Week Ending 5.18.2025
RESEARCH WATCH: 5.18.2025
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenge
This paper provides a critical distinction between AI Agents (narrow, task-specific systems driven by LLMs/LIMs) and Agentic AI (systems with multi-agent collaboration and orchestrated autonomy). The authors present a structured conceptual taxonomy and application mapping while analyzing challenges unique to each paradigm. This work offers valuable insights for researchers and developers building robust AI systems, particularly in domains requiring different levels of automation. Applications range from customer support and data summarization for AI Agents to research automation and medical decision support for more complex Agentic AI systems.
Authors: Ranjan Sapkota, Konstantinos I. Roumeliotis, Manoj Karkee
Link: https://arxiv.org/abs/2505.10468v1
Date: 2025-05-15
Summary:
This study critically distinguishes between AI Agents and Agentic AI, offering a structured conceptual taxonomy, application mapping, and challenge analysis to clarify their divergent design philosophies and capabilities. We begin by outlining the search strategy and foundational definitions, characterizing AI Agents as modular systems driven by Large Language Models (LLMs) and Large Image Models (LIMs) for narrow, task-specific automation. Generative AI is positioned as a precursor, with AI Agents advancing through tool integration, prompt engineering, and reasoning enhancements. In contrast, Agentic AI systems represent a paradigmatic shift marked by multi-agent collaboration, dynamic task decomposition, persistent memory, and orchestrated autonomy. Through a sequential evaluation of architectural evolution, operational mechanisms, interaction styles, and autonomy levels, we present a comparative analysis across both paradigms. Application domains such as customer support, scheduling, and data summarization are contrasted with Agentic AI deployments in research automation, robotic coordination, and medical decision support. We further examine unique challenges in each paradigm including hallucination, brittleness, emergent behavior, and coordination failure and propose targeted solutions such as ReAct loops, RAG, orchestration layers, and causal modeling. This work aims to provide a definitive roadmap for developing robust, scalable, and explainable AI agent and Agentic AI-driven systems. >AI Agents, Agent-driven, Vision-Language-Models, Agentic AI Decision Support System, Agentic-AI Applications
--------------------------------------------------------------------------------------------------------
Two Minds Better Than One: Collaborative Reward Modeling for LLM Alignment
This innovative paper addresses a critical challenge in LLM alignment - reward misgeneralization caused by noisy human preferences. The authors introduce Collaborative Reward Modeling (CRM), a framework combining peer review and curriculum learning where two reward models train in parallel and assess each other's data selections. By filtering out potential noise and structuring preference data from easy to hard, CRM significantly improves generalization accuracy. This approach has promising applications for developing more robust and reliably aligned AI systems, particularly when training data contains significant noise or inconsistencies.
Authors: Jiazheng Zhang, Wenqing Jing, Zizhuo Zhang, Zhiheng Xi, Shihan Dou, Rongxiang Weng, Jiahuan Li, Jingang Wang, MingXu Cai, Shibo Hong, Tao Gui, Qi Zhang
Link: https://arxiv.org/abs/2505.10597v1
Date: 2025-05-15
Summary:
Reward models (RMs) are essential for aligning large language models (LLMs) with human values. However, noisy preferences in human feedback often lead to reward misgeneralization, where RMs overfit to spurious patterns and provide misleading signals during policy optimization. We systematically analyze the training dynamics of preference pairs and identify that noisy examples are harder to fit and introduce instability. Empirical evidence shows that LLMs optimized using reward models trained on full noisy datasets perform worse than those trained on filtered, high-quality preferences. To address this, we propose Collaborative Reward Modeling (CRM), an online framework that enhances robustness by combining peer review and curriculum learning. Two reward models are trained in parallel and assess each other's data selections to filter out potential noise. Curriculum learning structures the preference data from easy to hard, ensuring synchronized training and stable feedback. Extensive experiments demonstrate that CRM improves generalization, with up to 9.94 points of accuracy gain on RewardBench under 40 percent label noise. CRM is also compatible with implicit-reward alignment methods, offering a practical and versatile strategy for robust alignment.
--------------------------------------------------------------------------------------------------------
WhatsAI: Transforming Meta Ray-Bans into an Extensible Generative AI Platform for Accessibility
WhatsAI represents a groundbreaking accessibility initiative that transforms Meta Ray-Ban smart glasses into an extensible visual assistance platform for blind and visually impaired (BVI) users. The paper introduces a hackable template integrating with WhatsApp to enable real-time scene description, object detection, and OCR through standard machine learning and visual language models. By democratizing the development of wearable visual accessibility technologies, WhatsAI empowers BVI enthusiasts to create personalized solutions addressing specific accessibility challenges, potentially fostering a community-driven approach to visual accessibility innovation led by BVI hackers themselves.
