Luyi Ma

Tao Zhang, Kehui Yao, Luyi Ma et al.

Evaluating large language models (LLMs) as judges is increasingly critical for building scalable and trustworthy evaluation pipelines. We present ScalingEval, a large-scale benchmarking study that systematically compares 36 LLMs, including GPT, Gemini, Claude, and Llama, across multiple product categories using a consensus-driven evaluation protocol. Our multi-agent framework aggregates pattern audits and issue codes into ground-truth labels via scalable majority voting, enabling reproducible comparison of LLM evaluators without human annotation. Applied to large-scale complementary-item recommendation, the benchmark reports four key findings: (i) Anthropic Claude 3.5 Sonnet achieves the highest decision confidence; (ii) Gemini 1.5 Pro offers the best overall performance across categories; (iii) GPT-4o provides the most favorable latency-accuracy-cost tradeoff; and (iv) GPT-OSS 20B leads among open-source models. Category-level analysis shows strong consensus in structured domains (Electronics, Sports) but persistent disagreement in lifestyle categories (Clothing, Food). These results establish ScalingEval as a reproducible benchmark and evaluation protocol for LLMs as judges, with actionable guidance on scaling, reliability, and model family tradeoffs.

Xiaohan Li, Zheng Liu, Luyi Ma et al.

Recent studies on Next-basket Recommendation (NBR) have achieved much progress by leveraging Personalized Item Frequency (PIF) as one of the main features, which measures the frequency of the user's interactions with the item. However, taking the PIF as an explicit feature incurs bias towards frequent items. Items that a user purchases frequently are assigned higher weights in the PIF-based recommender system and appear more frequently in the personalized recommendation list. As a result, the system will lose the fairness and balance between items that the user frequently purchases and items that the user never purchases. We refer to this systematic bias on personalized recommendation lists as frequency bias, which narrows users' browsing scope and reduces the system utility. We adopt causal inference theory to address this issue. Considering the influence of historical purchases on users' future interests, the user and item representations can be viewed as unobserved confounders in the causal diagram. In this paper, we propose a deconfounder model named FENDER (Frequency-aware Deconfounder for Next-basket Recommendation) to mitigate the frequency bias. With the deconfounder theory and the causal diagram we propose, FENDER decomposes PIF with a neural tensor layer to obtain substitute confounders for users and items. Then, FENDER performs unbiased recommendations considering the effect of these substitute confounders. Experimental results demonstrate that FENDER has derived diverse and fair results compared to ten baseline models on three datasets while achieving competitive performance. Further experiments illustrate how FENDER balances users' historical purchases and potential interests.

Luyi Ma, Wanjia Sherry Zhang, Zezhong Fan et al.

On online advertising platforms, newly introduced promotional ads face the cold-start problem, as they lack sufficient user feedback for model training. In this work, we propose LLM-HYPER, a novel framework that treats large language models (LLMs) as hypernetworks to directly generate the parameters of the click-through rate (CTR) estimator in a training-free manner. LLM-HYPER uses few-shot Chain-of-Thought prompting over multimodal ad content (text and images) to infer feature-wise model weights for a linear CTR predictor. By retrieving semantically similar past campaigns via CLIP embeddings and formatting them into prompt-based demonstrations, the LLM learns to reason about customer intent, feature influence, and content relevance. To ensure numerical stability and serviceability, we introduce normalization and calibration techniques that align the generated weights with production-ready CTR distributions. Extensive offline experiments show that LLM-HYPER significantly outperforms cold-start baselines in NDCG$@10$ by 55.9\%. Our real-world online A/B test on one of the top e-commerce platforms in the U.S. demonstrates the strong performance of LLM-HYPER, which drastically reduces the cold-start period and achieves competitive performance. LLM-HYPER has been successfully deployed in production.

Yanan Cao, Farnaz Fallahi, Murali Mohana Krishna Dandu et al.

Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning and prediction across different domains. Yet, their ability to infer temporal regularities from structured behavioral data remains underexplored. This paper presents a systematic study investigating whether LLMs can predict time intervals between recurring user actions, such as repeated purchases, and how different levels of contextual information shape their predictive behavior. Using a simple but representative repurchase scenario, we benchmark state-of-the-art LLMs in zero-shot settings against both statistical and machine-learning models. Two key findings emerge. First, while LLMs surpass lightweight statistical baselines, they consistently underperform dedicated machine-learning models, showing their limited ability to capture quantitative temporal structure. Second, although moderate context can improve LLM accuracy, adding further user-level detail degrades performance. These results challenge the assumption that "more context leads to better reasoning". Our study highlights fundamental limitations of today's LLMs in structured temporal inference and offers guidance for designing future context-aware hybrid models that integrate statistical precision with linguistic flexibility.

Zezhong Fan, Ziheng Chen, Luyi Ma et al.

