📚 Research Digest

Finding: The hybrid confirmation tree (HCT) method, where humans and AI make independent judgments that are combined through a tie-breaking process, consistently outperforms traditional AI-as-advisor approaches across 10 datasets because people cannot effectively discriminate between correct and incorrect AI advice.

Why it matters: This challenges fundamental assumptions about how to design human-AI collaboration systems and provides a concrete alternative framework that preserves human agency while improving decision accuracy.

Method: Compared HCT versus AI-as-advisor across 10 datasets from medical diagnostics to misinformation detection, using signal detection theory to analyze why HCT performs better.

Jargon: HCT = hybrid confirmation tree (independent human/AI judgments with human tie-breaker); Signal detection theory = framework for analyzing decision-making under uncertainty.

Finding: A collaborative AI-in-the-loop workflow for discussion facilitation in online learning communities enhanced social presence and higher-order cognitive processes when designed with reciprocal exchange, discourse-adaptive roles, and human review, but only when learners directly interacted with the AI rather than merely being exposed to AI-generated content.

Why it matters: This demonstrates specific design principles for human-AI collaboration in educational contexts that could inform broader HCI design for AI-supported teamwork and learning environments.

Method: Design-based research across two iterations in a 5-week cMOOC (N=606) testing different AI facilitation patterns with network-structure-driven targeting and mandatory human review before AI participation.

Jargon: cMOOCs are Connectivist Massive Open Online Courses emphasizing learner-driven networking rather than instructor-led content delivery; social/cognitive presence refer to learners' sense of community connection and depth of critical thinking in online discussions.

Finding: Structured intent representations (like 5W3H frameworks) significantly improve human-AI communication robustness across different AI models and languages, reducing cross-language variance from 0.470 to 0.020 and cutting user interaction rounds by 60% while increasing satisfaction from 3.16 to 4.04.

Why it matters: This demonstrates how structured communication protocols can enhance human-AI collaboration effectiveness, directly informing HCI design principles for AI interfaces and team collaboration tools.

Method: Evaluated 3,240 model outputs across 3 AI models, 3 languages, and multiple prompting frameworks, plus a 50-person user study of AI-assisted intent expansion in realistic settings.

Jargon: 5W3H = Who, What, When, Where, Why + How, How much, How many framework; PPS = Prompt Protocol Specification; CO-STAR and RISEN = alternative structured prompting frameworks.

Finding: This paper introduces psychophysical frameworks (meta-d' and signal detection theory) to measure AI systems' metacognitive abilities—their capacity to assess confidence in their own decisions and regulate behavior based on uncertainty and risk.

Why it matters: Understanding AI metacognition is crucial for designing human-AI collaboration systems where humans need to know when to trust AI recommendations and how AI systems can communicate their uncertainty effectively.

Method: Two experiments with three large language models testing confidence ratings on judgments and decision regulation under varying risk conditions, using established psychophysical measurement frameworks.

Jargon: Meta-d' = framework measuring metacognitive sensitivity (ability to distinguish correct from incorrect responses through confidence ratings); Signal Detection Theory (SDT) = method for measuring decision-making under uncertainty and risk.

Finding: KEditVis is a visual analytics system that helps users understand and optimize knowledge editing in large language models by providing interactive visualizations to select optimal editing layers and explore reasons behind ineffective edits.

Why it matters: This directly addresses human-AI collaboration and the critical UX challenges of making AI systems more transparent and controllable, which is essential for understanding how humans can effectively work with and improve AI systems.

Method: The system was evaluated through usage scenarios, expert interviews, and a user study to validate effectiveness and usability.

Jargon: Knowledge editing - techniques for correcting factual information stored in LLMs; Visual analytics - interactive data visualization combined with analytical reasoning.

Finding: AI use creates a "metacognitive decoupling" where people produce better outputs while becoming less accurate at judging their own competence, challenging the oversimplified view that AI just amplifies Dunning-Kruger effects. This creates a four-way gap between output quality, actual understanding, calibration accuracy, and self-assessed ability.

Why it matters: This framework is crucial for understanding how AI tools affect learning, skill development, and decision-making in educational and professional contexts, directly informing design of AI-assisted learning systems and assessment methods.

Method: Theoretical synthesis of existing evidence from human-AI interaction studies, learning research, and model evaluation to propose a new conceptual framework.

Jargon: Metacognitive decoupling = separation between actual understanding and awareness of one's competence; Calibration accuracy = how well self-assessments match actual performance; Dunning-Kruger effect = tendency for less competent people to overestimate their abilities.

