Stable and efficient training of ReLU networks with large depth is highly sensitive to weight initialization. Improper initialization can cause permanent neuron inactivation dying ReLU and exacerbate gradient instability as network depth increases. Methods such as He, Xavier, and orthogonal initialization preserve variance or promote approximate isometry. However, they do not necessarily regulate the pre-activation mean or control activation sparsity, and their effectiveness often diminishes in very deep architectures. This work introduces an orthogonal initialization specifically optimized for ReLU by solving an optimization problem on the Stiefel manifold, thereby preserving scale and calibrating the pre-activation statistics from the outset. A family of closed-form solutions and an efficient sampling scheme are derived. Theoretical analysis at initialization shows that prevention of the dying ReLU problem, slower decay of activation variance, and mitigation of gradient vanishing, which together stabilize signal and gradient flow in deep architectures. Empirically, across MNIST, Fashion-MNIST, multiple tabular datasets, few-shot settings, and ReLU-family activations, our method outperforms previous initializations and enables stable training in deep networks.

Robotic manipulation in unstructured environments requires systems that can generalize across diverse tasks while maintaining robust and reliable performance. We introduce {GVF-TAPE}, a closed-loop framework that combines generative visual foresight with task-agnostic pose estimation to enable scalable robotic manipulation. GVF-TAPE employs a generative video model to predict future RGB-D frames from a single side-view RGB image and a task description, offering visual plans that guide robot actions. A decoupled pose estimation model then extracts end-effector poses from the predicted frames, translating them into executable commands via low-level controllers. By iteratively integrating video foresight and pose estimation in a closed loop, GVF-TAPE achieves real-time, adaptive manipulation across a broad range of tasks. Extensive experiments in both simulation and real-world settings demonstrate that our approach reduces reliance on task-specific action data and generalizes effectively, providing a practical and scalable solution for intelligent robotic systems.

Surgical video understanding is crucial for facilitating Computer-Assisted Surgery (CAS) systems. Despite significant progress in existing studies, two major limitations persist, including inadequate visual content perception and insufficient temporal awareness in surgical videos, and hinder the development of versatile CAS solutions. In this work, we propose the SurgLLM framework, an effective large multimodal model tailored for versatile surgical video understanding tasks with enhanced spatial focus and temporal awareness. Specifically, to empower the spatial focus of surgical videos, we first devise Surgical Context-aware Multimodal Pretraining (Surg-Pretrain) for the video encoder of SurgLLM, by performing instrument-centric Masked Video Reconstruction (MV-Recon) and subsequent multimodal alignment. To incorporate surgical temporal knowledge into SurgLLM, we further propose Temporal-aware Multimodal Tuning (TM-Tuning) to enhance temporal reasoning with interleaved multimodal embeddings. Moreover, to accommodate various understanding tasks of surgical videos without conflicts, we devise a Surgical Task Dynamic Ensemble to efficiently triage a query with optimal learnable parameters in our SurgLLM. Extensive experiments performed on diverse surgical video understanding tasks, including captioning, general VQA, and temporal VQA, demonstrate significant improvements over the state-of-the-art approaches, validating the effectiveness of our SurgLLM in versatile surgical video understanding. The source code is available at https://github.com/franciszchen/SurgLLM.

Air pollution monitoring in resource-constrained regions remains challenging due to sparse sensor deployment and limited infrastructure. This work introduces AQFusionNet, a multimodal deep learning framework for robust Air Quality Index (AQI) prediction. The framework integrates ground-level atmospheric imagery with pollutant concentration data using lightweight CNN backbones (MobileNetV2, ResNet18, EfficientNet-B0). Visual and sensor features are combined through semantically aligned embedding spaces, enabling accurate and efficient prediction. Experiments on more than 8,000 samples from India and Nepal demonstrate that AQFusionNet consistently outperforms unimodal baselines, achieving up to 92.02% classification accuracy and an RMSE of 7.70 with the EfficientNet-B0 backbone. The model delivers an 18.5% improvement over single-modality approaches while maintaining low computational overhead, making it suitable for deployment on edge devices. AQFusionNet provides a scalable and practical solution for AQI monitoring in infrastructure-limited environments, offering robust predictive capability even under partial sensor availability.

Intensive studies have been conducted in recent years to integrate neural networks with physics models to balance model accuracy and interpretability. One recently proposed approach, named Physics-Enhanced Residual Learning (PERL), is to use learning to estimate the residual between the physics model prediction and the ground truth. Numeral examples suggested that integrating such residual with physics models in PERL has three advantages: (1) a reduction in the number of required neural network parameters; (2) faster convergence rates; and (3) fewer training samples needed for the same computational precision. However, these numerical results lack theoretical justification and cannot be adequately explained. This paper aims to explain these advantages of PERL from a theoretical perspective. We investigate a general class of problems with Lipschitz continuity properties. By examining the relationships between the bounds to the loss function and residual learning structure, this study rigorously proves a set of theorems explaining the three advantages of PERL. Several numerical examples in the context of automated vehicle trajectory prediction are conducted to illustrate the proposed theorems. The results confirm that, even with significantly fewer training samples, PERL consistently achieves higher accuracy than a pure neural network. These results demonstrate the practical value of PERL in real world autonomous driving applications where corner case data are costly or hard to obtain. PERL therefore improves predictive performance while reducing the amount of data required.

