🤖 machine learning

DiffGRM: Diffusion-based Generative Recommendation Model

This paper introduces DiffGRM, a diffusion-based generative recommendation model that overcomes the limitations of autoregressive methods by employing a masked discrete diffusion framework with specialized tokenization, training, and inference strategies to enable bidirectional context and balanced supervision for improved recommendation accuracy.

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

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a massive, endless library where every single book has a secret code instead of a title. To find the next book you'll love, a computer has to guess this code, one digit at a time. This is the world of Generative Recommendation, a branch of artificial intelligence that tries to predict what you want to see next—whether it's a video, a song, or a product—by "writing" the secret code for that item from scratch.

For a long time, these computers worked like a strict teacher reading a story aloud from left to right. They guessed the first number of the code, then the second, then the third, never looking back. But this method has a flaw: it treats every part of the code as if it's equally important and equally hard to guess. In reality, some parts of the code are easy clues (like "it's a shoe"), while others are tricky details (like "it's a size 9 left shoe"). The old method often got stuck on the hard parts because it couldn't look ahead or check its work, and it wasted time over-practicing the easy parts. The question researchers have been asking is: Can we build a smarter system that looks at the whole picture at once, figures out which parts are tricky, and fills in the blanks more efficiently?

Enter DiffGRM, a new approach from researchers at Kuaishou that swaps the old "left-to-right" reading style for something more like a group of detectives solving a mystery together. Instead of guessing one number at a time, DiffGRM uses a technique called discrete diffusion. Imagine a game where you start with a page of text that has been completely covered in black ink (masks). Your job is to figure out what the original words were. The computer starts by guessing what's under the ink everywhere at once. It doesn't just guess blindly; it looks at the whole sentence to see which words are most likely to be correct.

The paper introduces three clever tricks to make this work better than the old methods. First, they changed how the "secret codes" are made. Instead of a chain where one number depends on the one before it, they use a method called Parallel Semantic Encoding. Think of this like giving each detective a separate, independent clue card. This way, no single clue is stuck waiting for another, and the computer can guess all the numbers at the same time without getting confused by a bad guess earlier in the line.

Second, the team realized that not all guesses are created equal. Some parts of the code are easy to figure out, while others are a real puzzle. The old method treated them all the same, wasting energy on the easy stuff. DiffGRM uses a strategy called On-policy Coherent Noising. Imagine a teacher who knows exactly which students are struggling. Instead of asking the whole class the same easy questions, this teacher focuses the practice time specifically on the students who are having the hardest time. The computer does the same: it identifies the "hard" digits in the code and focuses its learning power there, ignoring the easy parts that it already knows.

Finally, when it's time to make the final recommendation, the computer needs to give you a list of options, not just one. The old way would pick the single best guess and stop. DiffGRM uses Confidence-guided Parallel Denoising. This is like a team of detectives who all shout out their best guesses at once. The team leader then picks the most confident guesses first, fills those in, and then uses that new information to solve the rest of the puzzle. This allows the system to generate a diverse list of top recommendations quickly and accurately.

The researchers tested this new system on real-world data from Amazon, looking at categories like Sports, Beauty, and Toys. They found that DiffGRM was significantly better than the previous best methods. In their tests, it improved the accuracy of recommendations by between 6.9% and 15.5% compared to the strongest existing models. The paper suggests that by letting the computer look at the whole code at once and focusing its energy on the difficult parts, we can build recommendation systems that understand us better and give us more relevant choices. It's not just a small tweak; it's a fundamental shift in how we teach computers to guess what we want next.

Technical Summary: DiffGRM: Diffusion-based Generative Recommendation Model

1. Problem Statement

Generative Recommendation (GR) has emerged as a paradigm where items are represented by semantic IDs (SIDs)—fixed-length sequences of $n$ digits—and predicted by autoregressively generating these digits. However, the authors identify two fundamental structural mismatches between SID representations and standard Autoregressive Models (ARMs):

Intra-item Consistency vs. Causal Constraints: An item's SID is a cohesive unit where all $n$ digits jointly specify a single item. ARMs, however, enforce a strict left-to-right causal dependency. This prevents the model from utilizing bidirectional context or cross-digit mutual verification during training, collapsing supervision into a single causal path and hindering the exploitation of intra-item consistency.
Inter-digit Heterogeneity vs. Uniform Supervision: SID digits encode different semantic granularities (e.g., Category, Brand, Type, Size) and exhibit varying levels of predictability ("easy" vs. "hard" digits). Standard next-token prediction objectives assign equal weight to all digits. This leads to an imbalance where easy digits are overtrained while hard digits receive insufficient supervision, degrading overall generation accuracy.