Finding: This paper introduces "concept frustration" - a geometric framework that detects when AI systems learn hidden concepts that contradict human conceptual understanding, and shows how incorporating these frustrating concepts can better align human and machine reasoning.

Why it matters: This directly addresses a core HCI challenge of making AI systems interpretable and aligned with human mental models, which is crucial for human-AI collaboration and interface design.

Method: Uses geometric analysis of foundation model embeddings combined with supervised concept models, validated on both synthetic data and real language/vision tasks with closed-form mathematical derivations.

Jargon: Foundation models = large pre-trained AI systems like GPT; concept frustration = when hidden AI concepts create contradictions in human conceptual frameworks; task-aligned geometry = mathematical representations optimized for specific tasks rather than general distance measures.

Finding: Semantic interaction enables analysts to directly manipulate AI-powered narrative visualizations, with map-based representations significantly outperforming timeline baselines for sensemaking, and users employing two distinct approaches (corrective and additive) to refine AI-extracted narratives.

Why it matters: This demonstrates how direct manipulation interfaces can improve human-AI collaboration in sensemaking tasks, offering design principles for interactive AI systems that support cognitive processes.

Method: Controlled user study with 33 participants comparing three conditions (timeline baseline, basic narrative map, interactive narrative map with semantic interaction) using insight-based evaluation metrics.

Jargon: Semantic interaction (SI) = direct manipulation of visualizations that feeds back into underlying AI models; narrative sensemaking = process of understanding and organizing story elements from complex information.

Finding: The Visual Analytics Context Protocol (VACP) makes visual analytics applications "agent-ready" by exposing application state and interactions to AI agents, enabling them to collaborate more effectively with humans in data analysis tasks compared to current computer vision-based approaches.

Why it matters: This directly addresses AI's impact on human-computer collaboration by creating structured protocols for AI agents to work alongside humans in visual analytics, which is crucial for understanding how AI transforms collaborative work practices.

Method: The authors developed a formal specification and library implementation compatible with major visualization frameworks, then evaluated it across representative visual analytics tasks comparing success rates and efficiency metrics.

Jargon: Visual Analytics (VA) - interactive data analysis combining automated analysis with human judgment through visual interfaces; DOM - Document Object Model, the structural representation of web pages.

Finding: LLMs systematically reproduce traditional caste hierarchies in matchmaking evaluations, rating same-caste matches up to 25% higher than inter-caste matches and ordering preferences according to historical social stratification patterns across five different model families.

Why it matters: This demonstrates how AI systems can perpetuate harmful social biases in high-stakes personal decisions, highlighting critical considerations for HCI design and human-AI collaboration in socially sensitive applications.

Method: Controlled audit varying caste identity and income levels across matrimonial profiles, testing five LLM families on social acceptance, marital stability, and cultural compatibility ratings.

Jargon: Caste audit - systematic evaluation of how AI systems treat different caste identities; Brahmin/Kshatriya/Vaishya/Shudra/Dalit - traditional Hindu caste categories from highest to lowest in historical hierarchy.

Finding: A human-AI collaboration system for small retail stores significantly improved point-of-purchase marketing text quality by 2.37 points on a 7-point scale, combining LLM text generation with human insights for customer understanding and creative refinement.

Why it matters: This demonstrates effective human-AI collaboration design where complementary strengths (AI's generative capacity, human creativity and domain knowledge) address each other's weaknesses in a real-world application.

Method: Experimental evaluation comparing POP text quality created with versus without the collaborative system support, using simulated personas for evaluation.

Jargon: POP (Point of Purchase) materials = marketing displays and texts at retail locations where customers make buying decisions.

Finding: The paper proposes an experiential AI literacy approach using storytelling and brainstorming to help workers identify practical AI applications in their workflows, moving beyond abstract AI knowledge to actionable understanding.

Why it matters: This directly addresses the critical challenge of translating AI knowledge into workplace practice, offering a concrete pedagogical method for teaching AI literacy that bridges theory and application.

Method: Uses storytelling and participatory brainstorming to generate grounded AI use cases based on participants' actual work experiences and contexts.

Jargon: Experiential learning - learning through direct experience and reflection rather than abstract instruction; Participatory design - involving end users in the design process to ensure relevance and usability.