Reinforcement Learning (RL) is known for its strong decision-making capabilities and has been widely applied in various real-world scenarios. However, with the increasing availability of offline datasets and the lack of well-designed online environments from human experts, the challenge of generalization in offline RL has become more prominent. Due to the limitations of offline data, RL agents trained solely on collected experiences often struggle to generalize to new tasks or environments. To address this challenge, we propose LLM-Driven Policy Diffusion (LLMDPD), a novel approach that enhances generalization in offline RL using task-specific prompts. Our method incorporates both text-based task descriptions and trajectory prompts to guide policy learning. We leverage a large language model (LLM) to process text-based prompts, utilizing its natural language understanding and extensive knowledge base to provide rich task-relevant context. Simultaneously, we encode trajectory prompts using a transformer model, capturing structured behavioral patterns within the underlying transition dynamics. These prompts serve as conditional inputs to a context-aware policy-level diffusion model, enabling the RL agent to generalize effectively to unseen tasks. Our experimental results demonstrate that LLMDPD outperforms state-of-the-art offline RL methods on unseen tasks, highlighting its effectiveness in improving generalization and adaptability in diverse settings.

Vision-Language-Action (VLA) models are a promising path to realizing generalist embodied agents that can quickly adapt to new tasks, modalities, and environments. However, methods for interpreting and steering VLAs fall far short of classical robotics pipelines, which are grounded in explicit models of kinematics, dynamics, and control. This lack of mechanistic insight is a central challenge for deploying learned policies in real-world robotics, where robustness and explainability are critical. Motivated by advances in mechanistic interpretability for large language models, we introduce the first framework for interpreting and steering VLAs via their internal representations, enabling direct intervention in model behavior at inference time. We project feedforward activations within transformer layers onto the token embedding basis, identifying sparse semantic directions - such as speed and direction - that are causally linked to action selection. Leveraging these findings, we introduce a general-purpose activation steering method that modulates behavior in real time, without fine-tuning, reward signals, or environment interaction. We evaluate this method on two recent open-source VLAs, Pi0 and OpenVLA, and demonstrate zero-shot behavioral control in simulation (LIBERO) and on a physical robot (UR5). This work demonstrates that interpretable components of embodied VLAs can be systematically harnessed for control - establishing a new paradigm for transparent and steerable foundation models in robotics.

We introduce GIER (Gap-driven Iterative Enhancement of Responses), a general framework for improving large language model (LLM) outputs through self-reflection and revision based on conceptual quality criteria. Unlike prompting strategies that rely on demonstrations, examples, or chain-of-thought templates, GIER utilizes natural language descriptions of reasoning gaps, and prompts a model to iteratively critique and refine its own outputs to better satisfy these criteria. Across three reasoning-intensive tasks (SciFact, PrivacyQA, and e-SNLI) and four LLMs (GPT-4.1, GPT-4o Mini, Gemini 1.5 Pro, and Llama 3.3 70B), GIER improves rationale quality, grounding, and reasoning alignment without degrading task accuracy. Our analysis demonstrates that models can not only interpret abstract conceptual gaps but also translate them into concrete reasoning improvements.

The development of alignment and reasoning capabilities in large language models has seen remarkable progress through two paradigms: instruction tuning and reinforcement learning from human feedback (RLHF) alignment paradigm, and distillation-based reasoning fine-tuning paradigm. While both approaches prove effective independently, the third paradigm of applying RLHF to distillation-trained models presents significant challenges. Our investigation reveals two critical phenomena that emerge in this paradigm: Sequence Length Collapse, where language generation dramatically reduces during early RLHF training, and the Reward Hockey Stick Curve, featuring severe reward score drops followed by gradual recovery. These instabilities fundamentally compromise the model's alignment and reasoning capabilities. To address these challenges, we propose Balanced Actor Initialization (BAI), a two-stage weighted model merging approach. BAI first merges instruction-following and distillation-based reasoning fine-tuned models, then further combines this intermediate model with the pretrained model to preserve foundational knowledge. Through comprehensive experiments across diverse benchmarks and detailed analysis of training experiments, we demonstrate that BAI resolves Sequence Length Collapse, mitigates the Reward Hockey Stick Curve, and enables continuous sequence length improvement during training. Additionally, our analysis reveals that balanced merging ratios achieve optimal trade-offs between training stability and reasoning capability preservation. Our work provides the effective solution for stable training in this third paradigm, enabling more capable reasoning models that combine distillation efficiency with RLHF alignment.

Modern urban spaces are equipped with an increasingly diverse set of sensors, all producing an abundance of multimodal data. Such multimodal data can be used to identify and reason about important incidents occurring in urban landscapes, such as major emergencies, cultural and social events, as well as natural disasters. However, such data may be fragmented over several sources and difficult to integrate due to the reliance on human-driven reasoning for identifying relationships between the multimodal data corresponding to an incident, as well as understanding the different components which define an incident. Such relationships and components are critical to identifying the causes of such incidents, as well as producing forecasting the scale and intensity of future incidents as they begin to develop. In this work, we create SIGMUS, a system for Semantic Integration for Knowledge Graphs in Multimodal Urban Spaces. SIGMUS uses Large Language Models (LLMs) to produce the necessary world knowledge for identifying relationships between incidents occurring in urban spaces and data from different modalities, allowing us to organize evidence and observations relevant to an incident without relying and human-encoded rules for relating multimodal sensory data with incidents. This organized knowledge is represented as a knowledge graph, organizing incidents, observations, and much more. We find that our system is able to produce reasonable connections between 5 different data sources (new article text, CCTV images, air quality, weather, and traffic measurements) and relevant incidents occurring at the same time and location.