Operator fusion has become a key optimization for deep learning, which combines multiple deep learning operators to improve data reuse and reduce global memory transfers. However, existing tensor compilers struggle to fuse complex reduction computations involving loop-carried dependencies, such as attention mechanisms. The paper introduces Neptune, a tensor compiler for advanced operator fusion for sequences of reduction operators. Neptune presents a new approach for advanced operator fusion, which intentionally breaks some existing dependencies and compensates by constructing algebraic correction expressions that allow the kernel to produce the correct result. On ten attention-based benchmarks, Neptune, starting from simple attention code and a high-level scheduling template, outperforms existing compilers like Triton, TVM, and FlexAttention, including Triton-based implementations of FlashAttention. Across four different GPU architectures from NVIDIA and AMD, Neptune-generated kernels have average speedup of $1.35\times$ over the next best alternative, demonstrating its effectiveness for deep learning workloads.

We propose a small-scale language model (SLM) based agent architecture, Planner-Caller-Generator (P-C-G), optimized for Korean tool use. P-C-G separates planning, calling, and generation by role: the Planner produces an initial batch plan with limited on-demand replanning; the Caller returns a normalized call object after joint schema-value validation; and the Generator integrates tool outputs to produce the final answer. We apply a Korean-first value policy to reduce execution failures caused by frequent Korean-to-English code switching in Korean settings. Evaluation assumes Korean queries and Korean tool/parameter specifications; it covers single-chain, multi-chain, missing-parameters, and missing-functions scenarios, and is conducted via an LLM-as-a-Judge protocol averaged over five runs under a unified I/O interface. Results show that P-C-G delivers competitive tool-use accuracy and end-to-end quality while reducing tokens and maintaining acceptable latency, indicating that role-specialized SLMs are a cost-effective alternative for Korean tool-use agents.

Large Language Models (LLMs) have shown impressive capabilities in many scenarios, but their performance depends, in part, on the choice of prompt. Past research has focused on optimizing prompts specific to a task. However, much less attention has been given to optimizing the general instructions included in a prompt, known as a system prompt. To address this gap, we propose SPRIG, an edit-based genetic algorithm that iteratively constructs prompts from prespecified components to maximize the model's performance in general scenarios. We evaluate the performance of system prompts on a collection of 47 different types of tasks to ensure generalizability. Our study finds that a single optimized system prompt performs on par with task prompts optimized for each individual task. Moreover, combining system and task-level optimizations leads to further improvement, which showcases their complementary nature. Experiments also reveal that the optimized system prompts generalize effectively across model families, parameter sizes, and languages. This study provides insights into the role of system-level instructions in maximizing LLM potential.

The integration of deep learning, particularly AI-Generated Content, with high-quality data derived from ab initio calculations has emerged as a promising avenue for transforming the landscape of scientific research. However, the challenge of designing molecular drugs or materials that incorporate multi-modality prior knowledge remains a critical and complex undertaking. Specifically, achieving a practical molecular design necessitates not only meeting the diversity requirements but also addressing structural and textural constraints with various symmetries outlined by domain experts. In this article, we present an innovative approach to tackle this inverse design problem by formulating it as a multi-modality guidance generation/optimization task. Our proposed solution involves a textural-structure alignment symmetric diffusion framework for the implementation of molecular generation/optimization tasks, namely 3DToMolo. 3DToMolo aims to harmonize diverse modalities, aligning them seamlessly to produce molecular structures adhere to specified symmetric structural and textural constraints by experts in the field. Experimental trials across three guidance generation settings have shown a superior hit generation performance compared to state-of-the-art methodologies. Moreover, 3DToMolo demonstrates the capability to generate novel molecules, incorporating specified target substructures, without the need for prior knowledge. This work not only holds general significance for the advancement of deep learning methodologies but also paves the way for a transformative shift in molecular design strategies. 3DToMolo creates opportunities for a more nuanced and effective exploration of the vast chemical space, opening new frontiers in the development of molecular entities with tailored properties and functionalities.

Understanding human intentions is critical for safe and effective human-robot collaboration. While state of the art methods for human goal prediction utilize learned models to account for the uncertainty of human motion data, that data is inherently stochastic and high variance, hindering those models' utility for interactions requiring coordination, including safety-critical or close-proximity tasks. Our key insight is that robot teammates can deliberately configure shared workspaces prior to interaction in order to reduce the variance in human motion, realizing classifier-agnostic improvements in goal prediction. In this work, we present an algorithmic approach for a robot to arrange physical objects and project "virtual obstacles" using augmented reality in shared human-robot workspaces, optimizing for human legibility over a given set of tasks. We compare our approach against other workspace arrangement strategies using two human-subjects studies, one in a virtual 2D navigation domain and the other in a live tabletop manipulation domain involving a robotic manipulator arm. We evaluate the accuracy of human motion prediction models learned from each condition, demonstrating that our workspace optimization technique with virtual obstacles leads to higher robot prediction accuracy using less training data.

