HealthBench consists of 5,000 multi-turn conversations between a model and an individual user or healthcare professional. Responses are evaluated using conversation-specific rubrics created by 262 physicians. Unlike previous multiple-choice or short-answer benchmarks, Health-Bench enables realistic, open-ended evaluation through 48,562 unique rubric criteria spanning several health contexts (e.g., emergencies, transforming clinical data, global health) and behavioral dimensions (e.g., accuracy, instruction following, communication). HealthBench performance over the last two years reflects steady initial progress (compare GPT-3.5 Turbo's 16% to GPT-4o's 32%) and more rapid recent improvements (o3 scores 60%). Smaller models have especially improved: GPT-4.1 nano outperforms GPT-4o and is 25 times cheaper. We additionally release two HealthBench variations: HealthBench Consensus, which includes 34 particularly important dimensions of model behavior validated via physician consensus, and HealthBench Hard, where the current top score is 32%. We hope that HealthBench grounds progress towards model development and applications that benefit human health.

Attention-based architectures have achieved superior performance in multivariate time series forecasting but are computationally expensive. Techniques such as patching and adaptive masking have been developed to reduce their sizes and latencies. In this work, we propose a structured pruning method, SPAT ($\textbf{S}$ensitivity $\textbf{P}$runer for $\textbf{At}$tention), which selectively removes redundant attention mechanisms and yields highly effective models. Different from previous approaches, SPAT aims to remove the entire attention module, which reduces the risk of overfitting and enables speed-up without demanding specialized hardware. We propose a dynamic sensitivity metric, $\textbf{S}$ensitivity $\textbf{E}$nhanced $\textbf{N}$ormalized $\textbf{D}$ispersion (SEND) that measures the importance of each attention module during the pre-training phase. Experiments on multivariate datasets demonstrate that SPAT-pruned models achieve reductions of 2.842% in MSE, 1.996% in MAE, and 35.274% in FLOPs. Furthermore, SPAT-pruned models outperform existing lightweight, Mamba-based and LLM-based SOTA methods in both standard and zero-shot inference, highlighting the importance of retaining only the most effective attention mechanisms. We have made our code publicly available https://anonymous.4open.science/r/SPAT-6042.

As large-scale social data explode and machine-learning methods evolve, scholars of entrepreneurship and innovation face new research opportunities but also unique challenges. This chapter discusses the difficulties of leveraging large-scale data to identify technological and commercial novelty, document new venture origins, and forecast competition between new technologies and commercial forms. It suggests how scholars can take advantage of new text, network, image, audio, and video data in two distinct ways that advance innovation and entrepreneurship research. First, machine-learning models, combined with large-scale data, enable the construction of precision measurements that function as system-level observatories of innovation and entrepreneurship across human societies. Second, new artificial intelligence models fueled by big data generate'digital doubles' of technology and business, forming laboratories for virtual experimentation about innovation and entrepreneurship processes and policies. The chapter argues for the advancement of theory development and testing in entrepreneurship and innovation by coupling big data with big models. Key words: Entrepreneurship, venture funding, creative destruction, big data, digital doubles, embeddings, virtual experiment, artificial intelligence (AI), large language models (LLMs), deep neural networks (DNNs).

We present MegaBeam-Mistral-7B, a language model that supports 512K-token context length. Our work addresses practical limitations in long-context training, supporting real-world tasks such as compliance monitoring and verification. Evaluated on three long-context benchmarks, our 7B-parameter model demonstrates superior in-context learning performance on HELMET and robust retrieval and tracing capability on RULER. It is currently the only open model to achieve competitive long-range reasoning on BABILong at 512K context length without RAG or targeted fine-tuning. Released as fully open source under the Apache 2.0 license, the model has been downloaded over 100,000 times on Hugging Face. Model available at: https://huggingface.co/aws-prototyping/MegaBeam-Mistral-7B-512k

Retrieval-Augmented Generation (RAG) is a cornerstone of modern question answering (QA) systems, enabling grounded answers based on external knowledge. Although recent progress has been driven by open-domain datasets, enterprise QA systems need datasets that mirror the concrete, domain-specific issues users raise in day-to-day support scenarios. Critically, evaluating end-to-end RAG systems requires benchmarks comprising not only question--answer pairs but also the specific knowledge base (KB) snapshot from which answers were derived. To address this need, we introduce WixQA, a benchmark suite featuring QA datasets precisely grounded in the released KB corpus, enabling holistic evaluation of retrieval and generation components. WixQA includes three distinct QA datasets derived from Wix.com customer support interactions and grounded in a snapshot of the public Wix Help Center KB: (i) WixQA-ExpertWritten, 200 real user queries with expert-authored, multi-step answers; (ii) WixQA-Simulated, 200 expert-validated QA pairs distilled from user dialogues; and (iii) WixQA-Synthetic, 6,222 LLM-generated QA pairs, with one pair systematically derived from each article in the knowledge base. We release the KB snapshot alongside the datasets under MIT license and provide comprehensive baseline results, forming a unique benchmark for evaluating enterprise RAG systems in realistic enterprise environments.

