Yaning Jia, Chunhui Zhang, Xingjian Diao, Xiangchi Yuan, Zhongyu Ouyang, Chiyu Ma, Soroush Vosoughi. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

What enables large language models (LLMs) to effectively model user preferences in sequential recommendation?Our investigation reveals that existing preference-alignment approaches largely rely on binary pairwise comparisons, overlooking two critical factors: preference intensity-the structured strength of affinity or aversion -and temporal context-the extent to which recent interactions better reflect a user's current intent.Through controlled experiments, we show that leveraging comprehensive feedback with structured preference signals substantially improves recommendation performance, indicating that binary modeling discards essential information.Motivated by these findings, we propose RecPO, a unified preference optimization framework that maps both explicit and implicit feedback into a common preference signal and constructs adaptive reward margins that jointly account for preference intensity and interaction recency.Experiments across five datasets show that RecPO consistently outperforms state-of-the-art baselines while exhibiting behavioral patterns aligned with human decision-making, including favoring immediate satisfaction, maintaining preference coherence, and avoiding dispreferred items.Our results highlight that preference intensity and temporal context are fundamental ingredients for effective LLM-based recommendation.Code:

Language models (LMs) require effective episodic grounding—the ability to learn from and apply past experiences—to perform well at physical planning tasks. While current approaches struggle with scalability and integration of episodic memory, which is particularly limited for medium-sized LMs (7B parameters), larger LMs (70-405B) offer untapped potential through their hierarchical representations and extensive pre-trained knowledge. Therefore, to unlock larger LMs’ potential for grounding, we present a scalable weak-to-strong episodic learning framework that efficiently transfers episodic behaviors from smaller to larger LMs. It uses Monte Carlo tree search for structured experience collection with a novel distillation method that preserves LM capabilities while incorporating episodic memory. This enables larger LMs to leverage their inherent advantages for improved physical planning. Experiments show our solution outperforms top proprietary LMs by 3.45% across diverse planning and question-answering tasks. Layer-wise probing reveals systematic improvements in task alignment, particularly in later LM layers. It shows stable generalization to even unseen scenarios, even as planning steps increase, whereas baselines deteriorate sharply beyond a complexity threshold of four planning steps.

Xingjian Diao, Chunhui Zhang, Keyi Kong, Weiyi Wu, Chiyu Ma, Zhongyu Ouyang, Peijun Qing, Soroush Vosoughi, Jiang Gui. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Theory-of-mind (ToM) enables humans to infer mental states—such as beliefs, desires, and intentions—forming the foundation of social cognition. Existing computational ToM methods rely on structured workflows with ToM-specific priors or deep model fine-tuning but struggle with scalability in multimodal environments. They remain trapped within the gravitational pull of multi-step planning complexity, failing to generalize as task demands increase. To overcome these limitations, we propose a scalable Bayesian ToM planner. It breaks down ToM complexity into stepwise Bayesian updates. Meanwhile, weak-to-strong control specializes smaller LMs to refine ToM-specific likelihood estimation, transferring their ToM reasoning behavior to larger LMs (7B to 405B) for social and world knowledge integration. This synergistic approach enables scalability, aligning large-model inference with human mental states with Bayesian principles. Extensive experiments demonstrate a 4.6% improvement in accuracy over state-of-the-art methods on multimodal ToM benchmarks, including unseen scenarios, establishing a new standard for modeling human mental states in complex environments.

In modern social media, recommender systems (RecSys) rely on the click-through rate (CTR) as the standard metric to evaluate user engagement. CTR prediction is traditionally framed as a binary classification task to predict whether a user will interact with a given item. However, this approach overlooks the complexity of real-world social modeling, where user, item, and their interactive features change dynamically in fast-paced online environments. This dynamic nature often leads to model instability, reflected in overfitting short-term fluctuations rather than higher-level interactive patterns. While overfitting calls for more scaled and refined supervisions, current solutions often rely on binary labels that overly simplify fine-grained user preferences through the thresholding process, which significantly reduces the richness of the supervision. Therefore, we aim to alleviate the overfitting problem by increasing the supervision bandwidth in CTR training. Specifically, (i) theoretically, we formulate the impact of fine-grained preferences on model stability as a Lipschitz constrain; (ii) empirically, we discover that scaling the supervision bandwidth can act as an implicit Lipschitz regularizer, stably optimizing existing CTR models to achieve better generalizability. Extensive experiments show that this scaled supervision significantly and consistently improves the optimization process and the performance of existing CTR models, even without the need for additional hyperparameter tuning.

