This paper revisits the implementation of $\textbf{L}$oad-$\textbf{b}$alancing $\textbf{L}$oss (LBL) when training Mixture-of-Experts (MoEs) models. Specifically, LBL for MoEs is defined as $N_E \sum_{i=1}^{N_E} f_i p_i$, where $N_E$ is the total number of experts, $f_i$ represents the frequency of expert $i$ being selected, and $p_i$ denotes the average gating score of the expert $i$. Existing MoE training frameworks usually employ the parallel training strategy so that $f_i$ and the LBL are calculated within a $\textbf{micro-batch}$ and then averaged across parallel groups. In essence, a micro-batch for training billion-scale LLMs normally contains very few sequences. So, the micro-batch LBL is almost at the sequence level, and the router is pushed to distribute the token evenly within each sequence. Under this strict constraint, even tokens from a domain-specific sequence ($\textit{e.g.}$, code) are uniformly routed to all experts, thereby inhibiting expert specialization. In this work, we propose calculating LBL using a $\textbf{global-batch}$ to loose this constraint. Because a global-batch contains much more diverse sequences than a micro-batch, which will encourage load balance at the corpus level. Specifically, we introduce an extra communication step to synchronize $f_i$ across micro-batches and then use it to calculate the LBL. Through experiments on training MoEs-based LLMs (up to $\textbf{42.8B}$ total parameters and $\textbf{400B}$ tokens), we surprisingly find that the global-batch LBL strategy yields excellent performance gains in both pre-training perplexity and downstream tasks. Our analysis reveals that the global-batch LBL also greatly improves the domain specialization of MoE experts.

Printed Electronics (PE) technology has emerged as a promising alternative to silicon-based computing. It offers attractive properties such as on-demand ultra-low-cost fabrication, mechanical flexibility, and conformality. However, PE are governed by large feature sizes, prohibiting the realization of complex printed Machine Learning (ML) classifiers. Leveraging PE's ultra-low non-recurring engineering and fabrication costs, designers can fully customize hardware to a specific ML model and dataset, significantly reducing circuit complexity. Despite significant advancements, state-of-the-art solutions achieve area efficiency at the expense of considerable accuracy loss. Our work mitigates this by designing area- and power-efficient printed ML classifiers with little to no accuracy degradation. Specifically, we introduce the first sequential Support Vector Machine (SVM) classifiers, exploiting the hardware efficiency of bespoke control and storage units and a single Multiply-Accumulate compute engine. Our SVMs yield on average 6x lower area and 4.6% higher accuracy compared to the printed state of the art.

Understanding time series data is crucial for multiple real-world applications. While large language models (LLMs) show promise in time series tasks, current approaches often rely on numerical data alone, overlooking the multimodal nature of time-dependent information, such as textual descriptions, visual data, and audio signals. Moreover, these methods underutilize LLMs' reasoning capabilities, limiting the analysis to surface-level interpretations instead of deeper temporal and multimodal reasoning. In this position paper, we argue that multimodal LLMs (MLLMs) can enable more powerful and flexible reasoning for time series analysis, enhancing decision-making and real-world applications. We call on researchers and practitioners to leverage this potential by developing strategies that prioritize trust, interpretability, and robust reasoning in MLLMs. Lastly, we highlight key research directions, including novel reasoning paradigms, architectural innovations, and domain-specific applications, to advance time series reasoning with MLLMs.

With the presence of robots increasing in the society, the need for interacting with robots is becoming necessary. The field of Human-Robot Interaction (HRI) has emerged important since more repetitive and tiresome jobs are being done by robots. In the recent times, the field of soft robotics has seen a boom in the field of research and commercialization. The Industry 5.0 focuses on human robot collaboration which also spurs the field of soft robotics. However the HRI for soft robotics is still in the nascent stage. In this work we review and then discuss how HRI is done for soft robots. We first discuss the control, design, materials and manufacturing of soft robots. This will provide an understanding of what is being interacted with. Then we discuss about the various input and output modalities that are used in HRI. The applications where the HRI for soft robots are found in the literature are discussed in detail. Then the limitations of HRI for soft robots and various research opportunities that exist in this field are discussed in detail. It is concluded that there is a huge scope for development for HRI for soft robots.

The development of biological data analysis tools and large language models (LLMs) has opened up new possibilities for utilizing AI in plant science research, with the potential to contribute significantly to knowledge integration and research gap identification. Nonetheless, current LLMs struggle to handle complex biological data and theoretical models in photosynthesis research and often fail to provide accurate scientific contexts. Therefore, this study proposed a photosynthesis research assistant (PRAG) based on OpenAI's GPT-4o with retrieval-augmented generation (RAG) techniques and prompt optimization. Vector databases and an automated feedback loop were used in the prompt optimization process to enhance the accuracy and relevance of the responses to photosynthesis-related queries. PRAG showed an average improvement of 8.7% across five metrics related to scientific writing, with a 25.4% increase in source transparency. Additionally, its scientific depth and domain coverage were comparable to those of photosynthesis research papers. A knowledge graph was used to structure PRAG's responses with papers within and outside the database, which allowed PRAG to match key entities with 63% and 39.5% of the database and test papers, respectively. PRAG can be applied for photosynthesis research and broader plant science domains, paving the way for more in-depth data analysis and predictive capabilities.