Finding: SCEIMA is a two-stage framework that systematically separates human-specific, model-specific, and interaction-driven factors in human-machine coordination by comparing human-human, human-model, and model-model interactions across empirical and simulation phases.

Why it matters: This provides HCI researchers with a rigorous method to understand whether coordination breakdowns stem from poor AI design, human adaptation to artificial partners, or emergent interaction dynamics—critical for designing better collaborative AI systems.

Method: The framework combines empirical studies (humans coordinating with both humans and AI models) with analytical simulations (AI models coordinating with each other) to isolate different sources of coordination variance.

Jargon: Sensorimotor synchronization = coordinating physical movements with external timing cues; contrastive instruments = using models as controlled comparisons to reveal human behavior patterns.

Finding: Voice delivery of empathetic messages reduces perceived empathy quality and recipients' feelings of being heard compared to text, with this "empathic voice penalty" being especially pronounced for AI empathizers due to increased feelings of uncanniness.

Why it matters: This challenges fundamental assumptions in HCI about multimodal interaction design and reveals critical considerations for developing empathetic AI systems and human-AI collaboration interfaces.

Method: Multiple experimental studies (N=799) comparing perceptions of empathy across delivery modes (voice vs. text) and sources (human vs. AI), measuring empathy quality, feeling heard, and uncanniness as mediator.

Jargon: Uncanniness refers to the unsettling feeling when something appears almost-but-not-quite human.

Finding: Socially assistive robots can serve as cognitive scaffolds that guide children with ASD through progressive social learning, transitioning from robot-led engagement to human-mediated skill generalization using a structured developmental framework.

Why it matters: This demonstrates how AI systems can be designed as adaptive learning partners that enhance rather than replace human interaction, offering insights for designing collaborative AI tools across educational and therapeutic contexts.

Method: Not applicable for theoretical framework paper.

Jargon: Socially assistive robots (SARs) - robots designed to help people through social interaction rather than physical tasks; cognitive scaffolds - external supports that help learners acquire skills they can eventually perform independently.

Finding: Social exclusion leads people to make more short-sighted decisions (choosing immediate over delayed rewards), and this effect occurs primarily through depletion of momentary self-control resources rather than through trait-level factors like impulsivity or general self-control ability.

Why it matters: This reveals a key mechanism underlying poor decision-making in social contexts, with direct implications for designing interventions that support better goal-setting and long-term thinking in teams and collaborative environments.

Method: Two experiments using network analysis to map psychological pathways - one correlational study of chronic exclusion and one manipulating acute exclusion with the Cyberball paradigm.

Jargon: Intertemporal choice = decisions between immediate vs. delayed rewards; temporal discounting = tendency to devalue future rewards; Cyberball = experimental paradigm where participants are excluded from a virtual ball-tossing game.

Finding: HapCompass is a wearable haptic device that uses a rotating actuator to provide directional feedback during robotic teleoperation, significantly improving task success rates, completion times, and force control compared to vision-only or non-directional feedback approaches.

Why it matters: This demonstrates how thoughtful haptic interface design can enhance human-robot collaboration and skill transfer, with implications for designing intuitive control interfaces and understanding multimodal feedback in HCI.

Method: Comparative user study evaluating teleoperation performance across feedback conditions, plus preliminary imitation learning evaluation showing improved demonstration quality.

Jargon: Linear resonant actuator (LRA) - a type of haptic motor that creates vibrations through linear motion; teleoperation - remotely controlling a robot; imitation learning - AI technique where robots learn by mimicking human demonstrations.

Finding: The paper proposes a Triadic Cognitive Architecture that grounds AI agent decision-making in physics-based constraints, using "cognitive friction" to prevent common failure modes like excessive tool use and prolonged deliberation that occur in current LLM-based autonomous agents.

Why it matters: This addresses critical issues in human-AI collaboration by creating more predictable and bounded AI behavior, which is essential for designing effective human-AI teams and understanding AI's impact on work dynamics.

Method: Uses mathematical modeling combining nonlinear filtering theory, Riemannian geometry, and optimal control theory, validated in a simulated Emergency Medical Diagnostic Grid environment.

Jargon: Cognitive friction = physics-inspired constraints on AI reasoning; HJB = Hamilton-Jacobi-Bellman equation from optimal control theory; ReAct = a common heuristic loop for AI agents combining reasoning and acting.

Finding: The YARN framework enhances LLMs' analogical reasoning about narratives by using four levels of abstraction to decompose stories into mappable units, consistently outperforming end-to-end LLM approaches and revealing systematic patterns in how analogical reasoning can be structured.

