LoopGR: Parameter-Efficient Scaling for Generative Recommendation via Recurrent Depth

1 Introduction

• The central challenge is allocating limited computation between interaction-history modeling and target SID generation under tight serving constraints.Increasing model capacity and training computation can improve recommendation, but recommender systems face tighter computational and serving constraints than LLMs.Source: §1 · pp. 1–2
• Existing designs progress from separate history and target computation to shared stacks and then target-centric computation.This architectural shift exposes a tension between expressive context modeling and efficient target generation, with efficient decoding often treating context as fixed.Source: §1 · pp. 1–2, Figure 1 · p. 2
• Looped Transformers decouple computational depth from parameter size through shared layers, but recurrent steps add latency and do not inherently separate history refinement from target generation.The paper motivates a recommendation-specific recurrent design that organizes those computations separately while supporting low-budget inference.Source: §1 · p. 2
• LoopGR interleaves shared decoder cycles with lightweight context-updater cycles, while recurrent trajectory self-distillation transfers predictive distributions and evolution directions from high- to low-budget trajectories.The decoder predicts target SID representations from current context, and the updater feeds emerging target state back to refine context.Source: §1 · pp. 2–3
• LoopGR outperforms strong baselines with fewer parameters and evaluates recurrent depth, model size, computation scaling, and trajectory evolution.The introduction reports these results and analyses without specifying a quantitative comparison.Source: §1 · p. 3

Generative Recommendation

• Generative recommendation approaches either align pretrained LLMs with recommendation objectives or train recommendation-tailored architectures from scratch.Other approaches introduce latent reasoning; OneRec-V2 reallocates computation toward target SID generation and avoids repeated expensive sequence encoding.Source: §2 · p. 3
• Most existing methods retain fixed-depth Transformer backbones, making greater representational capacity typically require additional layers and parameters.The paper characterizes this parameter-centric scaling as difficult to reconcile with recommendation systems' memory and latency requirements.Source: §2 · p. 3

Looped Transformers

• Looped Transformers reuse shared layers to increase effective depth, but recommender-system recurrence remains underexplored beyond discriminative CTR prediction.LoopCTR uses a parameter-shared prefix-attention module for CTR rather than autoregressive item generation; deeper recurrence also adds sequential computation.Source: §2 · p. 3

Semantic ID-based Generative Recommendation

• Semantic-ID recommendation predicts the next item in a user's interaction history rather than assigning each item an atomic identifier.An SID consists of semantic tokens generated by an RQ-VAE and an extra anti-collision token.Source: §3 · pp. 3–4

Recurrent Depth

• LoopGR's overview alternates parameter-shared decoding and lightweight context updates to predict target SIDs along recurrent depth.The computation forms a multi-step prediction trajectory used by the training objective.Source: Figure 2 · p. 4
• A recurrent architecture applies a parameter-shared transformation repeatedly to an initial hidden representation conditioned on context.The number of recurrent steps is the loop budget, and intermediate hidden states form a latent trajectory.Source: §3 · p. 4

4 Methodology

• LoopGR combines an encoder, shared decoder, and lightweight context updater in a hierarchical recurrent process for progressively predicting target SIDs.The architecture forms a multi-step prediction trajectory that its training objective exploits under different inference budgets.Source: §4 · p. 4, §4.1 · p. 5

4.1 LoopGR Architecture

• The encoder initializes contextual memory from historical SID tokens, while nested low-level cycles refine context and high-level cycles revisit target prediction.The updater uses prior prediction states to refine historical representations between decoding cycles; its parameters are shared across low-level cycles.Source: §4.1 · p. 5
• At each autoregressive generation step, the shared decoder repeats its procedure for T_H high-level cycles using the current context and target prefix.The decoder input is formed from the BOS token and target prefix, and each high-level cycle uses its contextual memory.Source: §4.1 · p. 5

4.2 Recurrent Trajectory Self-Distillation

• Recurrent trajectory self-distillation trains a high-budget teacher, initializes a lower-budget student from it, and aligns predictive distributions and their evolution directions.This transfers the teacher's refinement process so the student can approximate it with substantially fewer loop steps.Source: §4.2 · p. 6

5.1 Experimental Setup

• The experimental setup uses Amazon Reviews 2014 categories and a Tencent production recommendation dataset, comparing LoopGR with conventional and generative recommenders.The reported LoopGR-RTSD student operates at half the recurrent budget of its full-budget counterpart.Source: §5.1 · p. 6, §5.5 · p. 10

5.2 Overall Performance

• LoopGR and LoopGR (RTSD) achieve the best result in 8 of 12 evaluation settings, while LoopGR outperforms RPG on 11 of 12 metrics and ActionPiece on 8 of 12.The industrial evaluation also reports that LoopGR consistently outperforms the generative recommendation baselines used in Tencent’s recommendation system.Source: Table 1 · p. 7, §5.2 · p. 7, Table 2 · p. 7

