Masked Diffusion Challenges Left-to-Right Decoding in Generative Recommendation (Recsys'26)

Background: Limitations of Autoregressive Generative Recommendation

Generative recommendation (GR) models items as discrete semantic IDs (SIDs) and frames recommendation as sequence generation. Existing GR methods largely adopt autoregressive decoding from language models, generating tokens left-to-right. This paradigm mismatches recommendation in two ways: (1) the goal is whether the final SID corresponds to the target item, not the generation order, yet autoregressive models only see left prefixes and cannot enforce global consistency across semantic dimensions; (2) user interests are heterogeneous — different users attend to item attributes in different orders — but fixed decoding paths assume a universal attribute order.

Two mainstream GR approaches exist: residual-codebook autoregressive decoding (hierarchical SIDs, left-to-right) and parallel-codebook single-step decoding (simultaneous prediction of all positions). The former suffers from limited global consistency and rigid order; the latter ignores fine-grained inter-attribute constraints and still cannot adapt to heterogeneous interest structures. An ideal GR decoder should be order-agnostic, support multi-step refinement, and enable parallel generation.

MDGR: Masked Diffusion Generative Recommendation Framework

MDGR introduces masked diffusion to generative recommendation, modeling SID generation as a multi-step mask-denoising process. It redesigns three layers: codebook, training, and inference.

2.1 Parallel Codebooks

Residual quantization imposes hierarchical dependencies that conflict with bidirectional modeling. MDGR builds parallel codebooks using Optimized Product Quantization (OPQ). Given a pretrained encoder's item content representation v, it projects into M subspaces via linear layers W_m. Each subspace maintains a codebook C_m assigning the subspace vector to its nearest codeword. The parallel SID s = [c_1, ..., c_M] collects indices from all subspaces. This structure imposes no hierarchical dependencies, naturally supporting bidirectional modeling and independent masking.

2.2 Offline Training: Dynamic Noise Control Along Two Dimensions

MDGR uses an encoder-decoder architecture where the decoder replaces causal self-attention with bidirectional attention for discrete diffusion. The core design dynamically controls training noise along time and sample dimensions.

Time Dimension: Global Curriculum Noise Scheduling

Early in training, model capacity is limited; too many masked positions make reconstruction too hard. Inspired by curriculum learning, MDGR gradually increases the mask ratio. Let T be total training steps, t current step. Normalized progress τ = t/T. A smoothed cosine schedule with power transformation yields a stretched difficulty d(τ) = ((cos(πτ) + 1)/2)^γ, where γ controls the speed of d decreasing from 1 to 0. Given difficulty d, MDGR constructs a distribution over possible mask counts k ∈ {1,...,M}. Two monotonic scoring functions are defined: f_1(k) decreasing with k (prefers fewer masks), f_2(k) increasing with k (prefers more masks). Interpolation with d gives s(k; d) = d·f_1(k) + (1-d)·f_2(k). Normalizing yields sampling distribution p(k|d). Early training (large d) favors small k; later training shifts probability to large k, realizing easy-to-hard curriculum masking. To inform the decoder of current noise strength, MDGR adds a learnable difficulty-aware embedding: the sampled mask count k indexes an embedding table, and the resulting vector is added to every input token embedding.

Sample Dimension: History-Aware Mask Position Allocation

After deciding how many positions to mask, MDGR decides which positions. The intuition: tokens in the target SID that appear rarely in the user's history are harder to predict and should receive more masking supervision. For user u and target item i, count occurrences of each target SID token in the user's historical sequence: cnt_m = ∑_{j∈H_u} 1[s_j^m = s_i^m]. Take inverse and normalize to obtain a position sampling distribution p(m) ∝ 1/(cnt_m + ε). Sample k mask positions from this distribution and replace corresponding tokens with [MASK]. Thus, more masking budget goes to rare, challenging semantic dimensions, concentrating supervision on difficult semantics. The training objective reconstructs original tokens at masked positions: L = -∑_{m∈M_mask} log p(s_i^m | context).

2.3 Online Inference: Warmup-Based Two-Stage Parallel Decoding with Beam Search

Inference starts from an all- [MASK] SID and iteratively denoises conditioned on user history. Since recommendation requires a top-B candidate set, MDGR designs a parallel beam search decoding flow where each step: (1) decides how many positions to update, (2) selects specific positions by confidence, (3) performs beam search expansion on selected positions.

Anchor-Then-Parallel Two-Stage Decoding

At the fully noisy start, model confidence is low across positions; parallel decoding too early risks committing errors. MDGR sets a warmup step count W. At step t, the number of positions updated is n_t = 1 for t ≤ W (anchor stage) and n_t = P for t > W (parallel stage). During the first W steps, only the single highest-confidence position is decoded per step, freely choosing any position (not left-to-right) to lock the strongest global semantic anchor first. After warmup, the model switches to parallel denoising of P highest-confidence positions per step, drastically reducing decoding steps.

Confidence-Guided Position Selection and Beam Search Expansion

At each step, the model outputs a prediction distribution over the codebook for every unfilled position. The maximum probability at each position serves as its confidence. The top n_t positions by confidence are selected for decoding. On these positions, each beam retains the top-B candidate tokens and jointly expands; beams are ranked by cumulative log-probability, keeping the global top-B paths for the next step. After decoding completes, generated SIDs are mapped back to concrete items via the offline-built codebook index for top-K recommendation.

Experiments

On Amazon Electronics, Books public datasets, and an industrial dataset with 1B+ interactions, MDGR achieves the best results across all datasets and metrics, outperforming existing generative methods by 3.44% to 6.56% relative improvement. Gains come mainly from redesigned training supervision and decoding, not codebook changes, validating that masked diffusion enforces stronger cross-semantic global consistency than autoregressive decoding. Compared to contemporary diffusion-based GR methods DiffGRM and LLaDA-Rec — which largely reuse generic masked diffusion architectures without adapting to SID structure and interest heterogeneity — MDGR leads comprehensively due to its recommendation-specific designs at both training and inference.

Efficiency-Effectiveness Trade-off

The paper analyzes two key inference hyperparameters: warmup steps W and parallel decode positions per step P. Without parallel acceleration, MDGR matches baseline decoding steps (e.g., TIGER) while already leading in Recall@10, showing the framework itself brings gains. Enabling parallel decoding: larger P increases QPS; at the extreme single-step full decoding, QPS rises 85.74% but beam search covers far fewer candidate paths, causing >5% performance drop. Adding warmup steps recovers effectiveness. The final configuration W=2, P=3 achieves optimal effectiveness while delivering 13.66% QPS improvement, adopted as default.

Online Results

MDGR was deployed on Alibaba's e-commerce advertising platform in an online A/B test against the previously serving REG4Rec model. Results: ad revenue +1.20%, CTR +2.36%, GMV +3.69%, CVR +1.03%. This demonstrates that masked diffusion generative recommendation not only improves offline metrics but also yields stable business gains in large-scale industrial environments.

Paper Information

Title: Masked Diffusion Generative Recommendation<br> Authors: Lingyu Mu*, Hao Deng*, Haibo Xing, Jinxin Hu, Yu Zhang, Xiaoyi Zeng, Jing Zhang<br> Venue: Recsys 2026<br> Link:

https://arxiv.org/abs/2601.19501