Why it matters: This advances understanding of how AI systems can better support human-like reasoning and problem-solving, with implications for AI-assisted creativity, learning tools that use analogical thinking, and collaborative reasoning systems.

Method: Modular framework comparing LLM-derived abstractions at four levels against end-to-end baselines, with systematic error analysis of analogical reasoning patterns.

Jargon: Structural mapping - cognitive process of aligning elements between different scenarios to find analogies; Analogical reasoning - using similarities between different situations to solve problems or make arguments.

Finding: This study develops GEO-SFE, a framework that optimizes how AI-powered search engines cite content by engineering structural features (document architecture, information chunking, visual emphasis) rather than just semantic content, achieving 17.3% improvement in citation rates across six generative engines.

Why it matters: This directly addresses how AI systems impact information visibility and discovery, providing actionable insights for designing content that effectively collaborates with AI in information ecosystems.

Method: The researchers decomposed content structure into three hierarchical levels and tested architecture-aware optimization strategies across multiple generative engine platforms to validate their framework.

Jargon: GEO (Generative Engine Optimization) - techniques to improve content visibility in AI-powered search engines that generate direct answers rather than providing links.

Finding: XR functions as a neutral umbrella term for VR/AR/MR technologies rather than an abbreviation for "Extended Reality," and stable terminology requires collaborative governance across academia, industry, and standards organizations.

Why it matters: Understanding how XR terminology evolved and functions helps HCI researchers and educators communicate more clearly about immersive technologies and their human-centered applications.

Jargon: XR = umbrella term for Virtual Reality, Augmented Reality, and Mixed Reality technologies; MR = Mixed Reality (blending physical and digital environments).

Finding: This study introduces ATP-Bench, a benchmark for evaluating how well Multimodal Large Language Models can autonomously plan and use tools to generate interleaved text-and-image responses, revealing that current state-of-the-art models struggle significantly with coherent planning across visual-critical tasks.

Why it matters: This directly addresses AI's impact on human-computer interaction by examining how AI systems can become more autonomous collaborative partners that know when and how to use different tools, which is crucial for understanding future HCI design patterns.

Method: The researchers created a comprehensive benchmark with 7,702 QA pairs across 25 visual-critical intents and developed a Multi-Agent MLLM-as-a-Judge evaluation system to assess tool-planning capabilities without requiring ground-truth references.

Jargon: MLLM = Multimodal Large Language Model (AI that processes both text and images); Agentic Tool Planning = AI autonomously deciding which tools to use and when; Interleaved generation = producing mixed text-and-image content in sequence.

Finding: DIAL introduces a two-system architecture that separates high-level intent planning (using vision-language models for world modeling) from low-level action execution in robotic control, achieving state-of-the-art performance with 10x fewer demonstrations than existing methods.

Why it matters: This demonstrates how cognitive architectures (System-1/System-2 thinking) can be applied to human-AI collaboration design and provides insights into how separating planning from execution might improve learning efficiency in skill acquisition.

Method: Uses a two-stage training approach with decoupled warmup (separate system training) followed by end-to-end joint optimization, tested on robotic manipulation tasks and real humanoid robot deployment.

Jargon: VLA (Vision-Language-Action models) - AI systems that process visual and language inputs to generate robot actions; System-1/System-2 - dual-process theory distinguishing fast automatic responses from slow deliberate reasoning.

Finding: This paper develops a theoretical framework for understanding diffusion-based image editing systems by formalizing editing as guided transport on a learned manifold and analyzing key usability factors like controllability, faithfulness to user intent, and semantic consistency.

Why it matters: This provides crucial foundations for designing AI-powered creative tools that balance user control with system capability, directly informing HCI research on human-AI collaboration in creative workflows.

Method: The authors formalize editing as a conditional reverse-time generative process and derive mathematical bounds connecting guidance parameters to measurable quality metrics under Lipschitz assumptions.

Jargon: Diffusion models = AI systems that generate images by learning to reverse a noise-adding process; manifold = mathematical structure representing the space of possible images; inversion = process of finding the noise pattern that would generate a given image.

Finding: Query-based committee selection using recursive candidate set splitting can approximate optimal voting outcomes while dramatically reducing the cognitive burden on voters compared to requiring full preference rankings.

Why it matters: This directly applies to team formation, collaborative decision-making processes, and designing user interfaces that minimize cognitive load while maintaining decision quality.