Local explanation methods highlight the input tokens that have a considerable impact on the outcome of classifying the document at hand. For example, the Anchor algorithm applies a statistical analysis of the sensitivity of the classifier to changes in the token. Aggregating local explanations over a dataset provides a global explanation of the model. Such aggregation aims to detect words with the most impact, giving valuable insights about the model, like what it has learned in training and which adversarial examples expose its weaknesses. However, standard aggregation methods bear a high computational cost: a na\"ive implementation applies a costly algorithm to each token of each document, and hence, it is infeasible for a simple user running in the scope of a short analysis session. % We devise techniques for accelerating the global aggregation of the Anchor algorithm. Specifically, our goal is to compute a set of top-$k$ words with the highest global impact according to different aggregation functions. Some of our techniques are lossless and some are lossy. We show that for a very mild loss of quality, we are able to accelerate the computation by up to 30$\times$, reducing the computation from hours to minutes. We also devise and study a probabilistic model that accounts for noise in the Anchor algorithm and diminishes the bias toward words that are frequent yet low in impact.

In this paper, we present a new approach to mental state classification from EEG signals by combining signal processing techniques and machine learning (ML) algorithms. We evaluate the performance of the proposed method on a dataset of EEG recordings collected during a cognitive load task and compared it to other state-of-the-art methods. The results show that the proposed method achieves high accuracy in classifying mental states and outperforms state-of-the-art methods in terms of classification accuracy and computational efficiency.

Kitting refers to the task of preparing and grouping necessary parts and tools (or "kits") for assembly in a manufacturing environment. Automating this process simplifies the assembly task for human workers and improves efficiency. Existing automated kitting systems adhere to scripted instructions and predefined heuristics. However, given variability in the availability of parts and logistic delays, the inflexibility of existing systems can limit the overall efficiency of an assembly line. In this paper, we propose a bilevel optimization framework to enable a robot to perform task segmentation-based part selection, kit arrangement, and delivery scheduling to provide custom-tailored kits just in time - i.e., right when they are needed. We evaluate the proposed approach both through a human subjects study (n=18) involving the construction of a flat-pack furniture table and shop-flow simulation based on the data from the study. Our results show that the just-in-time kitting system is objectively more efficient, resilient to upstream shop flow delays, and subjectively more preferable as compared to baseline approaches of using kits defined by rigid task segmentation boundaries defined by the task graph itself or a single kit that includes all parts necessary to assemble a single unit.

Three new core algorithms are recently launched for the TBEA partnership project. This week, Matrix finished testing and launched three new core algorithms for the partnership project with TBEA. The new features include risky behaviour detection, equipment life and failure prediction, as well as AI adjustment of coal power generation efficiency. Matrix has successfully developed and launched all core functional modules for the project, and what follows will be the integration and optimization of the entire structure. The TBEA Unmanned Mine plays a crucial role in the energy section of the Belt and Road Initiative.

Inertial-sensor-based attitude estimation is a crucial technology in various applications, from human motion tracking to autonomous aerial and ground vehicles. Application scenarios differ in characteristics of the performed motion, presence of disturbances, and environmental conditions. Since state-of-the-art attitude estimators do not generalize well over these characteristics, their parameters must be tuned for the individual motion characteristics and circumstances. We propose RIANN, a real-time-capable neural network for robust IMU-based attitude estimation, which generalizes well across different motion dynamics, environments, and sampling rates, without the need for application-specific adaptations. We exploit two publicly available datasets for the method development and the training, and we add four completely different datasets for evaluation of the trained neural network in three different test scenarios with varying practical relevance. Results show that RIANN performs at least as well as state-of-the-art attitude estimation filters and outperforms them in several cases, even if the filter is tuned on the very same test dataset itself while RIANN has never seen data from that dataset, from the specific application, the same sensor hardware, or the same sampling frequency before. RIANN is expected to enable plug-and-play solutions in numerous applications, especially when accuracy is crucial but no ground-truth data is available for tuning or when motion and disturbance characteristics are uncertain. We made RIANN publicly available.