Authors: Nasif Zaman, Venkatesh Potluri, Brandon Biggs, James M. Coughlan
Link: https://arxiv.org/abs/2505.09823v1
Date: 2025-05-14
Summary:
Multi-modal generative AI models integrated into wearable devices have shown significant promise in enhancing the accessibility of visual information for blind or visually impaired (BVI) individuals, as evidenced by the rapid uptake of Meta Ray-Bans among BVI users. However, the proprietary nature of these platforms hinders disability-led innovation of visual accessibility technologies. For instance, OpenAI showcased the potential of live, multi-modal AI as an accessibility resource in 2024, yet none of the presented applications have reached BVI users, despite the technology being available since then. To promote the democratization of visual access technology development, we introduce WhatsAI, a prototype extensible framework that empowers BVI enthusiasts to leverage Meta Ray-Bans to create personalized wearable visual accessibility technologies. Our system is the first to offer a fully hackable template that integrates with WhatsApp, facilitating robust Accessible Artificial Intelligence Implementations (AAII) that enable blind users to conduct essential visual assistance tasks, such as real-time scene description, object detection, and Optical Character Recognition (OCR), utilizing standard machine learning techniques and cutting-edge visual language models. The extensible nature of our framework aspires to cultivate a community-driven approach, led by BVI hackers and innovators to tackle the complex challenges associated with visual accessibility.
--------------------------------------------------------------------------------------------------------
Fully analytical propagator for lunar satellite orbits in closed form
This technical paper presents a breakthrough in satellite orbit calculation - a fully analytical propagator for lunar artificial satellites that matches the accuracy of more complex semi-analytical systems. The authors derive an approximate analytical solution using Hamiltonian normal form in closed form, enabling precise position and velocity predictions without intermediate numerical propagation. This innovation has significant applications in spacecraft mission planning, lunar exploration, and communication satellite deployment, potentially reducing computational requirements while maintaining accuracy for orbital predictions spanning several decades. The system's open-source implementation makes it readily available for both research and practical space applications.
Authors: Rita Mastroianni, Edoardo Legnaro, Christos Efthymiopoulos
Link: https://arxiv.org/abs/2505.09241v1
Date: 2025-05-14
Summary:
We present a fully analytical propagator for the orbits of lunar artificial satellites in a lunar gravity and third-body model sufficiently precise for a wide range of practical applications. The gravity model includes the twelve most important lunar gravity harmonics as well as the Earth's quadrupole tidal terms with a precise representation of the Earth's lunicentric ephemeris, and it gives an accuracy comparable to the way more extended semi-analytical propagator SELENA [6] for satellite orbits at altitudes from 300 to 3000 km. Extra terms of a more complete gravity model are straightforward to include using the formulas of the presently discussed analytical theory. The theory is based on deriving an approximate analytical solution of the secular part of the equations of motion using a Hamiltonian normal form in closed form. In total, we have two types of element transformations: from osculating to mean elements (as in [6]), and from mean to proper elements. The solution of the problem in proper elements is trivial, and, through the inverses of the above transformations, it allows to recover the position and velocity of a satellite analytically at any time t given initial conditions of the osculating elements at time $t_0$ without any intermediate numerical propagation. The propagator model is valid in time spans of several decades, and for every initial condition leading to no-fall on the Moon's surface, except for identified thin zones around a set of secular resonances corresponding to commensurabilities between the satellite's secular frequencies and the secular frequencies of the lunicentric Earth's orbit. Open software python and symbolic routines implementing our propagator are provided in the repository [14]. Precision tests with respect to fully numerical orbital propagation in Cartesian coordinates are reported.
--------------------------------------------------------------------------------------------------------
Evaluating Explanation Quality in X-IDS Using Feature Alignment Metrics
This paper introduces novel metrics for evaluating explanations provided by explainable intrusion detection systems (X-IDSs). Rather than focusing solely on model-specific properties like fidelity, the authors propose measuring how well explanations align with domain-specific knowledge bases, making them more meaningful and actionable for security analysts. The research demonstrates how these metrics reveal quality differences across various X-IDS implementations and attack types. This approach has direct applications in cybersecurity operations, helping organizations select and improve intrusion detection systems that provide understandable, domain-relevant explanations of detected threats to enhance security decision-making.
Authors: Mohammed Alquliti, Erisa Karafili, BooJoong Kang
Link: https://arxiv.org/abs/2505.08006v1
Date: 2025-05-12
Summary:
Explainable artificial intelligence (XAI) methods have become increasingly important in the context of explainable intrusion detection systems (X-IDSs) for improving the interpretability and trustworthiness of X-IDSs. However, existing evaluation approaches for XAI focus on model-specific properties such as fidelity and simplicity, and neglect whether the explanation content is meaningful or useful within the application domain. In this paper, we introduce new evaluation metrics measuring the quality of explanations from X-IDSs. The metrics aim at quantifying how well explanations are aligned with predefined feature sets that can be identified from domain-specific knowledge bases. Such alignment with these knowledge bases enables explanations to reflect domain knowledge and enables meaningful and actionable insights for security analysts. In our evaluation, we demonstrate the use of the proposed metrics to evaluate the quality of explanations from X-IDSs. The experimental results show that the proposed metrics can offer meaningful differences in explanation quality across X-IDSs and attack types, and assess how well X-IDS explanations reflect known domain knowledge. The findings of the proposed metrics provide actionable insights for security analysts to improve the interpretability of X-IDS in practical settings.