5.3 Architecture Comparison

• At approximately 0.5 GFLOPs, LoopGR uses 1.3M parameters versus 2.7M for Encoder-Decoder, and it achieves the highest NDCG@5 across computation levels.At similar parameter counts, LoopGR also achieves higher NDCG@5 than alternative architectures; removing low-level recurrence degrades performance, while adding updater cycles improves Recall@5 and NDCG@5.Source: §5.3 · pp. 8–9, Figure 3 · p. 8

5.4 Analysis of Loop Dynamics

• LoopGR’s recurrent states make larger early transitions followed by localized refinement, while distilled students approximate longer trajectories with fewer, increasingly front-loaded updates.The 4-loop student’s trajectory is dominated by its initial transition; the 8-loop student makes progressively smaller updates toward its terminal state.Source: §5.4 · pp. 9–10, Figure 4 · p. 9

5.5 Ablation Study

• Identity updates reduce performance by about 8% on average relative to learnable updater variants, while RTSD reduces the gap to full-budget LoopGR to 3.2%.Table 4 reports LoopGR’s maximum relative improvements over Lazy Decoder, TIGER, and BARGE as 11.47%, 7.13%, and 3.91%, respectively.Source: §5.5 · pp. 10–11, Table 3 · p. 10, Table 4 · p. 11

• LoopGR decouples computational depth from parameter growth through hierarchical recurrent computation.Source: §6 · p. 11
• Recurrently reusing a decoder and context updater increases effective depth while dynamically refining historical context during SID generation.Source: §6 · p. 11
• LoopGR consistently outperforms strong baselines and surpasses representative Encoder-Decoder architectures with up to 52% fewer parameters.Source: §6 · p. 11

• Beauty contains 22363 users, 12101 items, and 198502 interactions, with an average interaction length of 8.88.Source: §A.1 · p. 16
• Sports and Outdoors contains 35598 users, 18357 items, and 296337 interactions, with an average interaction length of 8.32.Source: §A.1 · p. 16
• CDs and Vinyl is the largest dataset, with 75258 users, 64443 items, over one million interactions, and an average length of 14.58.Table 5 defines average length as the average number of interactions per user.Source: §A.1 · p. 16, Table 5 · p. 16

• Reducing the teacher from 16 cycles to 8 and 4 lowers computation from 17.408 GFLOPs to 8.415 and 3.918 GFLOPs, respectively.Source: §A.2 · p. 16
• The 4-cycle student surpasses the 16-cycle teacher on all four Sports metrics and remains competitive on Beauty; the 8-cycle student leads Beauty Recall@10 and NDCG@10.Source: Table 6 · p. 16, §A.2 · p. 16
• Inference latency falls from 211.08 ms to 104.01 ms and 49.82 ms at batch size 1, providing approximately 2.0× and 4.2× speedups.Both shorter trajectories retain comparable recommendation accuracy.Source: Table 6 · p. 16, §A.2 · p. 16

• Figure 5 compares a 16-layer Encoder-Decoder, a 16-loop LoopGR teacher, and distilled students with 8 and 4 recurrent cycles across three SID positions.Source: §A.3 · p. 16
• Across representative samples, LoopGR exhibits iterative refinement throughout recurrent computation.Source: §A.3 · p. 16
• The 8- and 4-cycle distilled students reach similar terminal regions through substantially shorter trajectories.Source: §A.3 · p. 16

Token-level Updaters

• Token-level updaters refine individual query-conditioned SID tokens; linear attention aggregates information across the history using a kernel-based mechanism.Its prediction-aware query enters only the query branch, while keys and values come from conditioned context representations.Source: §A.4 · p. 17
• Linear attention avoids explicit pairwise attention scores with computational complexity O(Lud^2), while token-level MLP and identity variants act independently on each SID token.The MLP applies a feed-forward network; identity leaves the input unchanged.Source: §A.4 · p. 17

Item-level Updaters

• The GRU pools each item’s SID tokens, combines projected item vectors with the prediction-aware query, and processes the history bidirectionally before broadcasting updates to all tokens.With hidden dimension d per direction, its sequence modeling and projection cost is O(nd^2).Source: §A.4 · p. 17
• Mixing preserves each item’s hierarchy-specific SID representations, while GRU summarizes them by mean pooling and shares one update across the item’s tokens.The two variants provide complementary inductive biases: GRU captures bidirectional dependencies across items, while Mixing adds no inter-item communication.Source: §A.4 · pp. 17–18
• Mixing groups each item’s SID tokens into an H × d matrix, mixes their features, applies hierarchy-specific feed-forward networks, and maps each output row back to its corresponding token.For a fixed number of mixing layers, its computational complexity is O(Lud^2).Source: §A.4 · p. 17