Graph knowledge has been proven effective in enhancing item rankings in recommender systems (RecSys), particularly during the retrieval stage.However, its application in the ranking stage, especially when richer contextual information in user-item interactions is available, remains underexplored.A major challenge lies in the substantial computational cost associated with repeatedly retrieving neighborhood information from billions of items stored in distributed systems.This resource-intensive requirement makes it difficult to scale graph-based methods in practical RecSys.To bridge this gap, we first demonstrate that incorporating graphs in the ranking stage improves ranking qualities.Notably, while the improvement is evident, we show that the substantial computational overheads entailed by graphs are prohibitively expensive for real-world recommendations.In light of this, we propose a non-parametric strategy that utilizes graph convolution for re-ranking only during test time.Our strategy circumvents the notorious computational overheads from graph convolution during training, and utilizes structural knowledge hidden in graphs on-the-fly during testing.It can be used as a plug-and-play module and easily employed to enhance the ranking ability of various ranking layers of a real-world RecSys with significantly reduced computational overhead.Through comprehensive experiments across four benchmark datasets with varying levels of sparsity, we demonstrate that our strategy yields noticeable improvements (i.e., 8.1% on average) during testing time with little to no additional computational overheads (i.e., 0.5% on average).

This study explores the inherent limitations of Large Language Models (LLMs) from a scaling perspective, focusing on the upper bounds of their cognitive capabilities.We integrate insights from cognitive science to quantitatively examine how LLMs perform on n-back tasks-a benchmark used to assess working memory, which involves temporarily holding and manipulating information.Our findings reveal that despite increased model size, LLMs still face significant challenges in holding and processing information effectively, especially under complex task conditions.We also assess various prompting strategies, revealing their diverse impacts on LLM performance.The results highlight the struggle of current LLMs to autonomously discover optimal problemsolving patterns without heavily relying on manually corrected prompts.To move beyond these constraints, fundamental improvements in the planning and search of LLMs are essential for them to reason autonomously.Improving these capabilities will reduce the reliance on external corrections and enable LLMs to become more autonomous in their problemsolving processes.

Self-supervised learning (SSL) has gained increasing attention in the graph learning community, owing to its capability of enabling powerful models pre-trained on large unlabeled graphs for general purposes, facilitating quick adaptation to specific domains. Though promising, existing graph SSL frameworks often struggle to capture both high-level abstract features and fine-grained features simultaneously, leading to sub-optimal generalization abilities across different downstream tasks. To bridge this gap, we present Multi-granularity Graph Semantic Ensemble via Knowledge Distillation, namely MGSE, a plug-and-play graph knowledge distillation framework that can be applied to any existing graph SSL framework to enhance its performance by incorporating the concept of multi-granularity. Specifically, MGSE captures multi-granular knowledge by employing multiple student models to learn from a single teacher model, conditioned by probability distributions with different granularities. We apply it to six state-of-the-art graph SSL frameworks and evaluate their performances over multiple graph datasets across different domains, the experimental results show that MGSE can consistently boost the performance of these existing graph SSL frameworks with up to 9.2% improvement.

Collaborative filtering (CF) is a prevalent technique utilized in recommender systems (RSs), and has been extensively deployed in various real-world applications. A recent study in CF focuses on improving the quality of representations from the perspective of alignment and uniformity on the hyperspheres for enhanced recommendation performance. It promotes alignment to increase the similarity between representations of interacting users and items, and enhances uniformity to have more uniformly distributed user and item representations within their respective hyperspheres. However, although alignment and uniformity are enforced by two different optimized objectives, respectively, they jointly constitute the supervised signals for model training. Models trained with only supervised signals in labeled data can inevitably overfit the noise introduced by label sampling variance, even with i.i.d. datasets. This overfitting to noise further compromises the model's generalizability and performance on unseen testing data. To address this issue, in this study, we aim to mitigate the effect caused by the sampling variance in labeled training data to improve representation generalizability from the perspective of alignment and uniformity. Representations with more generalized alignment and uniformity further lead to improved model performance on testing data. Specifically, we model the data as a user-item interaction bipartite graph, and apply a graph neural network (GNN) to learn the user and item representations. This graph modeling approach allows us to integrate self-supervised signals into the RS, by performing self-supervised contrastive learning on the user and item representations from the perspective of label-irrelevant alignment and uniformity. Since the representations are less dependent on label supervision, they can capture more label-irrelevant data structures and patterns, leading to more generalized alignment and uniformity. We conduct extensive experiments on three benchmark datasets to demonstrate the superiority of our framework (i.e., improved performance and faster convergence speed). Our codes: https://github.com/zyouyang/AUPlus