--------------------------------------------------------------------------------------------------------
Benchmarking Retrieval-Augmented Generation for Chemistry
This work introduces ChemRAG-Bench, a comprehensive benchmark for evaluating retrieval-augmented generation (RAG) in chemistry applications, alongside a specialized chemistry corpus integrating diverse knowledge sources including scientific literature, databases, and textbooks. The accompanying ChemRAG-Toolkit supports multiple retrieval algorithms and LLMs. Through extensive experiments, the authors demonstrate RAG's significant performance improvements over direct inference in chemistry tasks. This research enables more effective AI applications in chemistry, from drug discovery and materials science to chemical education and research assistance, by providing practical recommendations for implementing RAG systems tailored to chemistry's specialized knowledge requirements.
Authors: Xianrui Zhong, Bowen Jin, Siru Ouyang, Yanzhen Shen, Qiao Jin, Yin Fang, Zhiyong Lu, Jiawei Han
Link: https://arxiv.org/abs/2505.07671v1
Date: 2025-05-12
Summary:
Retrieval-augmented generation (RAG) has emerged as a powerful framework for enhancing large language models (LLMs) with external knowledge, particularly in scientific domains that demand specialized and dynamic information. Despite its promise, the application of RAG in the chemistry domain remains underexplored, primarily due to the lack of high-quality, domain-specific corpora and well-curated evaluation benchmarks. In this work, we introduce ChemRAG-Bench, a comprehensive benchmark designed to systematically assess the effectiveness of RAG across a diverse set of chemistry-related tasks. The accompanying chemistry corpus integrates heterogeneous knowledge sources, including scientific literature, the PubChem database, PubMed abstracts, textbooks, and Wikipedia entries. In addition, we present ChemRAG-Toolkit, a modular and extensible RAG toolkit that supports five retrieval algorithms and eight LLMs. Using ChemRAG-Toolkit, we demonstrate that RAG yields a substantial performance gain -- achieving an average relative improvement of 17.4% over direct inference methods. We further conduct in-depth analyses on retriever architectures, corpus selection, and the number of retrieved passages, culminating in practical recommendations to guide future research and deployment of RAG systems in the chemistry domain. The code and data is available at https://chemrag.github.io.
--------------------------------------------------------------------------------------------------------
Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking
This paper presents DeepFund, a groundbreaking benchmark tool for evaluating LLMs in real-time market conditions, eliminating the "time travel" problem where models inadvertently leverage future information from their training data. Using a multi-agent architecture connected to live stock market data, DeepFund evaluates models across multiple investment dimensions including ticker analysis, decision-making, portfolio management, and risk control. Their findings reveal significant practical limitations even in cutting-edge models like DeepSeek-V3 and Claude-3.7-Sonnet, which incurred net trading losses during evaluation. This work provides crucial insights for developing more effective AI-driven financial management systems and realistic performance expectations.
Authors: Changlun Li, Yao Shi, Chen Wang, Qiqi Duan, Runke Ruan, Weijie Huang, Haonan Long, Lijun Huang, Yuyu Luo, Nan Tang
Link: https://arxiv.org/abs/2505.11065v1
Date: 2025-05-16
Summary:
Large Language Models (LLMs) have demonstrated notable capabilities across financial tasks, including financial report summarization, earnings call transcript analysis, and asset classification. However, their real-world effectiveness in managing complex fund investment remains inadequately assessed. A fundamental limitation of existing benchmarks for evaluating LLM-driven trading strategies is their reliance on historical back-testing, inadvertently enabling LLMs to "time travel"-leveraging future information embedded in their training corpora, thus resulting in possible information leakage and overly optimistic performance estimates. To address this issue, we introduce DeepFund, a live fund benchmark tool designed to rigorously evaluate LLM in real-time market conditions. Utilizing a multi-agent architecture, DeepFund connects directly with real-time stock market data-specifically data published after each model pretraining cutoff-to ensure fair and leakage-free evaluations. Empirical tests on nine flagship LLMs from leading global institutions across multiple investment dimensions-including ticker-level analysis, investment decision-making, portfolio management, and risk control-reveal significant practical challenges. Notably, even cutting-edge models such as DeepSeek-V3 and Claude-3.7-Sonnet incur net trading losses within DeepFund real-time evaluation environment, underscoring the present limitations of LLMs for active fund management. Our code is available at https://github.com/HKUSTDial/DeepFund.
