To reveal when a large language model (LLM) is uncertain about a response, uncertainty quantification commonly produces percentage numbers along with the output. But is this all we can do? We argue that in the output space of LLMs, the space of strings, exist strings expressive enough to summarize the distribution over output strings the LLM deems possible. We lay a foundation for this new avenue of uncertainty explication and present SelfReflect, a theoretically-motivated metric to assess how faithfully a string summarizes an LLM's internal answer distribution. We show that SelfReflect is able to discriminate even subtle differences of candidate summary strings and that it aligns with human judgement, outperforming alternative metrics such as LLM judges and embedding comparisons. With SelfReflect, we investigate a number of self-summarization methods and find that even state-of-the-art reasoning models struggle to explicate their internal uncertainty. But we find that faithful summarizations can be generated by sampling and summarizing. To support the development of this universal form of LLM uncertainties, we publish our metric at https://github.com/apple/ml-selfreflect

As AI integrates in various types of human writing, calls for transparency around AI assistance are growing. However, if transparency operates on uneven ground and certain identity groups bear a heavier cost for being honest, then the burden of openness becomes asymmetrical. This study investigates how AI disclosure statement affects perceptions of writing quality, and whether these effects vary by the author's race and gender. Through a large-scale controlled experiment, both human raters (n = 1,970) and LLM raters (n = 2,520) evaluated a single human-written news article while disclosure statements and author demographics were systematically varied. This approach reflects how both human and algorithmic decisions now influence access to opportunities (e.g., hiring, promotion) and social recognition (e.g., content recommendation algorithms). We find that both human and LLM raters consistently penalize disclosed AI use. However, only LLM raters exhibit demographic interaction effects: they favor articles attributed to women or Black authors when no disclosure is present. But these advantages disappear when AI assistance is revealed. These findings illuminate the complex relationships between AI disclosure and author identity, highlighting disparities between machine and human evaluation patterns.

Large language models (LLMs) have demonstrated impressive capabilities as autonomous agents with rapidly expanding applications in various domains. As these agents increasingly engage in socioeconomic interactions, identifying their potential for undesirable behavior becomes essential. In this work, we examine scenarios where they can choose to collude, defined as secretive cooperation that harms another party. To systematically study this, we investigate the behavior of LLM agents acting as sellers in simulated continuous double auction markets. Through a series of controlled experiments, we analyze how parameters such as the ability to communicate, choice of model, and presence of environmental pressures affect the stability and emergence of seller collusion. We find that direct seller communication increases collusive tendencies, the propensity to collude varies across models, and environmental pressures, such as oversight and urgency from authority figures, influence collusive behavior. Our findings highlight important economic and ethical considerations for the deployment of LLM-based market agents.

We present a new dataset for chart question answering (CQA) constructed from visualization notebooks. The dataset features real-world, multi-view charts paired with natural language questions grounded in analytical narratives. Unlike prior benchmarks, our data reflects ecologically valid reasoning workflows. Benchmarking state-of-the-art multimodal large language models reveals a significant performance gap, with GPT-4.1 achieving an accuracy of 69.3%, underscoring the challenges posed by this more authentic CQA setting.

Automatic text simplification (ATS) aims to enhance language accessibility for various target groups, particularly persons with intellectual disabilities. Recent advancements in generative AI, especially large language models (LLMs), have substantially improved the quality of machine-generated text simplifications, thereby mitigating information barriers for the target group. However, existing LLM-based ATS systems do not incorporate preference feedback on text simplifications during training, resulting in a lack of personalization tailored to the specific needs of target group representatives. In this work, we extend the standard supervised fine-tuning (SFT) approach for adapting LLM-based ATS models by leveraging a computationally efficient LLM alignment technique -- direct preference optimization (DPO). Specifically, we post-train LLM-based ATS models using human feedback collected from persons with intellectual disabilities, reflecting their preferences on paired text simplifications generated by mainstream LLMs. Furthermore, we propose a pipeline for developing personalized LLM-based ATS systems, encompassing data collection, model selection, SFT and DPO post-training, and evaluation. Our findings underscore the necessity of active participation of target group persons in designing personalized AI accessibility solutions aligned with human expectations. This work represents a step towards personalizing inclusive AI systems at the target-group level, incorporating insights not only from text simplification experts but also from target group persons themselves.

As large language models (LLMs) increasingly underpin technological advancements, the privacy of their training data emerges as a critical concern. Differential Privacy (DP) serves as a rigorous mechanism to protect this data, yet its integration via Differentially Private Stochastic Gradient Descent (DP-SGD) introduces substantial challenges, primarily due to the complexities of per-sample gradient clipping. Current explicit methods, such as Opacus, necessitate extensive storage for per-sample gradients, significantly inflating memory requirements. Conversely, implicit methods like GhostClip reduce storage needs by recalculating gradients multiple times, which leads to inefficiencies due to redundant computations. This paper introduces FlashDP, an innovative cache-friendly per-layer DP-SGD that consolidates necessary operations into a single task, calculating gradients only once in a fused manner. This approach not only diminishes memory movement by up to \textbf{50\%} but also cuts down redundant computations by \textbf{20\%}, compared to previous methods. Consequently, FlashDP does not increase memory demands and achieves a \textbf{90\%} throughput compared to the Non-DP method on a four-A100 system during the pre-training of the Llama-13B model, while maintaining parity with standard per-layer clipped DP-SGD in terms of accuracy. These advancements establish FlashDP as a pivotal development for efficient and privacy-preserving training of LLMs. FlashDP's code has been open-sourced in https://github.com/kaustpradalab/flashdp.

