--Compound AI (cAI) systems chain multiple AI models to solve complex problems. Deploying cAI services on mobile edge platforms poses a significant challenge in scheduling concurrent DNN-transformer inference tasks, which arrive dynamically in an unknown sequence. Existing mobile edge AI inference strategies manage multi-DNN or transformer-only workloads, relying on design-time profiling, and cannot handle concurrent inference of DNNs and transformers required by cAI systems. In this work, we address the challenge of scheduling cAI systems on heterogeneous mobile edge platforms. We present Twill, a run-time framework to handle concurrent inference requests of cAI workloads through task affinity-aware cluster mapping and migration, priority-aware task freezing/unfreezing, and Dynamic V oltage/Frequency Scaling (DVFS), while minimizing inference latency within power budgets. We implement and deploy our Twill framework on the Nvidia Jetson Orin NX platform. We evaluate Twill against state-of-the-art edge AI inference techniques over contemporary DNNs and LLMs, reducing inference latency by 54% on average, while honoring power budgets. AI applications are rapidly evolving from monolithic models towards Compound Artificial Intelligence (cAI) systems, which integrate multiple task-specific models and components to solve complex problems [1]-[3]. Emerging cAI systems combine Large Language Models (LLMs) with Deep Neural Networks (DNNs) for providing novel services such as conversational language agents [2]-[5], augmented and virtual reality (AR/VR) gear, and interactive autonomous vehicles [6]. In this example, DNN models ( D1: VGG-19 and D2: ResNet-152) are used for image classification, and object detection, transformer models ( T1: Bert-base and T2: Bert-large) are used for text summarizing and classification, and generative transformers ( T3: OPT-350M and LLM: Deepseek-R1) are used for reasoning and report generation. Each model is responsible for extracting key features from the given input and sending the output to the subsequent models to perform collaborative tasks. T1, D1, and D2 are exclusive inference tasks that can run simultaneously, while T2, T3, and LLM are dependent on the outputs of other models. We deployed the exemplar cAI system on the Nvidia Jetson Orin NX platform.

The aim of this paper is to review the use of GenAI in scientometrics, and to begin a debate on the broader implications for the field. First, we provide an introduction on GenAI's generative and probabilistic nature as rooted in distributional linguistics. And we relate this to the debate on the extent to which GenAI might be able to mimic human 'reasoning'. Second, we leverage this distinction for a critical engagement with recent experiments using GenAI in scientometrics, including topic labelling, the analysis of citation contexts, predictive applications, scholars' profiling, and research assessment. GenAI shows promise in tasks where language generation dominates, such as labelling, but faces limitations in tasks that require stable semantics, pragmatic reasoning, or structured domain knowledge. However, these results might become quickly outdated. Our recommendation is, therefore, to always strive to systematically compare the performance of different GenAI models for specific tasks. Third, we inquire whether, by generating large amounts of scientific language, GenAI might have a fundamental impact on our field by affecting textual characteristics used to measure science, such as authors, words, and references. We argue that careful empirical work and theoretical reflection will be essential to remain capable of interpreting the evolving patterns of knowledge production.

Topic model and document-clustering evaluations either use automated metrics that align poorly with human preferences or require expert labels that are intractable to scale. We design a scalable human evaluation protocol and a corresponding automated approximation that reflect practitioners' real-world usage of models. Annotators -- or an LLM-based proxy -- review text items assigned to a topic or cluster, infer a category for the group, then apply that category to other documents. Using this protocol, we collect extensive crowdworker annotations of outputs from a diverse set of topic models on two datasets. We then use these annotations to validate automated proxies, finding that the best LLM proxies are statistically indistinguishable from a human annotator and can therefore serve as a reasonable substitute in automated evaluations. Package, web interface, and data are at https://github.com/ahoho/proxann

With the increasing interest in robotic synthesis in the context of organic chemistry, the automated extraction of chemical procedures from literature is critical. However, this task remains challenging due to the inherent ambiguity of chemical language and the high cost of human annotation required for developing reliable computer-aided extraction protocols. Here, we present ChemActor, a fully fine-tuned large language model (LLM), as a chemical executor to convert between unstructured experimental procedures and structured action sequences. We propose a sequential LLM-generated data framework to address the challenges of insufficient and low-quality annotated data. This framework integrates a data selection module that selects data based on distribution divergence, with a general-purpose LLM, to generate machine-executable actions from a single molecule input. Additionally, we introduce a novel multi-round LLMs circle review metric, which reflects the model's advanced understanding of chemical experimental procedures. Extensive experiments on reaction-to-description (R2D) and description-to-action (D2A) tasks demonstrate that ChemActor, augmented by LLM-generated data, achieves state-of-the-art performance, outperforming the baseline model by 10%. The code is available at: https://github.com/Zhanghahah/ChemActor.

Large language models (LLMs) are now widely accessible, reaching learners at all educational levels. This development has raised concerns that their use may circumvent essential learning processes and compromise the integrity of established assessment formats. In physics education, where problem solving plays a central role in instruction and assessment, it is therefore essential to understand the physics-specific problem-solving capabilities of LLMs. Such understanding is key to informing responsible and pedagogically sound approaches to integrating LLMs into instruction and assessment. This study therefore compares the problem-solving performance of a general-purpose LLM (GPT-4o, using varying prompting techniques) and a reasoning-optimized model (o1-preview) with that of participants of the German Physics Olympiad, based on a set of well-defined Olympiad problems. In addition to evaluating the correctness of the generated solutions, the study analyzes characteristic strengths and limitations of LLM-generated solutions. The findings of this study indicate that both tested LLMs (GPT-4o and o1-preview) demonstrate advanced problem-solving capabilities on Olympiad-type physics problems, on average outperforming the human participants. Prompting techniques had little effect on GPT-4o's performance, while o1-preview almost consistently outperformed both GPT-4o and the human benchmark. Based on these findings, the study discusses implications for the design of summative and formative assessment in physics education, including how to uphold assessment integrity and support students in critically engaging with LLMs.

