Multimodal Retrieval-Augmented Generation (MRAG) has shown promise in mitigating hallucinations in Multimodal Large Language Models (MLLMs) by incorporating external knowledge. However, existing methods typically adhere to rigid retrieval paradigms by mimicking fixed retrieval trajectories and thus fail to fully exploit the knowledge of different retrieval experts through dynamic interaction based on the model's knowledge needs or evolving reasoning states. To overcome this limitation, we introduce Mixture-of-Retrieval Experts (MoRE), a novel framework that enables MLLMs to collaboratively interact with diverse retrieval experts for more effective knowledge exploitation. Specifically, MoRE learns to dynamically determine which expert to engage with, conditioned on the evolving reasoning state. To effectively train this capability, we propose Stepwise Group Relative Policy Optimization (Step-GRPO), which goes beyond sparse outcome-based supervision by encouraging MLLMs to interact with multiple retrieval experts and synthesize fine-grained rewards, thereby teaching the MLLM to fully coordinate all experts when answering a given query. Experimental results on diverse open-domain QA benchmarks demonstrate the effectiveness of MoRE, achieving average performance gains of over 7% compared to competitive baselines. Notably, MoRE exhibits strong adaptability by dynamically coordinating heterogeneous experts to precisely locate relevant information, validating its capability for robust, reasoning-driven expert collaboration. All codes and data are released on https://github.com/OpenBMB/MoRE.

Evaluating autonomous vehicles’ performance in complex, long-tail traffic scenarios, especially under extreme conditions, often highlights the limitations of existing methods in generating realistic and challenging scenarios, which can affect vehicle safety and reliability. To address these gaps, this paper proposes a Scenario-Adaptive Gradient Adjustment (SAGA) network, an adversarial model specifically designed to generate intricate traffic scenarios that closely mimic real-world dynamics. The SAGA network includes a generator and a victim model, where the generator uses adversarial sequences based on the kinematic bicycle model to simulate dynamic vehicle characteristics and calculate precise gradients. These gradients are then used to perturb the victim model, creating safety-critical scenarios essential for evaluating autonomous vehicle performance, such as emergency evasions and complex intersection navigation. Additionally, we use a k-means++ clustering method tailored to categorize ten types of safety-critical scenarios for autonomous driving. The generated scenarios are comprehensive, diverse, and challenging, providing a robust testing environment for autonomous vehicles. Simulation results demonstrate the SAGA network’s effectiveness, significantly outperforming traditional non-transparent optimization and the KING method. SAGA achieved a 25% higher success rate than non-transparent optimization and a 2% improvement over the KING method in generating complex scenarios, along with an 8% increase in interpretability, reaching 80%. These findings highlight the capability of SAGA-generated scenarios to thoroughly assess autonomous vehicle performance under diverse traffic conditions, ensuring safer operations.

Kunlun Zhu, Yifan Luo, Dingling Xu, Yukun Yan, Zhenghao Liu, Shi Yu, Ruobing Wang, Shuo Wang, Yishan Li, Nan Zhang, Xu Han, Zhiyuan Liu, Maosong Sun. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The data that records road traffic accidents often hides important information. If we can analyze the main factors that affect the severity of the accident in the data, it has very good practical significance for avoiding the occurrence of major traffic accidents. In this paper, a method called CFS-BAS-BP is proposed to realize the identification of potential traffic accident risk factors based on combinatorial feature selection and bionic neural network. This method combines the advantages of Information Gain, Chi-square test and Random Forest in feature selection to screen the important features affecting traffic accidents to the greatest extent, and filter out irrelevant and redundant features in the data. As a result, 14 features that influence the occurrence and severity of traffic accidents are selected from 25 features in the dataset At the same time, the back propagation neural network optimized by the beetle antennae search algorithm is used to analyze the quantitative relationship between traffic data features and traffic accident risk level, so as to identify the main risk factors affecting the severity of traffic accidents in different scenarios. Through a large number of experiments on the real traffic accident data set STATS19, we prove that CFS-BAS-BP model has a significant effect on the selection of accident-related features, and its accuracy rate of identifying the risk factors affecting the severity of traffic accidents reaches 86.1%.

Providing reasonable explanations for a specific suggestion given by the recommender can help users trust the system more. As logic rule-based inference is concise, transparent, and aligned with human cognition, it can be adopted to improve the interpretability of recommendation models. Previous work that interprets user preference with logic rules merely focuses on the construction of rules while neglecting the usage of feature embeddings. This limits the model in capturing implicit relationships between features. In this article, we aim to improve both the effectiveness and explainability of recommendation models by simultaneously representing logic rules and feature embeddings. We propose a novel model-intrinsic explainable recommendation method named Feature-Enhanced Neural Collaborative Reasoning (FENCR) . The model automatically extracts representative logic rules from massive possibilities in a data-driven way. In addition, we utilize feature interaction-based neural modules to represent logic operators on embeddings. Experiments on two large public datasets show our model outperforms state-of-the-art neural logical recommendation models. Further case analyses demonstrate that FENCR can derive reasonable rules, indicating its high robustness and expandability. 1

The explainability of recommender systems has attracted significant attention in academia and industry. Many efforts have been made for explainable recommendations, yet evaluating the quality of the explanations remains a challenging and unresolved issue. In recent years, leveraging LLMs as evaluators presents a promising avenue in Natural Language Processing tasks (e.g., sentiment classification, information extraction), as they perform strong capabilities in instruction following and common-sense reasoning. However, evaluating recommendation explanatory texts is different from these NLG tasks, as its criteria are related to human perceptions and are usually subjective.

In this study, the latest released satellite based precipitation products - Global Precipitation Measurement (GPM) mission Level 3 product Integrated Multi-satellitE Retrievals for GPM (IMERG) and version 7 of Tropic Rainfall Measurement Mission (TRMM 3B42V7) data is evaluated and applied with a distributed hydrological model to examine the precipitation detection and performance in hydrological simulating over the Mekong River Basin (MRB) during 2014/4/1 to 2016/1/31. About 137 rain gauges stations were collected to carry out a pixel-point comparison between observation and satellite precipitation. Moreover, daily discharge observation data from five discharge gauges were used to evaluate the performance of hydrological simulation. The result demonstrate that: 1) IMERG data show more precision in both heavy and light rain detection than 3B42V7; 2) IMERG performs better than 3B42V7 when driving hydrologic model, giving the fact that the simulated discharge result from IMERG is more accurate and stable.

Decomposition is performed for the 4×4 SMAP radar channel covariance matrix and the correlation between resulting components, surface soil moisture and vegetation is examined. Globally, the first principal component is the most dominant and the correlation coefficients with respect to soil mortise is highest (R2≥ 0.8) in regions with fractional ground cover and sufficient temporal dynamics of soil moisture.