--------------------------------------------------------------------------------------------------------
How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference
This timely study introduces a framework for quantifying the environmental impact of LLM inference across 30 state-of-the-art models in commercial data centers. By combining API performance data with regional environmental metrics, the researchers reveal dramatic differences in resource consumption - from energy-intensive models like o3 and DeepSeek-R1 consuming 33 Wh per long prompt to more efficient options. The paper illustrates how widespread AI adoption drives substantial resource demands, with usage at scale requiring electricity equivalent to thousands of homes and significant water consumption. This research establishes a standardized methodology for environmental benchmarking, essential for sustainable AI development and deployment decisions.
Authors: Nidhal Jegham, Marwen Abdelatti, Lassad Elmoubarki, Abdeltawab Hendawi
Link: https://arxiv.org/abs/2505.09598v2
Date: 2025-05-15
Summary:
This paper introduces a novel infrastructure-aware benchmarking framework for quantifying the environmental footprint of LLM inference across 30 state-of-the-art models as deployed in commercial data centers. Our framework combines public API performance data with region-specific environmental multipliers and statistical inference of hardware configurations. We additionally utilize cross-efficiency Data Envelopment Analysis (DEA) to rank models by performance relative to environmental cost. Our results show that o3 and DeepSeek-R1 emerge as the most energy-intensive models, consuming over 33 Wh per long prompt, more than 70 times the consumption of GPT-4.1 nano, and that Claude-3.7 Sonnet ranks highest in eco-efficiency. While a single short GPT-4o query consumes 0.43 Wh, scaling this to 700 million queries/day results in substantial annual environmental impacts. These include electricity use comparable to 35,000 U.S. homes, freshwater evaporation matching the annual drinking needs of 1.2 million people, and carbon emissions requiring a Chicago-sized forest to offset. These findings illustrate a growing paradox: Although AI is becoming cheaper and faster, its global adoption drives disproportionate resource consumption. Our study provides a standardized, empirically grounded methodology for benchmarking the sustainability of LLM deployments, laying a foundation for future environmental accountability in AI development and sustainability standards.
--------------------------------------------------------------------------------------------------------
A Survey on the Safety and Security Threats of Computer-Using Agents: JARVIS or Ultron?
This paper provides a systematic analysis of safety and security threats posed by Computer-Using Agents (CUAs) - LLM-based systems that autonomously navigate desktop applications, web pages, and mobile apps. The authors establish a comprehensive taxonomy of safety threats, defensive strategies, and evaluation metrics for these increasingly capable autonomous systems. By identifying vulnerabilities in LLM-driven reasoning and complex multi-component interactions, this work provides researchers and practitioners with a structured foundation for exploring security vulnerabilities and implementing protective measures. The findings have broad implications for designing secure AI assistance systems as they gain greater capability to operate digital interfaces autonomously.
Authors: Ada Chen, Yongjiang Wu, Junyuan Zhang, Shu Yang, Jen-tse Huang, Kun Wang, Wenxuan Wang, Shuai Wang
Link: https://arxiv.org/abs/2505.10924v1
Date: 2025-05-16
Summary:
Recently, AI-driven interactions with computing devices have advanced from basic prototype tools to sophisticated, LLM-based systems that emulate human-like operations in graphical user interfaces. We are now witnessing the emergence of \emph{Computer-Using Agents} (CUAs), capable of autonomously performing tasks such as navigating desktop applications, web pages, and mobile apps. However, as these agents grow in capability, they also introduce novel safety and security risks. Vulnerabilities in LLM-driven reasoning, with the added complexity of integrating multiple software components and multimodal inputs, further complicate the security landscape. In this paper, we present a systematization of knowledge on the safety and security threats of CUAs. We conduct a comprehensive literature review and distill our findings along four research objectives: \textit{\textbf{(i)}} define the CUA that suits safety analysis; \textit{\textbf{(ii)} } categorize current safety threats among CUAs; \textit{\textbf{(iii)}} propose a comprehensive taxonomy of existing defensive strategies; \textit{\textbf{(iv)}} summarize prevailing benchmarks, datasets, and evaluation metrics used to assess the safety and performance of CUAs. Building on these insights, our work provides future researchers with a structured foundation for exploring unexplored vulnerabilities and offers practitioners actionable guidance in designing and deploying secure Computer-Using Agents.
--------------------------------------------------------------------------------------------------------
SEAL: Searching Expandable Architectures for Incremental Learning
This paper introduces SEAL, an innovative framework combining Neural Architecture Search with selective expansion for data-incremental learning scenarios where disjoint data samples arrive sequentially without storage. Unlike existing approaches that expand models with each new task, SEAL adapts structure dynamically based on capacity needs, preserving stability through cross-distillation. The framework's dual optimization of architecture and expansion policy demonstrates significant improvements in reducing forgetting while maintaining smaller model sizes compared to prior methods. SEAL offers promising applications in resource-constrained environments like mobile devices, IoT networks, and edge computing systems that require continuous adaptation to new data streams without compromising performance on previously learned tasks.
