PaperlayerView exact revision on arXiv

Source-linked AI summary

DiffGRM: Diffusion-based Generative Recommendation Model

Zhao Liu, Yichen Zhu, Yiqing Yang, Xiao Lv, Guoping Tang, Rui Huang, Qiang Luo, Ruiming Tang, Kun Gai, Guorui Zhou

arXiv:2510.21805v2cs.IRcs.AIcs.LG

TL;DR

Autoregressive recommendation models face a mismatch with SID structure: digits jointly identify items but differ in difficulty, while causal decoding and equal token weighting limit supervision. DiffGRM uses masked discrete diffusion with tailored tokenization, training, and inference. It improves NDCG@10 by 6.9%–15.5% over strong generative and discriminative baselines across multiple datasets.

  • Problem

    SID digits jointly specify an item but differ in semantic granularity and predictability, while autoregressive decoding limits cross-digit context and equal next-token weighting allocates supervision uniformly.

    Source: §1 · pp. 1–2

  • Method

    DiffGRM replaces autoregressive decoding with masked discrete diffusion and tailors tokenization, training, and inference to SID digits.

    Source: §1 · p. 2

  • Results

    DiffGRM improves NDCG@10 by 6.9%–15.5% over strong generative and discriminative recommendation baselines across multiple datasets.

    Source: §1 · p. 2

  • Takeaways & Limitations

    DiffGRM reconciles cross-digit semantics with parallel generation and produces robust Top-K recommendations.

    Source: §6 · p. 8

Abstract

from arXiv · show

Generative recommendation (GR) is an emerging paradigm that represents each item via a tokenizer as an n-digit semantic ID (SID) and predicts the next item by autoregressively generating its SID conditioned on the user's history. However, two structural properties of SIDs make ARMs ill-suited. First, intra-item consistency: the n digits jointly specify one item, yet the left-to-right causality trains each digit only under its prefix and blocks bidirectional cross-digit evidence, collapsing supervision to a single causal path. Second, inter-digit heterogeneity: digits differ in semantic granularity and predictability, while the uniform next-token objective assigns equal weight to all digits, overtraining easy digits and undertraining hard digits. To address these two issues, we propose DiffGRM, a diffusion-based GR model that replaces the autoregressive decoder with a masked discrete diffusion model (MDM), thereby enabling bidirectional context and any-order parallel generation of SID digits for recommendation. Specifically, we tailor DiffGRM in three aspects: (1) tokenization with Parallel Semantic Encoding (PSE) to decouple digits and balance per-digit information; (2) training with On-policy Coherent Noising (OCN) that prioritizes uncertain digits via coherent masking to concentrate supervision on high-value signals; and (3) inference with Confidence-guided Parallel Denoising (CPD) that fills higher-confidence digits first and generates diverse Top-K candidates. Experiments show consistent gains over strong generative and discriminative recommendation baselines on multiple datasets, improving NDCG@10 by 6.9%-15.5%. Code is available at https://github.com/liuzhao09/DiffGRM.

1 Introduction

Generative recommendation represents items as semantic-ID sequences and generates the next item from a user’s history. DiffGRM replaces autoregressive decoding with masked diffusion and adapts tokenization, training, and inference to SID structure.

1 Introduction

  • SID digits jointly specify one item but differ in semantic granularity and predictability, creating a mismatch with left-to-right autoregressive decoding and equal next-token supervision.Causal attention restricts each digit to its prefix context, while equal weighting undertrains hard digits and overtrains easy ones.Source: §1 · pp. 1–2
  • DiffGRM uses Parallel Semantic Encoding to decouple digits and balance per-digit information, On-policy Coherent Noising to prioritize hard digits, and Confidence-guided Parallel Denoising to generate diverse Top-K candidates.The method replaces autoregressive decoding with masked diffusion, using bidirectional context and parallel SID-digit generation.Source: §1 · p. 2
  • DiffGRM reports 6.9%–15.5% higher NDCG@10 than strong generative and discriminative recommendation baselines.The paper presents DiffGRM as a diffusion-based generative recommendation framework using a masked discrete diffusion model over SID digits.Source: §1 · p. 2

2 Background

The background formulates next-item recommendation as conditional generation of an item’s SID from a user’s SID history. It then describes masked diffusion, which corrupts SID digits and predicts masked digits in parallel.

