Bridging Past and Future: End-to-End Autonomous Driving with Historical Prediction and Planning

End-to-end autonomous driving unifies tasks in a differentiable framework, enabling planning-oriented optimization and attracting growing attention. Current methods aggregate historical information either through dense historical bird's-eye-view (BEV) features or by querying a sparse memory bank, following paradigms inherited from detection. However, we argue that these paradigms either omit historical information in motion planning or fail to align with its multi-step nature, which requires predicting or planning multiple future time steps. In line with the philosophy of future is a continuation of past, we propose BridgeAD, which reformulates motion and planning queries as multi-step queries to differentiate the queries for each future time step. This design enables the effective use of historical prediction and planning by applying them to the appropriate parts of the end-to-end system based on the time steps, which improves both perception and motion planning. Specifically, historical queries for the current frame are combined with perception, while queries for future frames are integrated with motion planning. In this way, we bridge the gap between past and future by aggregating historical insights at every time step, enhancing the overall coherence and accuracy of the end-to-end autonomous driving pipeline. Extensive experiments on the nuScenes dataset in both open-loop and closed-loop settings demonstrate that BridgeAD achieves state-of-the-art performance.

UniMamba: Unified Spatial-Channel Representation Learning with Group-Efficient Mamba for LiDAR-based 3D Object Detection

Recent advances in LiDAR 3D detection have demonstrated the effectiveness of Transformer-based frameworks in capturing the global dependencies from point cloud spaces, which serialize the 3D voxels into the flattened 1D sequence for iterative self-attention. However, the spatial structure of 3D voxels will be inevitably destroyed during the serialization process. Besides, due to the considerable number of 3D voxels and quadratic complexity of Transformers, multiple sequences are grouped before feeding to Transformers, leading to a limited receptive field. Inspired by the impressive performance of State Space Models (SSM) achieved in the field of 2D vision tasks, in this paper, we propose a novel Unified Mamba (UniMamba), which seamlessly integrates the merits of 3D convolution and SSM in a concise multi-head manner, aiming to perform "local and global" spatial context aggregation efficiently and simultaneously. Specifically, a UniMamba block is designed which mainly consists of spatial locality modeling, complementary Z-order serialization and local-global sequential aggregator. The spatial locality modeling module integrates 3D submanifold convolution to capture the dynamic spatial position embedding before serialization. Then the efficient Z-order curve is adopted for serialization both horizontally and vertically. Furthermore, the local-global sequential aggregator adopts the channel grouping strategy to efficiently encode both "local and global" spatial inter-dependencies using multi-head SSM. Additionally, an encoder-decoder architecture with stacked UniMamba blocks is formed to facilitate multi-scale spatial learning hierarchically. Extensive experiments are conducted on three popular datasets: nuScenes, Waymo and Argoverse 2. Particularly, our UniMamba achieves 70.2 mAP on the nuScenes dataset.

Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning

Training visual reinforcement learning (RL) in practical scenarios presents a significant challenge, $\textit{i.e.,}$ RL agents suffer from low sample efficiency in environments with variations. While various approaches have attempted to alleviate this issue by disentanglement representation learning, these methods usually start learning from scratch without prior knowledge of the world. This paper, in contrast, tries to learn and understand underlying semantic variations from distracting videos via offline-to-online latent distillation and flexible disentanglement constraints. To enable effective cross-domain semantic knowledge transfer, we introduce an interpretable model-based RL framework, dubbed Disentangled World Models (DisWM). Specifically, we pretrain the action-free video prediction model offline with disentanglement regularization to extract semantic knowledge from distracting videos. The disentanglement capability of the pretrained model is then transferred to the world model through latent distillation. For finetuning in the online environment, we exploit the knowledge from the pretrained model and introduce a disentanglement constraint to the world model. During the adaptation phase, the incorporation of actions and rewards from online environment interactions enriches the diversity of the data, which in turn strengthens the disentangled representation learning. Experimental results validate the superiority of our approach on various benchmarks.

Multi-Layer Visual Feature Fusion in Multimodal LLMs: Methods, Analysis, and Best Practices

Multimodal Large Language Models (MLLMs) have made significant advancements in recent years, with visual features playing an increasingly critical role in enhancing model performance. However, the integration of multi-layer visual features in MLLMs remains underexplored, particularly with regard to optimal layer selection and fusion strategies. Existing methods often rely on arbitrary design choices, leading to suboptimal outcomes. In this paper, we systematically investigate two core aspects of multi-layer visual feature fusion: (1) selecting the most effective visual layers and (2) identifying the best fusion approach with the language model. Our experiments reveal that while combining visual features from multiple stages improves generalization, incorporating additional features from the same stage typically leads to diminished performance. Furthermore, we find that direct fusion of multi-layer visual features at the input stage consistently yields superior and more stable performance across various configurations. We make all our code publicly available: https://github.com/EIT-NLP/Layer_Select_Fuse_for_MLLM.