Generative recommendation (GeneRec) has introduced a new paradigm that represents items as discrete semantic tokens and predicts items in a generative manner. Despite its strong performance across multiple recommendation tasks, existing GeneRec approaches still suffer from severe popularity bias and may even exacerbate it. In this work, we conduct a comprehensive empirical analysis to uncover the root causes of this phenomenon, yielding two core insights: 1) imbalanced tokenization inherits and can further amplify popularity bias from historical item interactions; 2) current training procedures disproportionately favor popular tokens while neglecting semantic relationships among tokens, thereby intensifying popularity bias. Building on these insights, we propose CRAB, a post-hoc debiasing strategy for GeneRec that alleviates popularity bias by mitigating frequency imbalance among semantic tokens. Specifically, given a well-trained model, we first rebalance the codebook by splitting over-popular tokens while preserving their hierarchical semantic structure. Based on the adjusted codebook, we further introduce a tree-structured regularizer to enhance semantic consistency, encouraging more informative representations for unpopular tokens during training. Experiments on real-world datasets demonstrate that CRAB significantly improves recommendation performance by effectively alleviating popularity bias.

Tianning Dong, Luyi Ma, Varun Vasudevan et al.

Multimodal Large Language Models (MLLMs) are reshaping how modern agentic systems reason over sequential user-behavior data. However, whether textual or image representations of user behavior data are more effective for maximizing MLLM performance remains underexplored. We present \texttt{BehaviorLens}, a systematic benchmarking framework for assessing modality trade-offs in user-behavior reasoning across six MLLMs by representing transaction data as (1) a text paragraph, (2) a scatter plot, and (3) a flowchart. Using a real-world purchase-sequence dataset, we find that when data is represented as images, MLLMs next-purchase prediction accuracy is improved by 87.5% compared with an equivalent textual representation without any additional computational cost.

Luyi Ma, Xiaohan Li, Zezhong Fan et al.

Integrating diverse data modalities is crucial for enhancing the performance of personalized recommendation systems. Traditional models, which often rely on singular data sources, lack the depth needed to accurately capture the multifaceted nature of item features and user behaviors. This paper introduces a novel framework for multi-behavior recommendations, leveraging the fusion of triple-modality, which is visual, textual, and graph data through alignment with large language models (LLMs). By incorporating visual information, we capture contextual and aesthetic item characteristics; textual data provides insights into user interests and item features in detail; and graph data elucidates relationships within the item-behavior heterogeneous graphs. Our proposed model called Triple Modality Fusion (TMF) utilizes the power of LLMs to align and integrate these three modalities, achieving a comprehensive representation of user behaviors. The LLM models the user's interactions including behaviors and item features in natural languages. Initially, the LLM is warmed up using only natural language-based prompts. We then devise the modality fusion module based on cross-attention and self-attention mechanisms to integrate different modalities from other models into the same embedding space and incorporate them into an LLM. Extensive experiments demonstrate the effectiveness of our approach in improving recommendation accuracy. Further ablation studies validate the effectiveness of our model design and benefits of the TMF.

Najmeh Forouzandehmehr, Yijie Cao, Nikhil Thakurdesai et al.

The outfit generation problem involves recommending a complete outfit to a user based on their interests. Existing approaches focus on recommending items based on anchor items or specific query styles but do not consider customer interests in famous characters from movie, social media, etc. In this paper, we define a new Character-based Outfit Generation (COG) problem, designed to accurately interpret character information and generate complete outfit sets according to customer specifications such as age and gender. To tackle this problem, we propose a novel framework LVA-COG that leverages Large Language Models (LLMs) to extract insights from customer interests (e.g., character information) and employ prompt engineering techniques for accurate understanding of customer preferences. Additionally, we incorporate text-to-image models to enhance the visual understanding and generation (factual or counterfactual) of cohesive outfits. Our framework integrates LLMs with text-to-image models and improves the customer's approach to fashion by generating personalized recommendations. With experiments and case studies, we demonstrate the effectiveness of our solution from multiple dimensions.

Luyi Ma, Nikhil Thakurdesai, Jiao Chen et al.

Data processing is one of the fundamental steps in machine learning pipelines to ensure data quality. Majority of the applications consider the user-defined function (UDF) design pattern for data processing in databases. Although the UDF design pattern introduces flexibility, reusability and scalability, the increasing demand on machine learning pipelines brings three new challenges to this design pattern -- not low-code, not dependency-free and not knowledge-aware. To address these challenges, we propose a new design pattern that large language models (LLMs) could work as a generic data operator (LLM-GDO) for reliable data cleansing, transformation and modeling with their human-compatible performance. In the LLM-GDO design pattern, user-defined prompts (UDPs) are used to represent the data processing logic rather than implementations with a specific programming language. LLMs can be centrally maintained so users don't have to manage the dependencies at the run-time. Fine-tuning LLMs with domain-specific data could enhance the performance on the domain-specific tasks which makes data processing knowledge-aware. We illustrate these advantages with examples in different data processing tasks. Furthermore, we summarize the challenges and opportunities introduced by LLMs to provide a complete view of this design pattern for more discussions.