2.1 SID Generation for Recommendation

  • Given a user’s interaction history, the model learns a conditional generator for the next item’s SID, instantiated as a masked-diffusion decoder trained with digit-wise supervision on masked digits.Items are represented by content-derived embeddings discretized into SIDs, and user history is converted into a sequence of SIDs.Source: §2.1 · pp. 2–3

2.2 Masked Diffusion for Parallel Token Prediction

  • Masked diffusion replaces selected SID digits with [MASK] under a time-dependent schedule, then predicts all masked digits in parallel from the corrupted sequence.The masked-digit cross-entropy objective removes causal constraints and provides richer supervision signals through discrete masking during training.Source: §2.2 · p. 3

3 Method

DiffGRM uses a history encoder and a bidirectional masked-diffusion decoder to predict SID digits in parallel. PSE decouples the digits, OCN allocates masking to uncertain digits, and CPD generates Top-K candidates by confidence-guided denoising.

3.1 Overview

  • The encoder summarizes user history while the MD-Decoder uses non-causal attention to predict masked SID digits in parallel.Encoder-side keys and values are cached and reused across views and denoising steps.Source: §3.1 · p. 3

3.2 Parallel Semantic Encoding

  • PSE independently assigns partitioned representation subvectors to per-digit codebooks, removing residual sequential dependence and enabling parallel digit prediction.The method uses an OPQ-based partition-and-quantize scheme.Source: §3.2 · p. 4

3.3 On-policy Coherent Noising

  • OCN constructs a small set of nested views by masking uncertainty-ranked hard digits, concentrating supervision on higher-value signals while avoiding combinatorial masking patterns.Views progress from light to heavy corruption, with the hardest digits masked first.Source: §3.3 · p. 4

3.4 Confidence-guided Parallel Denoising

  • CPD performs global beam-search denoising, filling SID digits in descending model confidence to produce Top-K candidates.It departs from left-to-right expansion; generated sequences are deduplicated and ranked by final scores.Source: §3.4 · p. 5

3.5.1 Why not RQ.

  • PSE is chosen over residual quantization because residual levels can imbalance digit information and hierarchical dependence conflicts with parallel any-order prediction.PSE instead factorizes representations into independent subspaces.Source: §3.5.1 · p. 5

3.5.2 Complexity of Training.

  • Training complexity is encoder-dominated for long histories and a small number of views, making ARM and DiffGRM training costs essentially the same overall.DiffGRM reuses one encoder run across its views.Source: §3.5.2 · p. 5

3.5.3 Complexity of Inference.

  • DiffGRM and ARM have similar end-to-end inference cost, with a modest extra decoder factor that increases compute but typically leaves wall-clock latency close to ARM.The relative encoder and decoder cost depends on history length compared with nB_act.Source: §3.5.3 · p. 5

4 Experiments

Experiments compare DiffGRM with recommendation baselines on three Amazon Reviews datasets, then evaluate OCN, component ablations, and inference choices. DiffGRM leads on most overall metrics, while analyses examine sample efficiency and the effects of its design choices.

4.1 Experimental Setup

  • Evaluation uses Sports, Beauty, and Toys Amazon Reviews sequences with leave-last-out testing, reporting Recall@K and NDCG@K for K=5 and 10.The final item is held out for testing, the second-to-last for validation, and preceding interactions for training; checkpoints are selected on validation performance.Source: §4.1 · p. 6

4.2 Overall Performance (RQ1)

  • DiffGRM ranks first on 11 of 12 metrics; versus the strongest baseline, NDCG@10 improves 15.53% on Sports, 8.19% on Beauty, and 6.94% on Toys.Recall@10 improves 10.00% on Sports and 8.28% on Beauty, but is 4.03% lower on Toys; DiffGRM still has higher NDCG at K=5 and K=10.Source: §4.2 · pp. 6–7

4.3 Effectiveness of OCN (RQ2)

  • OCN achieves better results at the same or lower effective sample passes by selecting uncertain positions along coherent paths for training.Effective sample passes (ESP) counts best-epoch training views per sample; increasing coherent paths raises performance but also raises ESP.Source: §4.3 · p. 7, Figure 3 · p. 7