The escalating computational costs of Large Language Model (LLM) inference have become a critical barrier to their widespread and sustainable deployment. While existing optimization strategies are effective, they are predominantly based on statistical heuristics or architectural modifications, lacking a guiding cognitive theory to manage the inference process itself. This paper aims to bridge this gap by introducing a novel paradigm: the Cognitive Load-Aware Inference (CLAI) framework, which operationalizes principles from Cognitive Load Theory (CLT) and neuroscience for LLM inference. We formalize the concepts of Intrinsic Cognitive Load, Extraneous Cognitive Load, and Germane Cognitive Load into quantifiable LLM metrics ($ICL_{LLM}$, $ECL_{LLM}$, and $GCL_{LLM}$), thereby reframing the inference process as a cognitive economics optimization problem: based on the intrinsic complexity of a problem ($ICL_{LLM}$), minimize wasteful computation ($ECL_{LLM}$), and strategically allocate the token budget to productive reasoning ($GCL_{LLM}$). We propose two implementation paths: CLAI-Prompt, a zero-shot method that guides a base LLM through cognitive control steps via a structured meta-prompt, and CLAI-Tune, a fine-tuned model that internalizes these principles for spontaneous cognitive economy. Across a range of benchmarks in complex reasoning, long-context question answering, and code generation, our methods achieve significant reductions in token consumption (up to 45\%) without sacrificing accuracy. Furthermore, CLAI-Tune exhibits an emergent ability to autonomously decompose difficult problems, a key characteristic of human expert cognition. This work demonstrates that by emulating the brain's resource management strategies, we can build more efficient, robust, and capable artificial intelligence systems.

In this paper, we introduce an innovative vector quantization based action tokenizer built upon the largest-scale action trajectory dataset to date, leveraging over 100 times more data than previous approaches. This extensive dataset enables our tokenizer to capture rich spatiotemporal dynamics, resulting in a model that not only accelerates inference but also generates smoother and more coherent action outputs. Once trained, the tokenizer can be seamlessly adapted to a wide range of downstream tasks in a zero-shot manner, from short-horizon reactive behaviors to long-horizon planning. A key finding of our work is that the domain gap between synthetic and real action trajectories is marginal, allowing us to effectively utilize a vast amount of synthetic data during training without compromising real-world performance. To validate our approach, we conducted extensive experiments in both simulated environments and on real robotic platforms. The results demonstrate that as the volume of synthetic trajectory data increases, the performance of our tokenizer on downstream tasks improves significantly-most notably, achieving up to a 30% higher success rate on two real-world tasks in long-horizon scenarios. These findings highlight the potential of our action tokenizer as a robust and scalable solution for real-time embodied intelligence systems, paving the way for more efficient and reliable robotic control in diverse application domains.Project website: https://xiaoxiao0406.github.io/vqvla.github.io

As large language models (LLMs) are increasingly used in high-stakes domains, accurately assessing their confidence is crucial. Humans typically express confidence through epistemic markers (e.g., "fairly confident") instead of numerical values. However, it remains unclear whether LLMs consistently use these markers to reflect their intrinsic confidence due to the difficulty of quantifying uncertainty associated with various markers. To address this gap, we first define marker confidence as the observed accuracy when a model employs an epistemic marker. We evaluate its stability across multiple question-answering datasets in both in-distribution and out-of-distribution settings for open-source and proprietary LLMs. Our results show that while markers generalize well within the same distribution, their confidence is inconsistent in out-of-distribution scenarios. These findings raise significant concerns about the reliability of epistemic markers for confidence estimation, underscoring the need for improved alignment between marker based confidence and actual model uncertainty. Our code is available at https://github.com/HKUST-KnowComp/MarCon.

The deployment of large language models (LLMs) is largely hindered by their large number of parameters. Structural pruning has emerged as a promising solution. Prior structured pruning methods directly remove unimportant parameters based on certain metrics, which often causes knowledge loss and necessitates extensive retraining. To overcome this, we introduce a novel pruning method TRSP: Two-Stage Regularization-Based Structured Pruning for LLMs. Specifically, we multiply the output of each transformer layer by an initial learnable weight and iteratively learn these weights by adding their $\ell_1$-norm as a regularization term to the loss function, serving as the first-stage regularization. Subsequently, we apply additional regularization to the difference between the output and input of layers with smaller weights, encouraging the shift of knowledge to the preserved layers. This serves as the second-stage regularization. TRSP retains more knowledge and better preserves model performance than direct parameter elimination. Through extensive experimentation we show that TRSP outperforms strong layer-wise structured pruning methods without requiring retraining. As a layer-wise pruning method, it delivers notable end-to-end acceleration, making it a promising solution for efficient LLM deployment.

Cross-cultural research in perception and cognition has shown that individuals from different cultural backgrounds process visual information in distinct ways. East Asians, for example, tend to adopt a holistic perspective, attending to contextual relationships, whereas Westerners often employ an analytical approach, focusing on individual objects and their attributes. In this study, we investigate whether Vision-Language Models (VLMs) trained predominantly on different languages, specifically Japanese and English, exhibit similar culturally grounded attentional patterns. Using comparative analysis of image descriptions, we examine whether these models reflect differences in holistic versus analytic tendencies. Our findings suggest that VLMs not only internalize the structural properties of language but also reproduce cultural behaviors embedded in the training data, indicating that cultural cognition may implicitly shape model outputs.