AI for PDEs has garnered significant attention, particularly Physics-Informed Neural Networks (PINNs). However, PINNs are typically limited to solving specific problems, and any changes in problem conditions necessitate retraining. Therefore, we explore the generalization capability of transfer learning in the strong and energy form of PINNs across different boundary conditions, materials, and geometries. The transfer learning methods we employ include full finetuning, lightweight finetuning, and Low-Rank Adaptation (LoRA). The results demonstrate that full finetuning and LoRA can significantly improve convergence speed while providing a slight enhancement in accuracy.

This paper addresses the critical disconnect between prediction and decision quality in portfolio optimization by integrating Large Language Models (LLMs) with decision-focused learning. We demonstrate both theoretically and empirically that minimizing the prediction error alone leads to suboptimal portfolio decisions. We aim to exploit the representational power of LLMs for investment decisions. An attention mechanism processes asset relationships, temporal dependencies, and macro variables, which are then directly integrated into a portfolio optimization layer. This enables the model to capture complex market dynamics and align predictions with the decision objectives. Extensive experiments on S\&P100 and DOW30 datasets show that our model consistently outperforms state-of-the-art deep learning models. In addition, gradient-based analyses show that our model prioritizes the assets most crucial to decision making, thus mitigating the effects of prediction errors on portfolio performance. These findings underscore the value of integrating decision objectives into predictions for more robust and context-aware portfolio management.

Physics-informed neural networks (PINNs) face significant challenges in modeling multi-frequency wavefields in complex velocity models due to their slow convergence, difficulty in representing high-frequency details, and lack of generalization to varying frequencies and velocity scenarios. To address these issues, we propose Meta-LRPINN, a novel framework that combines low-rank parameterization using singular value decomposition (SVD) with meta-learning and frequency embedding. Specifically, we decompose the weights of PINN's hidden layers using SVD and introduce an innovative frequency embedding hypernetwork (FEH) that links input frequencies with the singular values, enabling efficient and frequency-adaptive wavefield representation. Meta-learning is employed to provide robust initialization, improving optimization stability and reducing training time. Additionally, we implement adaptive rank reduction and FEH pruning during the meta-testing phase to further enhance efficiency. Numerical experiments, which are presented on multi-frequency scattered wavefields for different velocity models, demonstrate that Meta-LRPINN achieves much fast convergence speed and much high accuracy compared to baseline methods such as Meta-PINN and vanilla PINN. Also, the proposed framework shows strong generalization to out-of-distribution frequencies while maintaining computational efficiency. These results highlight the potential of our Meta-LRPINN for scalable and adaptable seismic wavefield modeling.

In this manuscript, we study the learning of deep attention neural networks, defined as the composition of multiple self-attention layers, with tied and low-rank weights. We first establish a mapping of such models to sequence multi-index models, a generalization of the widely studied multi-index model to sequential covariates, for which we establish a number of general results. In the context of Bayesian-optimal learning, in the limit of large dimension $D$ and commensurably large number of samples $N$, we derive a sharp asymptotic characterization of the optimal performance as well as the performance of the best-known polynomial-time algorithm for this setting --namely approximate message-passing--, and characterize sharp thresholds on the minimal sample complexity required for better-than-random prediction performance. Our analysis uncovers, in particular, how the different layers are learned sequentially. Finally, we discuss how this sequential learning can also be observed in a realistic setup.

Many approaches to multi-robot coordination are susceptible to failure due to communication loss and uncertainty in estimation. We present a real-time communication-free distributed algorithm for navigating robots to their desired goals certified by control barrier functions, that model and control the onboard sensing behavior to keep neighbors in the limited field of view for position estimation. The approach is robust to temporary tracking loss and directly synthesizes control in real time to stabilize visual contact through control Lyapunov-barrier functions. The main contributions of this paper are a continuous-time robust trajectory generation and control method certified by control barrier functions for distributed multi-robot systems and a discrete optimization procedure, namely, MPC-CBF, to approximate the certified controller. In addition, we propose a linear surrogate of high-order control barrier function constraints and use sequential quadratic programming to solve MPC-CBF efficiently. We demonstrate results in simulation with 10 robots and physical experiments with 2 custom-built UAVs. To the best of our knowledge, this work is the first of its kind to generate a robust continuous-time trajectory and controller concurrently, certified by control barrier functions utilizing piecewise splines.