Among the existing self-supervised learning (SSL) methods for graphs, graph contrastive learning (GCL) frameworks usually automatically generate supervision by transforming the same graph into different views through graph augmentation operations. The computation-efficient augmentation techniques enable the prevalent usage of GCL to alleviate the supervision shortage issue. Despite the remarkable performance of those GCL methods, the InfoMax principle used to guide the optimization of GCL has been proven to be insufficient to avoid redundant information without losing important features. In light of this, we introduce the Graph Contrastive Learning with Cross-View Reconstruction (GCVR), aiming to learn robust and sufficient representation from graph data. Specifically, GCVR introduces a cross-view reconstruction mechanism based on conventional graph contrastive learning to elicit those essential features from raw graphs. Besides, we introduce an extra adversarial view perturbed from the original view in the contrastive loss to pursue the intactness of the graph semantics and strengthen the representation robustness. We empirically demonstrate that our proposed model outperforms the state-of-the-art baselines on graph classification tasks over multiple benchmark datasets.

Graph Neural Networks (GNNs) are powerful models for non-Euclidean data, but their training is often accentuated by massive unnecessary computation: on the one hand, training on non-Euclidean data has relatively high computational cost due to its irregular density properties; on the other hand, the class imbalance property often associated with non-Euclidean data cannot be alleviated by the massiveness of the data, thus hindering the generalisation of the models. To address the above issues, theoretically, we start with a hypothesis about the effectiveness of using a subset of training data for GNNs, which is guaranteed by the gradient distance between the subset and the full set. Empirically, we also observe that a subset of the data can provide informative gradients for model optimization and which changes over time dynamically. We name this phenomenon dynamic data sparsity. Additionally, we find that pruned sparse contrastive models may miss valuable information, leading to a large loss value on the informative subset. Motivated by the above findings, we develop a unified data model dynamic sparsity framework called Data Decantation (DataDec) to address the above challenges. The key idea of DataDec is to identify the informative subset dynamically during the training process by applying sparse graph contrastive learning. The effectiveness of DataDec is comprehensively evaluated on graph benchmark datasets and we also verify its generalizability on image data.

Opioids (e.g., oxycodone and morphine) are highly addictive prescription (aka Rx) drugs which can be easily overprescribed and lead to opioid overdose. Recently, the opioid epidemic is increasingly serious across the US as its related deaths have risen at alarming rates. To combat the deadly opioid epidemic, a state-run prescription drug monitoring program (PDMP) has been established to alleviate the drug over-prescribing problem in the US. Although PDMP provides a detailed prescription history related to opioids, it is still not enough to prevent opioid overdose because it cannot predict over-prescribing risk. In addition, existing machine learning-based methods mainly focus on drug doses while ignoring other prescribing patterns behind patients' historical records, thus resulting in suboptimal performance. To this end, we propose a novel model DDHGNN - Disentangled Dynamic Heterogeneous Graph Neural Network, for over-prescribing prediction. Specifically, we abstract the PDMP data into a dynamic heterogeneous graph which comprehensively depicts the prescribing and dispensing (P&D) relationships. Then, we design a dynamic heterogeneous graph neural network to learn patients' representations. Furthermore, we devise an adversarial disentangler to learn a disentangled representation which is particularly related to the prescribing patterns. Extensive experiments on a 1-year anonymous PDMP data demonstrate that DDHGNN outperforms state-of-the-art methods, revealing its promising future in preventing opioid overdose.