4.4 Ablation Study (RQ3)

  • PSE variants show that residual-quantized digit dependencies conflict with masked diffusion, while random tokens discard semantic structure and further hurt performance.These ablations support PSE's role in making tokenization compatible with masked diffusion.Source: §4.4 · pp. 7–8
  • Removing OCN's coherent masking and on-policy selection disperses supervision and leaves infrequent long-tail samples insufficiently trained; removing only on-policy selection performs better but remains below DiffGRM.Replacing CPD with random fixed-order beam search also reduces performance across all three datasets.Source: §4.4 · p. 8

4.5.1 OCN Strategy Analysis.

  • Least-confidence digit selection outperforms most-confidence selection under the same refresh setting, while static ordering outperforms re-estimating uncertainty after every denoising step.The largest degradation occurs for M-R when CPD is replaced; the paper attributes the drop to hard digits being undertrained when the easy-first order changes.Source: §4.5.1 · p. 8, Table 5 · p. 8

4.5.2 CPD Beam-Size Analysis.

  • Increasing CPD beam size across 32, 64, 128, and 256 improves NDCG@10 by mitigating local optima.Figure 4 reports DiffGRM NDCG@10 as a function of beam size.Source: §4.5.2 · p. 8, Figure 4 · p. 8

4.5.3 Hidden Dimension Analysis.

  • A model width of 256 suits Sports and Beauty, whereas Toys benefits from width 1024; these settings are used in the main experiments.Source: §4.5.3 · p. 8

5 Related Work

Prior work develops autoregressive generative recommenders and discrete diffusion models, but the paper distinguishes GR's structured SID and Top-K needs from diffusion language models' free-form, single-output text focus.

Generative Recommendation Models

  • Autoregressive recommenders generate item SIDs token by token, while representative work explores quantization, context-aware tokenization, parallel prediction, and language-modeling techniques.The passage names TIGER, HSTU, GenNewsRec, MTGRec, ETEGRec, ActionPiece, and RPG.Source: §5 · p. 8
  • The paper identifies left-to-right decoding mismatch and imbalance as motivation for diffusion-based recommendation alternatives.It notes that generative recommenders unify representation and prediction and support open-vocabulary recommendation.Source: §5 · p. 8

Discrete Diffusion Language Models

  • Discrete diffusion language models have advanced through CTMC formulations, score-based objectives, and reverse-sampling strategies, approaching autoregressive performance on some tasks.The passage contrasts those models' free-form, single-output text target with GR's structured n-digit SIDs and Top-K output.Source: §5 · p. 8

6 Conclusion

The conclusion presents DiffGRM as a discrete masked-diffusion recommender combining tokenization, training, and inference components, and reports gains over strong baselines.

6 Conclusion

  • DiffGRM improves NDCG@10 by 6.9%–15.5% over strong baselines and produces robust Top-K recommendations.The conclusion says the framework reconciles cross-digit semantics with parallel generation.Source: §6 · p. 8
  • The framework combines OPQ-based PSE to balance information and remove cross-digit dependence, OCN to emphasize uncertain digits, and CPD to generate diverse Top-K candidates.The conclusion characterizes CPD as confidence-guided parallel beam search.Source: §6 · p. 8
  • The authors identify inference efficiency and scalability as topics for future investigation.Source: §6 · p. 8

A Notations

  • Table 6 is presented as a summary of the paper's notations and explanations.Source: A Notations · p. 10

B Implementation Details

The implementation combines prior and reproduced baselines with a Transformer-encoder/MD-Decoder DiffGRM setup. Training uses AdamW and validation-based early stopping, followed by masked-diffusion decoding and beam search.

B Implementation Details

  • Baseline results come from prior work on the same Amazon splits [20] [21] [24] [48], while remaining methods are reproduced with public implementations and tuned to their original papers.Implementations primarily use RecBole [70] or authors’ official code.Source: B Implementation Details · p. 10
  • DiffGRM uses a PyTorch implementation, FAISS OPQ-quantized text-derived item embeddings as SIDs, and a Transformer encoder with an MD-Decoder.Dataset-specific hyperparameters are summarized in Table 7.Source: B Implementation Details · p. 10, Table 7 · p. 11
  • Training uses AdamW for at most 100 epochs, 10,000-step linear warmup, dropout 0.1, label smoothing, and early stopping with patience 15.Early stopping uses validation score 0.8 NDCG@10 + 0.2 Recall@10, and the best checkpoint is tested.Source: B Implementation Details · p. 11
  • SID candidates are generated by stepwise masked-diffusion denoising followed by beam search.The sentence-T5-base semantic encoder is used for fairness with prior work, with bge-large-en-v1.5 and gte-large-en-v1.5 variants also evaluated.Source: B Implementation Details · p. 11

C Expressive Ability Analysis

The analysis reports that larger semantic encoders improve DiffGRM's NDCG@10, and contrasts autoregressive supervision with masked diffusion's larger target-context signal space.