ULTHO: Ultra-Lightweight yet Efficient Hyperparameter Optimization in Deep Reinforcement Learning

Hyperparameter optimization (HPO) is a billion-dollar problem in machine learning, which significantly impacts the training efficiency and model performance. However, achieving efficient and robust HPO in deep reinforcement learning (RL) is consistently challenging due to its high non-stationarity and computational cost. To tackle this problem, existing approaches attempt to adapt common HPO techniques (e.g., population-based training or Bayesian optimization) to the RL scenario. However, they remain sample-inefficient and computationally expensive, which cannot facilitate a wide range of applications. In this paper, we propose ULTHO, an ultra-lightweight yet powerful framework for fast HPO in deep RL within single runs. Specifically, we formulate the HPO process as a multi-armed bandit with clustered arms (MABC) and link it directly to long-term return optimization. ULTHO also provides a quantified and statistical perspective to filter the HPs efficiently. We test ULTHO on benchmarks including ALE, Procgen, MiniGrid, and PyBullet. Extensive experiments demonstrate that the ULTHO can achieve superior performance with simple architecture, contributing to the development of advanced and automated RL systems.

Unified Arbitrary-Time Video Frame Interpolation and Prediction

Video frame interpolation and prediction aim to synthesize frames in-between and subsequent to existing frames, respectively. Despite being closely-related, these two tasks are traditionally studied with different model architectures, or same architecture but individually trained weights. Furthermore, while arbitrary-time interpolation has been extensively studied, the value of arbitrary-time prediction has been largely overlooked. In this work, we present uniVIP - unified arbitrary-time Video Interpolation and Prediction. Technically, we firstly extend an interpolation-only network for arbitrary-time interpolation and prediction, with a special input channel for task (interpolation or prediction) encoding. Then, we show how to train a unified model on common triplet frames. Our uniVIP provides competitive results for video interpolation, and outperforms existing state-of-the-arts for video prediction. Codes will be available at: https://github.com/srcn-ivl/uniVIP

Exploring Simple Siamese Network for High-Resolution Video Quality Assessment

In the research of video quality assessment (VQA), two-branch network has emerged as a promising solution. It decouples VQA with separate technical and aesthetic branches to measure the perception of low-level distortions and high-level semantics respectively. However, we argue that while technical and aesthetic perspectives are complementary, the technical perspective itself should be measured in semantic-aware manner. We hypothesize that existing technical branch struggles to perceive the semantics of high-resolution videos, as it is trained on local mini-patches sampled from videos. This issue can be hidden by apparently good results on low-resolution videos, but indeed becomes critical for high-resolution VQA. This work introduces SiamVQA, a simple but effective Siamese network for highre-solution VQA. SiamVQA shares weights between technical and aesthetic branches, enhancing the semantic perception ability of technical branch to facilitate technical-quality representation learning. Furthermore, it integrates a dual cross-attention layer for fusing technical and aesthetic features. SiamVQA achieves state-of-the-art accuracy on high-resolution benchmarks, and competitive results on lower-resolution benchmarks. Codes will be available at: https://github.com/srcn-ivl/SiamVQA

Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs

We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

SoFar: Language-Grounded Orientation Bridges Spatial Reasoning and Object Manipulation

Spatial intelligence is a critical component of embodied AI, promoting robots to understand and interact with their environments. While recent advances have enhanced the ability of VLMs to perceive object locations and positional relationships, they still lack the capability to precisely understand object orientations-a key requirement for tasks involving fine-grained manipulations. Addressing this limitation not only requires geometric reasoning but also an expressive and intuitive way to represent orientation. In this context, we propose that natural language offers a more flexible representation space than canonical frames, making it particularly suitable for instruction-following robotic systems. In this paper, we introduce the concept of semantic orientation, which defines object orientations using natural language in a reference-frame-free manner (e.g., the ''plug-in'' direction of a USB or the ''handle'' direction of a knife). To support this, we construct OrienText300K, a large-scale dataset of 3D models annotated with semantic orientations that link geometric understanding to functional semantics. By integrating semantic orientation into a VLM system, we enable robots to generate manipulation actions with both positional and orientational constraints. Extensive experiments in simulation and real world demonstrate that our approach significantly enhances robotic manipulation capabilities, e.g., 48.7% accuracy on Open6DOR and 74.9% accuracy on SIMPLER.

TINQ: Temporal Inconsistency Guided Blind Video Quality Assessment

Blind video quality assessment (BVQA) has been actively researched for user-generated content (UGC) videos. Recently, super-resolution (SR) techniques have been widely applied in UGC. Therefore, an effective BVQA method for both UGC and SR scenarios is essential. Temporal inconsistency, referring to irregularities between consecutive frames, is relevant to video quality. Current BVQA approaches typically model temporal relationships in UGC videos using statistics of motion information, but inconsistencies remain unexplored. Additionally, different from temporal inconsistency in UGC videos, such inconsistency in SR videos is amplified due to upscaling algorithms. In this paper, we introduce the Temporal Inconsistency Guided Blind Video Quality Assessment (TINQ) metric, demonstrating that exploring temporal inconsistency is crucial for effective BVQA. Since temporal inconsistencies vary between UGC and SR videos, they are calculated in different ways. Based on this, a spatial module highlights inconsistent areas across consecutive frames at coarse and fine granularities. In addition, a temporal module aggregates features over time in two stages. The first stage employs a visual memory capacity block to adaptively segment the time dimension based on estimated complexity, while the second stage focuses on selecting key features. The stages work together through Consistency-aware Fusion Units to regress cross-time-scale video quality. Extensive experiments on UGC and SR video quality datasets show that our method outperforms existing state-of-the-art BVQA methods. Code is available at https://github.com/Lighting-YXLI/TINQ.