Advancements in artificial intelligence (AI) are transforming pathology by integrat-ing large language models (LLMs) with retrieval-augmented generation (RAG) and domain-specific foundation models. This study explores the application of RAG-enhanced LLMs coupled with pathology foundation models for thyroid cytology diagnosis, addressing challenges in cytological interpretation, standardization, and diagnostic accuracy. By leveraging a curated knowledge base, RAG facilitates dy-namic retrieval of relevant case studies, diagnostic criteria, and expert interpreta-tion, improving the contextual understanding of LLMs. Meanwhile, pathology foun-dation models, trained on high-resolution pathology images, refine feature extrac-tion and classification capabilities. The fusion of these AI-driven approaches en-hances diagnostic consistency, reduces variability, and supports pathologists in dis-tinguishing benign from malignant thyroid lesions. Our results demonstrate that integrating RAG with pathology-specific LLMs significantly improves diagnostic efficiency and interpretability, paving the way for AI-assisted thyroid cytopathology, with foundation model UNI achieving AUC 0.73-0.93 for correct prediction of surgi-cal pathology diagnosis from thyroid cytology samples.

GPT has become nearly synonymous with large language models (LLMs), an increasingly popular term in AIED proceedings. A simple keyword-based search reveals that 61% of the 76 long and short papers presented at AIED 2024 describe novel solutions using LLMs to address some of the long-standing challenges in education, and 43% specifically mention GPT. Although LLMs pioneered by GPT create exciting opportunities to strengthen the impact of AI on education, we argue that the field's predominant focus on GPT and other resource-intensive LLMs (with more than 10B parameters) risks neglecting the potential impact that small language models (SLMs) can make in providing resource-constrained institutions with equitable and affordable access to high-quality AI tools. Supported by positive results on knowledge component (KC) discovery, a critical challenge in AIED, we demonstrate that SLMs such as Phi-2 can produce an effective solution without elaborate prompting strategies. Hence, we call for more attention to developing SLM-based AIED approaches.

Grasping unknown objects in unstructured environments remains a fundamental challenge in robotics, requiring both semantic understanding and spatial reasoning. Existing methods often rely on dense training datasets or explicit geometric modeling, limiting their scalability to real-world tasks. Recent advances in Large Multimodal Models (LMMs) offer new possibilities for integrating vision and language understanding, but their application to autonomous robotic grasping remains largely unexplored. We present ORACLE-Grasp, a zero-shot framework that leverages LMMs as semantic oracles to guide grasp selection without requiring additional training or human input. The system formulates grasp prediction as a structured, iterative decision process, using dual-prompt tool calling to first extract high-level object context and then select task-relevant grasp regions. By discretizing the image space and reasoning over candidate areas, ORACLE-Grasp mitigates the spatial imprecision common in LMMs and produces human-like, task-driven grasp suggestions. Early stopping and depth-based refinement steps further enhance efficiency and physical grasp reliability. Experiments demonstrate that the predicted grasps achieve low positional and orientation errors relative to human-annotated ground truth and lead to high success rates in real-world pick up tasks. These results highlight the potential of combining language-driven reasoning with lightweight vision techniques to enable robust, autonomous grasping without task-specific datasets or retraining.

Recent advancements in integrating tactile sensing with vision-language models (VLMs) have demonstrated remarkable potential for robotic multimodal perception. However, existing tactile descriptions remain limited to superficial attributes like texture, neglecting critical contact states essential for robotic manipulation. To bridge this gap, we propose CLTP, an intuitive and effective language tactile pretraining framework that aligns tactile 3D point clouds with natural language in various contact scenarios, thus enabling contact-state-aware tactile language understanding for contact-rich manipulation tasks. We first collect a novel dataset of 50k+ tactile 3D point cloud-language pairs, where descriptions explicitly capture multidimensional contact states (e.g., contact location, shape, and force) from the tactile sensor's perspective. CLTP leverages a pre-aligned and frozen vision-language feature space to bridge holistic textual and tactile modalities. Experiments validate its superiority in three downstream tasks: zero-shot 3D classification, contact state classification, and tactile 3D large language model (LLM) interaction. To the best of our knowledge, this is the first study to align tactile and language representations from the contact state perspective for manipulation tasks, providing great potential for tactile-language-action model learning. Code and datasets are open-sourced at https://sites.google.com/view/cltp/.

--Neighborhood environments include physical and environmental conditions such as housing quality, roads, and sidewalks, which significantly influence human health and wellbeing. Traditional methods for assessing these environments, including field surveys and geographic information systems (GIS), are resource-intensive and challenging to evaluate neighborhood environments at scale. Although machine learning offers potential for automated analysis, the laborious process of labeling training data and the lack of accessible models hinder scalability. This study explores the feasibility of large language models (LLMs) such as ChatGPT and Gemini as tools for decoding neighborhood environments (e.g., sidewalk and powerline) at scale. We train a robust YOLOv11-based model, which achieves an average accuracy of 99.13% in detecting six environmental indicators, including streetlight, sidewalk, powerline, apartment, single-lane road, and multilane road. We then evaluate four LLMs, including ChatGPT, Gemini, Claude, and Grok, to assess their feasibility, robustness, and limitations in identifying these indicators, with a focus on the impact of prompting strategies and fine-tuning. We apply majority voting with the top three LLMs to achieve over 88% accuracy, which demonstrates LLMs could be a useful tool to decode the neighborhood environment without any training effort. Neighborhood environments refer to the community where people live and participate in daily life, including its physical and environmental conditions, which play a critical role in shaping human health, behavior, and quality of life [1]- [3]. Those environmental indicators include housing quality, streetlights, parks, sidewalks, green space, power lines, etc. Research studies have shown the impact of neighborhood environments on health outcomes (e.g., obesity, diabetes, and mortality rates) [4], [5] and well-being factors (e.g., physical activity and access to nutritious foods) [4], [6].