Method: Experimental evaluation across multiple election models comparing different query strategies and cost functions that model cognitive effort required for preference elicitation.

Jargon: Multiwinner voting rules - methods for selecting multiple winners/committee members from candidates; k-Borda - voting system where each voter ranks candidates and points are awarded based on position.

Finding: This study uses partial information decomposition to quantitatively analyze how large vision-language models actually combine visual and text information, revealing that models fall into two distinct processing strategies (fusion-centric vs. language-centric) and identifying when multimodal fusion is learned during training.

Why it matters: Understanding how AI systems internally process and combine different types of information is crucial for designing effective human-AI collaboration interfaces and predicting when these systems will complement vs. substitute human multimodal reasoning.

Method: Novel application of partial information decomposition framework to decompose model decisions into redundant, unique, and synergistic information components across 26 models and four datasets.

Jargon: Partial Information Decomposition (PID) - mathematical framework that separates shared information into redundant (overlapping), unique (source-specific), and synergistic (emergent from combination) components.

Finding: iPoster is an interactive framework that uses graph-enhanced diffusion models to generate poster layouts based on user-specified constraints, allowing designers to provide partial input (element categories, sizes, positions) and receive refined, context-aware layouts that respect those constraints.

Why it matters: This demonstrates how AI can augment human creativity in design workflows by maintaining user agency and creative control while providing intelligent assistance, relevant to both HCI design tools and human-AI collaboration in creative work.

Method: Uses a unified graph-enhanced diffusion architecture with masking strategies to preserve user constraints during denoising, plus cross content-aware attention to align elements with canvas regions.

Jargon: Diffusion models - generative AI models that create content by iteratively removing noise; denoising - the process of progressively refining generated content by removing artificial noise.

Finding: PRISM is a 270K-sample video dataset designed to train embodied vision-language models for retail environments, focusing on spatial knowledge, physical dynamics, and embodied action understanding rather than just visual recognition.

Why it matters: This addresses a key HCI challenge of how AI systems can better understand and interact with physical environments, directly relevant to human-AI collaboration and the design of embodied interfaces.

Method: Uses multi-view video data (egocentric, exocentric, 360°) from real supermarkets with structured evaluation across embodied reasoning, spatial perception, and intuitive physics capabilities.

Jargon: Embodied VLMs = vision-language models that can understand and act in physical spaces; SFT = supervised fine-tuning for training AI models on specific tasks.

Finding: Frontier LLMs (GPT-5, Claude) can match official legislative reasoning on standardized tasks but fail on novel political proposals, exhibiting "cascading bounded rationality" where cognitive limitations compound across human principals, AI agents, and evaluators.

Why it matters: This demonstrates how AI collaboration breaks down when moving from routine to creative/novel tasks, with implications for human-AI teamwork and decision-making in complex domains.

Method: Compared 6 LLMs against official Romanian legislative explanations using both LLM-as-Judge semantic scoring and programmatic text similarity metrics.

Jargon: Principal-agent theory - framework where one party (principal) delegates tasks to another (agent) under information asymmetry; Expuneri de motive - official explanatory memoranda for laws; Cascading bounded rationality - how cognitive limitations amplify across multiple decision-makers in a system.

Finding: EAGLE is a federated learning algorithm that ensures fairness by minimizing disparities in "loss gaps" - the difference between global model performance and each client's optimal local model performance - rather than just equalizing raw performance across clients.

Why it matters: This addresses key issues in distributed AI collaboration where participants have different data and goals, relevant to understanding fairness and motivation in human-AI collaborative systems.

Method: Proposes a regularization approach with theoretical convergence guarantees and introduces a novel heterogeneity measure to characterize performance relative to standard federated learning.

Jargon: Loss gap = difference between global model performance and best possible local model performance for each client; Federated learning = distributed machine learning where multiple parties train a shared model without sharing raw data.

Finding: This paper presents a hybrid edge-based video analysis system that combines privacy-aware skeleton detection with semantic AI models for public safety, focusing on system-level design trade-offs rather than new algorithms.

Why it matters: Demonstrates practical HCI considerations around AI deployment including privacy, latency, and human-AI collaboration in safety-critical applications.

Method: Demonstrator-based evaluation comparing skeleton-based motion analysis versus vision-language models on GPU-enabled edge devices under realistic constraints.

Jargon: Edge computing = processing data locally on devices rather than cloud servers; Zero-shot reasoning = AI making decisions about scenarios it wasn't specifically trained on.

