Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) are prone to hallucinations, i.e., the generated content that is nonsensical or unfaithful to input sources. Unlike in LLMs, hallucinations in MLLMs often stem from the sensitivity of text decoder to visual tokens, leading to a phenomenon akin to "amnesia" about visual information. To address this issue, we propose MemVR, a novel decoding paradigm inspired by common cognition: when the memory of an image seen the moment before is forgotten, people will look at it again for factual answers. Following this principle, we treat visual tokens as supplementary evidence, re-injecting them into the MLLM through Feed Forward Network (FFN) as “key-value memory” at the middle trigger layer. This look-twice mechanism occurs when the model exhibits high uncertainty during inference, effectively enhancing factual alignment. Comprehensive experimental evaluations demonstrate that MemVR significantly mitigates hallucination across various MLLMs and excels in general benchmarks without incurring additional time overhead.

Given that substantial amounts of domain-specific knowledge are stored in structured formats, such as web data organized through HTML, Large Language Models (LLMs) are expected to fully comprehend this structured information to broaden their applications in various real-world downstream tasks. Current approaches for applying LLMs to structured data fall into two main categories: serialization-based and operation-based methods. Both approaches, whether relying on serialization or using SQL-like operations as an intermediary, encounter difficulties in fully capturing structural relationships and effectively handling sparse data. To address these unique characteristics of structured data, we propose HyperG, a hypergraph-based generation framework aimed at enhancing LLMs' ability to process structured knowledge. Specifically, HyperG first augment sparse data with contextual information, leveraging the generative power of LLMs, and incorporate a prompt-attentive hypergraph learning (PHL) network to encode both the augmented information and the intricate structural relationships within the data. To validate the effectiveness and generalization of HyperG, we conduct extensive experiments across two different downstream tasks requiring structured knowledge. Our code is publicly available at: https://github.com/s1ruihuang/HyperG.

Sequential recommender systems (SRSs) are designed to suggest relevant items to users by analyzing their interaction sequences. However, SRSs often suffer from exposure bias in these sequences due to imbalanced item exposure and varied user activity levels, creating a self-reinforcing loop favoring popular items regardless of their true relevance. Most SRSs only focus on item dependencies to address exposure bias, while overlooking user-side exposure bias and the rich semantics behind interactions. These oversights result in a limited understanding of less active users' preferences and inaccurate preference capture for less exposed items, exacerbating exposure biases. Towards this end, we propose a novel methodLLM-enhancedDualPropensity ScoreEstimation (LDPE), which synergistically integrates Large Language Models (LLMs) and causal inference. First, LDPE leverages LLMs' superior ability in capturing rich semantics from textual data and then integrates collaborative information to generate debiased semantic-rich LLM-based user/item embeddings. With these debiased item/user embeddings, LDPE estimates time-aware debiased propensity scores from both the item and user sides. These dual propensity scores can fully mitigate exposure bias by considering item popularity, user activity levels, and temporal dynamics. Lastly, LDPE employs the transformer as the backbone of our method, incorporating estimated dual propensity scores for accurately predicting users' true preferences. Extensive experiments show that our LDPE outperforms state-of-the-art baselines in terms of recommendation performance.

Predicting user satisfaction over time is crucial in news recommendations, as users’ preferences are significantly influenced by various time-variant factors. Traditional correlation-based recommenders often suffer from redundant relationships, which can undermine their effectiveness over time. This work takes a time-aware causal approach to news recommendations, treating exposed news at a predicted time as the treatment variable and the resulting user satisfaction as the outcome variable. Capturing the evolving causal effects of exposed news items on user satisfaction poses significant challenges, particularly stemming from the need to model complex dependencies among time-variant covariates, such as news popularity and recency, as well as to effectively leverage the inherent user preferences embedded in time-invariant covariates. To these ends, we propose the CA u S al T ime-aware Rec ommender, named CAST-Rec , which accounts for the causal influences of both time-variant and time-invariant covariates. Specifically, we model the intricate causal dependencies among time-variant covariates through a series of transformer-based causal blocks. For time-invariant covariates, we utilize the semantic understanding and generative capabilities of Large Language Models (LLMs) to infer inherent user preferences while mitigating potential confounding effects. Extensive experiments demonstrate the superior performance of CAST-Rec compared to various news recommendation models and across multiple LLM implementations.