Authors: Matteo Gambella, Vicente Javier Castro Solar, Manuel Roveri
Link: https://arxiv.org/abs/2505.10457v1
Date: 2025-05-15
Summary:
Incremental learning is a machine learning paradigm where a model learns from a sequential stream of tasks. This setting poses a key challenge: balancing plasticity (learning new tasks) and stability (preserving past knowledge). Neural Architecture Search (NAS), a branch of AutoML, automates the design of the architecture of Deep Neural Networks and has shown success in static settings. However, existing NAS-based approaches to incremental learning often rely on expanding the model at every task, making them impractical in resource-constrained environments. In this work, we introduce SEAL, a NAS-based framework tailored for data-incremental learning, a scenario where disjoint data samples arrive sequentially and are not stored for future access. SEAL adapts the model structure dynamically by expanding it only when necessary, based on a capacity estimation metric. Stability is preserved through cross-distillation training after each expansion step. The NAS component jointly searches for both the architecture and the optimal expansion policy. Experiments across multiple benchmarks demonstrate that SEAL effectively reduces forgetting and enhances accuracy while maintaining a lower model size compared to prior methods. These results highlight the promise of combining NAS and selective expansion for efficient, adaptive learning in incremental scenarios.
--------------------------------------------------------------------------------------------------------
This research implements an innovative rule-based strategy approach to train Deep Reinforcement Learning (DRL) algorithms for trading Bitcoin and Ripple cryptocurrencies. By integrating Deep Q-Network, Double Deep Q-Network, Dueling Deep Q-learning networks, and Advantage Actor-Critic algorithms, the authors create optimal trading policies evaluated through portfolio wealth and trade signal metrics. Their experimental results demonstrate that Duelling and Double Deep Q-Network particularly excel with XRP trading. This work has practical applications for cryptocurrency traders, investment firms developing automated trading systems, and financial technology companies seeking to leverage AI for volatile asset management while addressing challenges posed by the complex price dynamics of digital currencies.
Authors: Dieu-Donne Fangnon, Armandine Sorel Kouyim Meli, Verlon Roel Mbingui, Phanie Dianelle Negho, Regis Konan Marcel Djaha
Link: https://arxiv.org/abs/2505.07660v1
Date: 2025-05-12
Summary:
Artificial intelligence (AI) has demonstrated remarkable success across various applications. In light of this trend, the field of automated trading has developed a keen interest in leveraging AI techniques to forecast the future prices of financial assets. This interest stems from the need to address trading challenges posed by the inherent volatility and dynamic nature of asset prices. However, crafting a flawless strategy becomes a formidable task when dealing with assets characterized by intricate and ever-changing price dynamics. To surmount these formidable challenges, this research employs an innovative rule-based strategy approach to train Deep Reinforcement Learning (DRL). This application is carried out specifically in the context of trading Bitcoin (BTC) and Ripple (XRP). Our proposed approach hinges on the integration of Deep Q-Network, Double Deep Q-Network, Dueling Deep Q-learning networks, alongside the Advantage Actor-Critic algorithms. Each of them aims to yield an optimal policy for our application. To evaluate the effectiveness of our Deep Reinforcement Learning (DRL) approach, we rely on portfolio wealth and the trade signal as performance metrics. The experimental outcomes highlight that Duelling and Double Deep Q-Network outperformed when using XRP with the increasing of the portfolio wealth. All codes are available in this \href{https://github.com/VerlonRoelMBINGUI/RL_Final_Projects_AMMI2023}{\color{blue}Github link}.
--------------------------------------------------------------------------------------------------------
PhiNet v2: A Mask-Free Brain-Inspired Vision Foundation Model from Video
PhiNet v2 introduces a novel Transformer-based architecture that processes temporal visual input without relying on strong augmentation, representing a significant advancement in brain-inspired self-supervised learning. Unlike its predecessor that operated on static images, this model leverages variational inference to learn robust visual representations from continuous input streams, more closely mimicking human visual processing. Through extensive experimentation, the authors demonstrate competitive performance compared to state-of-the-art vision foundation models. PhiNet v2 has potential applications in video understanding, autonomous systems, surveillance, and any computer vision task requiring temporal reasoning with reduced dependence on artificial data augmentation techniques.
Authors: Makoto Yamada, Kian Ming A. Chai, Ayoub Rhim, Satoki Ishikawa, Mohammad Sabokrou, Yao-Hung Hubert Tsai
Link: https://arxiv.org/abs/2505.11129v1
Date: 2025-05-16
Summary:
Recent advances in self-supervised learning (SSL) have revolutionized computer vision through innovative architectures and learning objectives, yet they have not fully leveraged insights from biological visual processing systems. Recently, a brain-inspired SSL model named PhiNet was proposed; it is based on a ResNet backbone and operates on static image inputs with strong augmentation. In this paper, we introduce PhiNet v2, a novel Transformer-based architecture that processes temporal visual input (that is, sequences of images) without relying on strong augmentation. Our model leverages variational inference to learn robust visual representations from continuous input streams, similar to human visual processing. Through extensive experimentation, we demonstrate that PhiNet v2 achieves competitive performance compared to state-of-the-art vision foundation models, while maintaining the ability to learn from sequential input without strong data augmentation. This work represents a significant step toward more biologically plausible computer vision systems that process visual information in a manner more closely aligned with human cognitive processes.