Patient cohort retrieval is a pivotal task in medical research and clinical practice, enabling the identification of specific patient groups from extensive electronic health records (EHRs). In this work, we address the challenge of cohort retrieval in the echocardiography domain by applying Dense Passage Retrieval (DPR), a prominent methodology in semantic search. We propose a systematic approach to transform an echocardiographic EHR dataset of unstructured nature into a Query-Passage dataset, framing the problem as a Cohort Retrieval task. Additionally, we design and implement evaluation metrics inspired by real-world clinical scenarios to rigorously test the models across diverse retrieval tasks. Furthermore, we present a custom-trained DPR embedding model that demonstrates superior performance compared to traditional and off-the-shelf SOTA methods.To our knowledge, this is the first work to apply DPR for patient cohort retrieval in the echocardiography domain, establishing a framework that can be adapted to other medical domains.

Cardiac biosignals, such as electrocardiograms (ECG) and photoplethysmograms (PPG), are of paramount importance for the diagnosis, prevention, and management of cardiovascular diseases, and have been extensively used in a variety of clinical tasks. Conventional deep learning approaches for analyzing these signals typically rely on homogeneous datasets and static bespoke models, limiting their robustness and generalizability across diverse clinical settings and acquisition protocols. In this study, we present a cardiac sensing foundation model (CSFM) that leverages advanced transformer architectures and a generative, masked pretraining strategy to learn unified representations from vast, heterogeneous health records. Our model is pretrained on an innovative multi-modal integration of data from multiple large-scale datasets (including MIMIC-III-WDB, MIMIC-IV-ECG, and CODE), comprising cardiac signals and the corresponding clinical or machine-generated text reports from approximately 1.7 million individuals. We demonstrate that the embeddings derived from our CSFM not only serve as effective feature extractors across diverse cardiac sensing scenarios, but also enable seamless transfer learning across varying input configurations and sensor modalities. Extensive evaluations across diagnostic tasks, demographic information recognition, vital sign measurement, clinical outcome prediction, and ECG question answering reveal that CSFM consistently outperforms traditional one-modal-one-task approaches. Notably, CSFM exhibits robust performance across multiple ECG lead configurations from standard 12-lead systems to single-lead setups, and in scenarios where only ECG, only PPG, or a combination thereof is available. These findings highlight the potential of CSFM as a versatile and scalable solution, for comprehensive cardiac monitoring.

The incessant advent of online services demands high speed and efficient recommender systems (ReS) that can maintain real-time performance along with processing very complex user-item interactions. The present study, therefore, considers computational bottlenecks involved in hybrid Graph Neural Network (GNN) and Large Language Model (LLM)-based ReS with the aim optimizing their inference latency and training efficiency. An extensive methodology was used: hybrid GNN-LLM integrated architecture-optimization strategies(quantization, LoRA, distillation)-hardware acceleration (FPGA, DeepSpeed)-all under R 4.4.2. Experimental improvements were significant, with the optimal Hybrid + FPGA + DeepSpeed configuration reaching 13.6% more accuracy (NDCG@10: 0.75) at 40-60ms of latency, while LoRA brought down training time by 66% (3.8 hours) in comparison to the non-optimized baseline. Irrespective of domain, such as accuracy or efficiency, it can be established that hardware-software co-design and parameter-efficient tuning permit hybrid models to outperform GNN or LLM approaches implemented independently. It recommends the use of FPGA as well as LoRA for real-time deployment. Future work should involve federated learning along with advanced fusion architectures for better scalability and privacy preservation. Thus, this research marks the fundamental groundwork concerning next-generation ReS balancing low-latency response with cutting-edge personalization.

With the development of generative artificial intelligence and instruction tuning techniques, multimodal large language models (MLLMs) have made impressive progress on general reasoning tasks. Benefiting from the chain-of-thought (CoT) methodology, MLLMs can solve the visual reasoning problem step-by-step. However, existing MLLMs still face significant challenges when applied to pathology visual reasoning tasks: (1) LLMs often underperforms because they lack domain-specific information, which can lead to model hallucinations. (2) The additional reasoning steps in CoT may introduce errors, leading to the divergence of answers. To address these limitations, we propose PathCoT, a novel zero-shot CoT prompting method which integrates the pathology expert-knowledge into the reasoning process of MLLMs and incorporates self-evaluation to mitigate divergence of answers. Specifically, PathCoT guides the MLLM with prior knowledge to perform as pathology experts, and provides comprehensive analysis of the image with their domain-specific knowledge. By incorporating the experts' knowledge, PathCoT can obtain the answers with CoT reasoning. Furthermore, PathCoT incorporates a self-evaluation step that assesses both the results generated directly by MLLMs and those derived through CoT, finally determining the reliable answer. The experimental results on the PathMMU dataset demonstrate the effectiveness of our method on pathology visual understanding and reasoning.