154,135 users, 48,106,880 items, and 5,730,321,793 interactions (Xing et al., 21 Aug 2025). Metrics are Recall@K and NDCG@K. Baselines include SASRec, S$N$2-Rec, TIGER, COBRA, ReaRec, and STREAM.

Across all four datasets, REG4Rec is reported as the best-performing model. Relative to the best baseline, the improvements are $N$3 $N$4, $N$5 $N$6, $N$7 $N$8, and $N$9 $v_i \in \mathbb{R}^d$0 on Beauty; $v_i \in \mathbb{R}^d$1 $v_i \in \mathbb{R}^d$2, $v_i \in \mathbb{R}^d$3 $v_i \in \mathbb{R}^d$4, $v_i \in \mathbb{R}^d$5 $v_i \in \mathbb{R}^d$6, and $v_i \in \mathbb{R}^d$7 $v_i \in \mathbb{R}^d$8 on Sports; $v_i \in \mathbb{R}^d$9 $M$0, $M$1 $M$2, $M$3 $M$4, and $M$5 $M$6 on Toys; and $M$7 $M$8, $M$9 $r$0, $r$1 $r$2, and $r$3 $r$4 on the industrial dataset (Xing et al., 21 Aug 2025). The paper further summarizes the average relative improvements over the best baseline as approximately $r$5 on $r$6, $r$7 on $r$8, $r$9 on $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$0, and $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$1 on $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$2.

The industrial-dataset ablation study quantifies each component’s contribution. Full REG4Rec yields $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$3, $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$4, $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$5, and $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$6. Replacing MPQ with RQ-VAE gives $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$7, $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$8, $\mathcal{C}^r = \{z^r_1, z^r_2, \ldots, z^r_D\},$9, $e^r_i = E_r(v_i),$0, described as an approximately $e^r_i = E_r(v_i),$1 average performance drop. Removing CRSS gives $e^r_i = E_r(v_i),$2, $e^r_i = E_r(v_i),$3, $e^r_i = E_r(v_i),$4, $e^r_i = E_r(v_i),$5, an approximately $e^r_i = E_r(v_i),$6 drop. Removing PARS yields $e^r_i = E_r(v_i),$7, $e^r_i = E_r(v_i),$8, $e^r_i = E_r(v_i),$9, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$00, an approximately $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$01 drop. Removing MSRA gives $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$02, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$03, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$04, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$05. Removing CORP gives $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$06, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$07, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$08, $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$09 (Xing et al., 21 Aug 2025). These results support the paper’s interpretation that parallel unordered tokenization, dynamic ordering, preference-aligned RL, multi-step reward shaping, and self-reflective pruning are complementary rather than redundant.

The sensitivity analysis reports that codebook size $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$10 is stable from 200 to 500 and slightly worse at 100; the time-decay coefficient $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$11 performs best around 0.8; the future-horizon parameter $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$12 performs best around 3; and the CORP threshold $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$13 performs best around 0.06 (Xing et al., 21 Aug 2025). Reasoning-depth scaling, achieved by varying $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$14 from 2 to 12, shows monotonic increases in both $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$15 and $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$16, indicating that the model can exploit deeper reasoning chains rather than merely tolerate them.

The online A/B test on a major Southeast Asian e-commerce advertising platform ran from July 18 to July 22, 2025, with 15% of users in each group (Xing et al., 21 Aug 2025). Against a SASRec-based control, the REG4Rec treatment improves Advertising Revenue by $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$17, CTR by $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$18, and GMV by $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$19, with all gains reported as statistically significant at $$\mathcal{S}_u = (i_1, i_2, \ldots, i_T), \quad i_t \in \mathcal{I},$$20.

6. Limitations, related directions, and terminological scope

The paper identifies several limitations. Reward design in PARS and MSRA relies on handcrafted components such as step-hit, category-hit, JS consistency, and path robustness, and these may need retuning across domains (Xing et al., 21 Aug 2025). MPQ depends on high-quality multimodal embeddings, so weak feature extraction would degrade token semantics. The full system also introduces nontrivial training and engineering overhead through pretraining, RL post-training, the category predictor, CORP, LADQ, and large-scale infrastructure. Cold-start behavior, cross-domain transfer, and user-facing explainability are described as insufficiently explored.

The authors list several future directions: self-adaptive reward mechanisms such as inverse RL, explainability and causal reasoning over multi-path traces, richer multimodal and cross-domain extensions, better support for cold-start scenarios, and human feedback in the loop (Xing et al., 21 Aug 2025). These directions are consistent with adjacent work. ReRec studies LLM-based recommendation assistants under reinforcement fine-tuning, using dual-graph enhanced reward shaping, reasoning-aware advantage estimation, and an online curriculum scheduler, while emphasizing preservation of instruction-following and general knowledge (Huang et al., 9 Apr 2026). Re4 addresses a different problem—multi-interest recommendation in the matching phase—through backward-flow regularization via Re-contrast, Re-attend, and Re-construct (Zhang et al., 2022). A plausible implication is that REG4Rec occupies a middle position between these lines: more industrial and token-centric than LLM-only recommendation assistants, but more explicitly reasoning-oriented than conventional multi-interest recommenders.

The name itself also has a broader terminological footprint. In a separate visual grounding context, REG4Rec is used descriptively for a paradigm in which referring-expression generation is explicitly optimized against a downstream referring-expression comprehension module, as instantiated by the Interactive REG model (Ye et al., 2023). That usage concerns interactive language generation for visual localization rather than sequential recommendation. Within recommendation research, however, the term denotes the reasoning-enhanced generative model centered on MPQ, CRSS, PARS, MSRA, CORP, and LADQ (Xing et al., 21 Aug 2025).

In summary, REG4Rec recasts large-scale next-item prediction as reasoning over multiple unordered semantic tokens, then constrains that reasoning with recommendation-specific rewards and inference-time consistency checks. Its contribution lies not only in replacing singular item tokenization with a richer semantic substrate, but in treating recommendation paths themselves as trainable, evaluable, and prunable objects (Xing et al., 21 Aug 2025).