--------------------------------------------------------------------------------------------------------
This groundbreaking paper challenges current machine unlearning methods by demonstrating that adversaries can distinguish between models produced by unlearning techniques and control models retrained without forgotten data. The authors propose "computational unlearning," a formal definition requiring that adversaries cannot differentiate unlearned models from retraining from scratch except with negligible probability. Their analysis reveals fundamental limitations, including the impossibility of deterministic computational unlearning for entropic learning algorithms and severe utility tradeoffs in differential privacy approaches. This work has significant implications for privacy-preserving machine learning, regulatory compliance, and the development of more rigorous unlearning techniques to support the right to be forgotten.
Authors: Brennon Brimhall, Philip Mathew, Neil Fendley, Yinzhi Cao, Matthew Green
Link: https://arxiv.org/abs/2505.08138v1
Date: 2025-05-13
Summary:
Machine unlearning methods take a model trained on a dataset and a forget set, then attempt to produce a model as if it had only been trained on the examples not in the forget set. We empirically show that an adversary is able to distinguish between a mirror model (a control model produced by retraining without the data to forget) and a model produced by an unlearning method across representative unlearning methods from the literature. We build distinguishing algorithms based on evaluation scores in the literature (i.e. membership inference scores) and Kullback-Leibler divergence. We propose a strong formal definition for machine unlearning called computational unlearning. Computational unlearning is defined as the inability for an adversary to distinguish between a mirror model and a model produced by an unlearning method. If the adversary cannot guess better than random (except with negligible probability), then we say that an unlearning method achieves computational unlearning. Our computational unlearning definition provides theoretical structure to prove unlearning feasibility results. For example, our computational unlearning definition immediately implies that there are no deterministic computational unlearning methods for entropic learning algorithms. We also explore the relationship between differential privacy (DP)-based unlearning methods and computational unlearning, showing that DP-based approaches can satisfy computational unlearning at the cost of an extreme utility collapse. These results demonstrate that current methodology in the literature fundamentally falls short of achieving computational unlearning. We conclude by identifying several open questions for future work.
--------------------------------------------------------------------------------------------------------
Enhancing Thyroid Cytology Diagnosis with RAG-Optimized LLMs and Pa-thology Foundation Models
This medical research paper explores the integration of retrieval-augmented generation (RAG) enhanced LLMs with pathology foundation models for thyroid cytology diagnosis. By leveraging a curated knowledge base of case studies, diagnostic criteria, and expert interpretations alongside specialized image analysis, the combined approach enhances diagnostic consistency and accuracy in distinguishing benign from malignant thyroid lesions. The UNI foundation model achieved impressive AUC scores of 0.73-0.93 for predicting surgical pathology outcomes from cytology samples. This technology has direct applications in clinical pathology, potentially improving diagnostic efficiency, reducing variability in interpretation, and supporting pathologists in making more accurate diagnoses.
Authors: Hussien Al-Asi, Jordan P Reynolds, Shweta Agarwal, Bryan J Dangott, Aziza Nassar, Zeynettin Akkus
Link: https://arxiv.org/abs/2505.08590v1
Date: 2025-05-13
Summary:
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.
--------------------------------------------------------------------------------------------------------
This insightful study introduces a benchmark using narrative brainteasers to investigate how large language models (LLMs) solve problems, focusing not only on correctness but on solution quality and creativity. The researchers explore multiple aspects of the reasoning process, including semantic parsing, solution generation, self-correction, step-by-step sketching, and utilization of hints. Their findings reveal that while LLMs can often find creative, insightful solutions to complex problems, they sometimes resort to brute force approaches despite the availability of more elegant solutions. This research provides valuable insights for AI developers seeking to enhance reasoning capabilities in LLMs for applications requiring creative problem-solving and novel approaches.