Luyi Ma, Wanjia Zhang, Kai Zhao et al.

Generative models have recently demonstrated strong potential in multi-behavior recommendation systems, leveraging the expressive power of transformers and tokenization to generate personalized item sequences. However, their adoption is hindered by (1) the lack of explicit information for token reasoning, (2) high computational costs due to quadratic attention complexity and dense sequence representations after tokenization, and (3) limited multi-scale modeling over user history. In this work, we propose GRACE (Generative Recommendation via journey-aware sparse Attention on Chain-of-thought tokEnization), a novel generative framework for multi-behavior sequential recommendation. GRACE introduces a hybrid Chain-of-Thought (CoT) tokenization method that encodes user-item interactions with explicit attributes from product knowledge graphs (e.g., category, brand, price) over semantic tokenization, enabling interpretable and behavior-aligned generation. To address the inefficiency of standard attention, we design a Journey-Aware Sparse Attention (JSA) mechanism, which selectively attends to compressed, intra-, inter-, and current-context segments in the tokenized sequence. Experiments on two real-world datasets show that GRACE significantly outperforms state-of-the-art baselines, achieving up to +106.9% HR@10 and +106.7% NDCG@10 improvement over the state-of-the-art baseline on the Home domain, and +22.1% HR@10 on the Electronics domain. GRACE also reduces attention computation by up to 48% with long sequences.

Zezhong Fan, Xiaohan Li, Luyi Ma et al.

Designing realistic multi-object scenes requires not only generating images, but also planning spatial layouts that respect semantic relations and physical plausibility. On one hand, while recent advances in diffusion models have enabled high-quality image generation, they lack explicit spatial reasoning, leading to unrealistic object layouts. On the other hand, traditional spatial planning methods in robotics emphasize geometric and relational consistency, but they struggle to capture semantic richness in visual scenes. To bridge this gap, in this paper, we propose LayoutAgent, an agentic framework that unifies vision-language reasoning with compositional diffusion for layout generation. Given multiple input images with target objects in them, our method first employs visual-language model to preprocess the inputs through segmentation, object size estimation, scene graph construction, and prompt rewriting. Then we leverage compositional diffusion-a method traditionally used in robotics-to synthesize bounding boxes that respect object relations encoded in the scene graph for spatial layouts. In the end, a foreground-conditioned image generator composes the complete scene by rendering the objects into the planned layout guided by designed prompts. Experiments demonstrate that LayoutAgent outperforms other state-of-the-art layout generation models in layout coherence, spatial realism and aesthetic alignment.

Jiao Chen, Luyi Ma, Xiaohan Li et al.

Knowledge Graphs (KGs) play a crucial role in enhancing e-commerce system performance by providing structured information about entities and their relationships, such as complementary or substitutable relations between products or product types, which can be utilized in recommender systems. However, relation labeling in KGs remains a challenging task due to the dynamic nature of e-commerce domains and the associated cost of human labor. Recently, breakthroughs in Large Language Models (LLMs) have shown surprising results in numerous natural language processing tasks. In this paper, we conduct an empirical study of LLMs for relation labeling in e-commerce KGs, investigating their powerful learning capabilities in natural language and effectiveness in predicting relations between product types with limited labeled data. We evaluate various LLMs, including PaLM and GPT-3.5, on benchmark datasets, demonstrating their ability to achieve competitive performance compared to humans on relation labeling tasks using just 1 to 5 labeled examples per relation. Additionally, we experiment with different prompt engineering techniques to examine their impact on model performance. Our results show that LLMs significantly outperform existing KG completion models in relation labeling for e-commerce KGs and exhibit performance strong enough to replace human labeling.

Luyi Ma, Jianpeng Xu, Jason H. D. Cho et al.

The complementary item recommender system (CIRS) recommends the complementary items for a given query item. Existing CIRS models consider the item co-purchase signal as a proxy of the complementary relationship due to the lack of human-curated labels from the huge transaction records. These methods represent items in a complementary embedding space and model the complementary relationship as a point estimation of the similarity between items vectors. However, co-purchased items are not necessarily complementary to each other. For example, customers may frequently purchase bananas and bottled water within the same transaction, but these two items are not complementary. Hence, using co-purchase signals directly as labels will aggravate the model performance. On the other hand, the model evaluation will not be trustworthy if the labels for evaluation are not reflecting the true complementary relatedness. To address the above challenges from noisy labeling of the copurchase data, we model the co-purchases of two items as a Gaussian distribution, where the mean denotes the co-purchases from the complementary relatedness, and covariance denotes the co-purchases from the noise. To do so, we represent each item as a Gaussian embedding and parameterize the Gaussian distribution of co-purchases by the means and covariances from item Gaussian embedding. To reduce the impact of the noisy labels during evaluation, we propose an independence test-based method to generate a trustworthy label set with certain confidence. Our extensive experiments on both the publicly available dataset and the large-scale real-world dataset justify the effectiveness of our proposed model in complementary item recommendations compared with the state-of-the-art models.