The paper argues that these limitations necessitate a shift from autoregressive decoding to a framework capable of bidirectional context and parallel generation.

2. Methodology: DiffGRM

The authors propose DiffGRM, a diffusion-based generative recommendation model that replaces the autoregressive decoder with a Masked Discrete Diffusion Model (MDM). The framework consists of three tailored components:

A. Tokenization: Parallel Semantic Encoding (PSE)

To address the coupling issues inherent in standard tokenizers (like Residual Quantization), DiffGRM employs Parallel Semantic Encoding (PSE).

Mechanism: It uses an Orthogonal Procrustes Quantization (OPQ) scheme. A continuous item embedding is rotated and partitioned into $n$ independent sub-vectors. Each sub-vector is quantized independently against its own codebook.
Benefit: This decouples the digits, removing residual sequential dependencies. It balances information across digits and enables fully parallel prediction, aligning with the bidirectional nature of diffusion models.

B. Training: On-policy Coherent Noising (OCN)

Standard random masking in diffusion models is inefficient for recommendation due to the combinatorial explosion of masking patterns and the sparsity of long-tail items. DiffGRM introduces On-policy Coherent Noising (OCN).

Mechanism: Instead of random masking, OCN uses the current model's predictions to estimate the difficulty (uncertainty) of each digit. It constructs a nested set of "views" (training samples) where the most uncertain (hardest) digits are masked first.
Strategy: The masking schedule progresses from light to heavy corruption, prioritizing high-value signals (hard digits) while keeping the visible context coherent. This focuses the training budget on the most informative supervision signals rather than scattering effort across easy digits or random patterns.

C. Inference: Confidence-guided Parallel Denoising (CPD)

Recommendation requires a diverse set of Top- $K$ candidates, unlike the single-output nature of standard text generation. DiffGRM proposes Confidence-guided Parallel Denoising (CPD).

Mechanism: Starting from a fully masked input, the model performs a global parallel beam search. At each denoising step, it identifies the most confident digit-token pairs across all active beams and fills them, while remasking the rest.
Strategy: This approach fills higher-confidence digits first and explores diverse paths for the remaining digits, generating a diverse and accurate Top- $K$ set of SID candidates without the sequential bottleneck of autoregressive beam search.

3. Key Contributions

The paper claims the following contributions:

First Diffusion-based GR Framework: DiffGRM is the first generative recommendation model to replace the autoregressive decoder with a masked discrete diffusion model over SID digits, effectively removing left-to-right constraints and leveraging bidirectional cross-digit context.
Novel Training and Inference Strategies:
- OCN: A training strategy that curbs the combinatorial explosion of supervision signals by coherently masking uncertainty-ranked hard digits, significantly improving sample efficiency.
- CPD: A confidence-guided global parallel beam search for inference that yields diverse Top- $K$ SID candidates, tailored specifically to the requirements of recommendation tasks.
State-of-the-Art Performance: The model achieves superior results across multiple datasets, demonstrating the effectiveness of aligning the generation mechanism with the structural properties of semantic IDs.

4. Experimental Results

The authors evaluated DiffGRM on three Amazon Reviews datasets (Sports, Beauty, Toys) against strong baselines, including discriminative models (e.g., SASRec, BERT4Rec) and generative models (e.g., TIGER, RPG).

Performance: DiffGRM achieved the best results on 11 out of 12 metrics (Recall@5, Recall@10, NDCG@5, NDCG@10).
Improvements: Compared to the strongest baseline, DiffGRM improved NDCG@10 by 6.9% to 15.5% across the datasets. Specifically, it saw a 15.53% improvement on Sports, 8.19% on Beauty, and 6.94% on Toys.
Efficiency: Ablation studies confirmed that OCN improves sample efficiency, achieving better performance with fewer effective sample passes (ESP) compared to random or fixed-path masking. The complexity analysis indicates that while DiffGRM adds a factor of $n$ to the decoder term, the overall inference cost remains comparable to ARMs due to the dominance of the encoder computation and the small size of $n$ .

5. Significance and Claims

The paper positions DiffGRM as a solution to the structural mismatch between item representation (SIDs) and generation mechanisms (ARMs). By adopting a diffusion-based approach, the model successfully:

Exploits bidirectional intra-item semantics that ARMs cannot access.
Addresses inter-digit heterogeneity by dynamically allocating supervision to difficult digits via OCN.
Provides diverse Top- $K$ recommendations through parallel denoising, avoiding the error propagation inherent in sequential generation.

The authors conclude that DiffGRM reconciles cross-digit semantics with parallel generation, offering a robust and accurate alternative to autoregressive frameworks for generative recommendation. They note that future work will focus on further investigating inference efficiency and scalability.