Authors: Simeng Han, Stephen Xia, Grant Zhang, Howard Dai, Chen Liu, Lichang Chen, Hoang Huy Nguyen, Hongyuan Mei, Jiayuan Mao, R. Thomas McCoy
Link: https://arxiv.org/abs/2505.10844v1
Date: 2025-05-16
Summary:
Accuracy remains a standard metric for evaluating AI systems, but it offers limited insight into how models arrive at their solutions. In this work, we introduce a benchmark based on brainteasers written in long narrative form to probe more deeply into the types of reasoning strategies that models use. Brainteasers are well-suited for this goal because they can be solved with multiple approaches, such as a few-step solution that uses a creative insight or a longer solution that uses more brute force. We investigate large language models (LLMs) across multiple layers of reasoning, focusing not only on correctness but also on the quality and creativity of their solutions. We investigate many aspects of the reasoning process: (1) semantic parsing of the brainteasers into precise mathematical competition style formats; (2) generating solutions from these mathematical forms; (3) self-correcting solutions based on gold solutions; (4) producing step-by-step sketches of solutions; and (5) making use of hints. We find that LLMs are in many cases able to find creative, insightful solutions to brainteasers, suggesting that they capture some of the capacities needed to solve novel problems in creative ways. Nonetheless, there also remain situations where they rely on brute force despite the availability of more efficient, creative solutions, highlighting a potential direction for improvement in the reasoning abilities of LLMs.
--------------------------------------------------------------------------------------------------------
ZENN: A Thermodynamics-Inspired Computational Framework for Heterogeneous Data-Driven Modeling
ZENN (zentropy-enhanced neural network) introduces a groundbreaking approach to machine learning by extending zentropy theory into data science, enabling more effective learning from heterogeneous data sources. The framework simultaneously captures both energy and intrinsic entropy components of multi-source data through a redesigned neural network architecture. Experimental results demonstrate superior generalization capabilities and robustness in classification tasks and energy landscape reconstructions, particularly for high-order derivatives. When applied to reconstructing the Helmholtz energy landscape of Fe3Pt, ZENN successfully captured complex material behaviors including negative thermal expansion. This approach has promising applications in materials science, physics, chemistry, and any field requiring machine learning from diverse, heterogeneous datasets.
Authors: Shun Wang, Shun-Li Shang, Zi-Kui Liu, Wenrui Hao
Link: https://arxiv.org/abs/2505.09851v1
Date: 2025-05-14
Summary:
Traditional entropy-based methods - such as cross-entropy loss in classification problems - have long been essential tools for quantifying uncertainty and disorder in data and developing artificial intelligence algorithms. However, the rapid growth of data across various domains has introduced new challenges, particularly the integration of heterogeneous datasets with intrinsic disparities. In this paper, we extend zentropy theory into the data science domain by introducing intrinsic entropy, enabling more effective learning from heterogeneous data sources. We propose a zentropy-enhanced neural network (ZENN) that simultaneously learns both energy and intrinsic entropy components, capturing the underlying structure of multi-source data. To support this, we redesign the neural network architecture to better reflect the intrinsic properties and variability inherent in diverse datasets. We demonstrate the effectiveness of ZENN on classification tasks and energy landscape reconstructions, showing its superior generalization capabilities and robustness-particularly in predicting high-order derivatives. As a practical application, we employ ZENN to reconstruct the Helmholtz energy landscape of Fe3Pt using data generated from DFT and capture key material behaviors, including negative thermal expansion and the critical point in the temperature-pressure space. Overall, our study introduces a novel approach for data-driven machine learning grounded in zentropy theory, highlighting ZENN as a versatile and robust deep learning framework for scientific problems involving complex, heterogeneous datasets.
--------------------------------------------------------------------------------------------------------
Behind the Noise: Conformal Quantile Regression Reveals Emergent Representations
This innovative research presents a machine learning approach that not only denoises low-quality scientific imaging data but also uncovers meaningful structure in the latent space. Using ensembles of lightweight neural networks trained via conformal quantile regression, the method performs reliable denoising with calibrated uncertainty bounds while revealing interpretable spatial and chemical features without requiring labels or segmentation. Validated on geobiochemical imaging data, this approach has broad applications in scientific fields where acquisition time constraints result in noisy measurements, enabling confident interpretation of data and guiding experimental design under resource limitations while extracting meaningful patterns that might otherwise remain hidden.
Authors: Petrus H. Zwart, Tamas Varga, Odeta Qafoku, James A. Sethian
Link: https://arxiv.org/abs/2505.08176v1
Date: 2025-05-13
Summary:
Scientific imaging often involves long acquisition times to obtain high-quality data, especially when probing complex, heterogeneous systems. However, reducing acquisition time to increase throughput inevitably introduces significant noise into the measurements. We present a machine learning approach that not only denoises low-quality measurements with calibrated uncertainty bounds, but also reveals emergent structure in the latent space. By using ensembles of lightweight, randomly structured neural networks trained via conformal quantile regression, our method performs reliable denoising while uncovering interpretable spatial and chemical features -- without requiring labels or segmentation. Unlike conventional approaches focused solely on image restoration, our framework leverages the denoising process itself to drive the emergence of meaningful representations. We validate the approach on real-world geobiochemical imaging data, showing how it supports confident interpretation and guides experimental design under resource constraints.