C Expressive Ability Analysis

  • NDCG@10 rises with semantic encoder size and capacity across three benchmarks, with DiffGRM benefiting the most.The analysis fixes input fields to Title, Price, Brand, Category, and Description, and holds other configurations fixed.Source: C Expressive Ability Analysis · p. 11

D Supervision Signals for ARM and MDM

  • A teacher-forced autoregressive sample provides exactly n supervision signals, one for each target digit under its strict-prefix context.The passage says one sample covers all n digit-context pairs, so the minimum sample count is one.Source: D Supervision Signals for ARM and MDM · p. 11
  • Masked diffusion defines signals for each target digit in a nonempty masked set, yielding n2^(n-1) distinct target-context pairs and requiring 2^n−1 configurations for full coverage.The paper says it prioritizes high-value masked sets in practice rather than enumerating all patterns.Source: D Supervision Signals for ARM and MDM · p. 11

E ARM Sample Expansion

This comparison tests whether repeated identical training instances benefit an autoregressive recommender. ARM duplication adds no new supervision, unlike masked diffusion views that can expose additional masked-digit signals.

E ARM Sample Expansion

  • Masked diffusion views expose additional masked-digit supervision, whereas duplicating an autoregressive training instance adds no new supervision signals.The ARM supervises all n SID digits in one pass; the diffusion model supervises only masked digits per view.Source: E ARM Sample Expansion · p. 11
  • On Toys, the ARM comparison uses the GRID framework [26] with the main system's encoder–decoder backbone and RQ-KMeans quantization.The 4x setting duplicates each original instance four times without changing targets or token order; other hyperparameters and early stopping remain unchanged.Source: E ARM Sample Expansion · p. 11
  • Table 11 reports negligible performance differences between original and fourfold-duplicated ARM training instances.This supports the interpretation that the MDM gains come from additional supervision signals rather than more copies of identical supervision.Source: E ARM Sample Expansion · p. 11

F Sliding Window Data Augmentation

Sliding-window augmentation expands user histories into contiguous subsequences for next-item training. The reported results show consistent gains across datasets as the training-sample count increases.

F Sliding Window Data Augmentation

  • Sliding-window augmentation expands each user interaction sequence into all contiguous subsequences, each targeting the next item through the masked diffusion decoder.The added item–SID correspondences and richer contexts are described as helping generalizable pattern learning, mitigating overfitting, and improving robustness to sparsity or noise.Source: F Sliding Window Data Augmentation · p. 11
  • Figure 5 contrasts learnable supervision signals for the autoregressive model and masked diffusion model in a three-layer codebook example.The example uses a three-layer codebook and semantic ID; its caption identifies the two model paradigms being compared.Source: Figure 5 · p. 12, F Sliding Window Data Augmentation · p. 12
  • Results show consistent gains across datasets as the training-sample count increases.Training samples count sequence instances actually used for optimization after splitting or augmentation; Table 12 reports the augmentation effects on sample count and performance.Source: F Sliding Window Data Augmentation · pp. 11–12, Table 12 · p. 12

G Hidden Dimension Analysis

The hidden-dimension analysis finds different accuracy–efficiency tradeoffs across datasets and uses dataset-specific dimensions in the main experiments.

G Hidden Dimension Analysis

  • Sports and Beauty show a performance knee near d_m=256, while Toys continues improving through d_m=1024.The analysis varies d_m across 64, 128, 256, 512, and 1024 with other settings fixed; larger dimensions also increase parameter, memory, and compute costs, while best epoch decreases.Source: G Hidden Dimension Analysis · p. 12
  • The main experiments use d_m=256 for Sports and Beauty and d_m=1024 for Toys to balance accuracy and efficiency.The selected settings follow the dataset-specific performance trends in the hidden-dimension analysis.Source: G Hidden Dimension Analysis · p. 12

Related papers

Loading 2510.21805v2…