--------------------------------------------------------------------------------------------------------
The Way We Prompt: Conceptual Blending, Neural Dynamics, and Prompt-Induced Transitions in LLMs
This philosophical paper explores the parallels between human cognition and large language models through the lens of Conceptual Blending Theory (CBT). By systematically investigating Prompt-Induced Transitions and Hallucinations, the author reveals how LLMs blend and compress meaning in ways structurally similar to human cognitive processes. The framework bridges linguistics, neuroscience, and AI research, suggesting that prompt engineering serves not merely as a technical tool but as a scientific method for investigating the deep structure of meaning. This approach offers new perspectives for cognitive scientists, AI researchers, and human-AI collaboration specialists seeking to understand the mechanisms behind LLM behaviors and their relationship to human intelligence.
Authors: Makoto Sato
Link: https://arxiv.org/abs/2505.10948v1
Date: 2025-05-16
Summary:
Large language models (LLMs), inspired by neuroscience, exhibit behaviors that often evoke a sense of personality and intelligence-yet the mechanisms behind these effects remain elusive. Here, we operationalize Conceptual Blending Theory (CBT) as an experimental framework, using prompt-based methods to reveal how LLMs blend and compress meaning. By systematically investigating Prompt-Induced Transitions (PIT) and Prompt-Induced Hallucinations (PIH), we uncover structural parallels and divergences between artificial and biological cognition. Our approach bridges linguistics, neuroscience, and empirical AI research, demonstrating that human-AI collaboration can serve as a living prototype for the future of cognitive science. This work proposes prompt engineering not just as a technical tool, but as a scientific method for probing the deep structure of meaning itself.
--------------------------------------------------------------------------------------------------------
Rhetorical XAI: Explaining AI's Benefits as well as its Use via Rhetorical Design
This paper reconceptualizes Explainable AI (XAI) through the framework of Rhetorical Design, emphasizing that explanations function as a form of argumentation that shapes user perceptions of system usefulness and credibility. By analyzing explanations through the lens of logical reasoning, projected credibility, and emotional resonance, the authors demonstrate how rhetorical appeals influence user trust and facilitate AI adoption. The research synthesizes design strategies from prior XAI work and highlights opportunities for integrating rhetorical principles into explanation systems. This approach has applications in designing more persuasive and effective AI explanations across domains including healthcare, finance, and decision support systems where user trust and adoption are critical.
Authors: Houjiang Liu, Yiheng Su, Matthew Lease
Link: https://arxiv.org/abs/2505.09862v1
Date: 2025-05-14
Summary:
This paper explores potential benefits of incorporating Rhetorical Design into the design of Explainable Artificial Intelligence (XAI) systems. While XAI is traditionally framed around explaining individual predictions or overall system behavior, explanations also function as a form of argumentation, shaping how users evaluate system perceived usefulness, credibility, and foster appropriate trust. Rhetorical Design offers a useful framework to analyze the communicative role of explanations between AI systems and users, focusing on: (1) logical reasoning conveyed through different types of explanations, (2) credibility projected by the system and its developers, and (3) emotional resonance elicited in users. Together, these rhetorical appeals help us understand how explanations influence user perceptions and facilitate AI adoption. This paper synthesizes design strategies from prior XAI work that align with these three rhetorical appeals and highlights both opportunities and challenges of integrating rhetorical design into XAI design.
--------------------------------------------------------------------------------------------------------
Explainable Reinforcement Learning Agents Using World Models
This paper introduces an innovative technique for generating explanations in Model-Based Deep Reinforcement Learning agents using World Models. Recognizing that understanding what a user wanted the agent to do isn't sufficient to explain why the agent chose differently, the authors augment these agents with a Reverse World Model that predicts what the world state should have been for the agent to prefer a given counterfactual action. Their research demonstrates that explanations showing what the world should have looked like significantly improve user understanding of agent policies. This approach has applications in autonomous systems, robotics, and human-AI collaboration scenarios where non-expert users need to comprehend and potentially influence AI agent behavior.
Authors: Madhuri Singh, Amal Alabdulkarim, Gennie Mansi, Mark O. Riedl
Link: https://arxiv.org/abs/2505.08073v1
Date: 2025-05-12
Summary:
Explainable AI (XAI) systems have been proposed to help people understand how AI systems produce outputs and behaviors. Explainable Reinforcement Learning (XRL) has an added complexity due to the temporal nature of sequential decision-making. Further, non-AI experts do not necessarily have the ability to alter an agent or its policy. We introduce a technique for using World Models to generate explanations for Model-Based Deep RL agents. World Models predict how the world will change when actions are performed, allowing for the generation of counterfactual trajectories. However, identifying what a user wanted the agent to do is not enough to understand why the agent did something else. We augment Model-Based RL agents with a Reverse World Model, which predicts what the state of the world should have been for the agent to prefer a given counterfactual action. We show that explanations that show users what the world should have been like significantly increase their understanding of the agent policy. We hypothesize that our explanations can help users learn how to control the agents execution through by manipulating the environment.
--------------------------------------------------------------------------------------------------------