Finding: Instructors, teaching assistants, and students have significantly misaligned perceptions of what constitutes academic dishonesty in computing courses, with instructors attributing cheating to laziness while students cite knowledge gaps and time pressures as primary motivations.

Why it matters: These perception misalignments reveal critical communication breakdowns that affect learning environments, motivation systems, and collaborative relationships between instructors and students.

Method: Survey study with 6 instructors, 21 TAs, and 538 undergraduate students evaluating cheating scenarios and reporting motivations through open-ended responses.

Jargon: Academic dishonesty - violations of academic integrity policies including plagiarism, unauthorized collaboration, and use of prohibited resources.

Finding: The paper argues that contradictions are emerging between the traditional ideology of computing as empowerment and current AI development driven by corporate-university partnerships, questioning who actually benefits from AI technologies.

Why it matters: This ideological analysis is relevant for understanding how AI impacts work and society, particularly regarding power dynamics in human-AI collaboration and the broader social implications of AI systems.

Method: Theoretical analysis using ideology as a conceptual lens to examine AI development practices.

Jargon: Ideology here refers to systems of ideas that represent and shape social, economic, and political relationships in technology development.

Finding: An AR system using HoloLens 2 achieved markerless surface registration for surgical guidance, reducing margin relocalization error from 14.2mm (verbal guidance) to 3.2mm (AR guidance) in tests with medical trainees.

Why it matters: This demonstrates how AR interfaces can significantly improve precision in high-stakes professional tasks, providing insights for HCI design in expert domains and spatial user interfaces.

Method: Comparative study with 6 medical trainees using silicone phantoms, comparing markerless AR registration against marker-based baseline and AR guidance against verbal communication.

Jargon: Target registration error - measure of spatial accuracy between virtual and real objects in AR; markerless registration - AR tracking without physical reference markers.

Finding: This study examines the security vulnerabilities, cyber threats, and unintended consequences that emerge when AI systems are deployed across industrial IoT networks and edge-cloud computing environments, distinguishing between software-level and infrastructure-level risks.

Why it matters: Understanding AI system vulnerabilities and unintended consequences is crucial for teaching about responsible AI design, human-AI collaboration challenges, and the broader societal impacts of AI deployment.

Jargon: IIoT (Industrial Internet of Things) - networked industrial devices and sensors; Edge-Fog-Cloud continuum - distributed computing architecture from local devices to remote servers.

Finding: This paper develops a computational framework linking cognitive behavioral therapy interventions to reinforcement learning principles, proposing three hierarchical layers that capture different learning and decision-making aberrations in psychiatric conditions.

Why it matters: The framework offers insights into how computational models can inform behavioral interventions and decision-making processes, relevant to understanding cognitive biases and designing better human-AI collaboration systems.

Jargon: Meta-cognition refers to thinking about thinking; reinforcement learning is a machine learning approach where agents learn through trial and error feedback.

Finding: Monte Carlo simulation studies used to evaluate statistical methods in psychology and behavioral sciences lack systematic attention to repeatability (replication, reproduction, robustness), despite their foundational role in determining which analytical approaches are adopted across fields.

Why it matters: This directly impacts behavioral science research quality and decision-making about statistical methods, which affects how we study cognitive biases, learning, and other core topics in these domains.

Method: Theoretical framework combining repeatability terminology with Monte Carlo simulation stages to identify verification needs.

Jargon: Monte Carlo simulations - computational studies that use random sampling to evaluate how statistical methods perform under different conditions.

Finding: Lonely individuals with higher resting parasympathetic nervous system activity show more flexible approach-avoidance behavior, being less likely to approach negative stimuli while maintaining normal approach to positive stimuli, suggesting better regulatory capacity for managing loneliness-driven motivations.

Why it matters: This provides insights into individual differences in decision-making and motivation that could inform design of social technologies and interventions for supporting collaboration and team dynamics.

Method: 107 participants completed an approach-avoidance task with positive/negative social and non-social images while measuring resting parasympathetic activity and loneliness levels.

Jargon: Parasympathetic nervous system (PNS) - part of the autonomic nervous system associated with rest, recovery, and regulatory capacity; approach-avoidance - behavioral tendency to move toward or away from stimuli.

Finding: This scoping review analyzes 891 papers on smart hospital ecosystems, identifying three key clusters (AI-driven healthcare, privacy-preserving digital health, and cloud-edge infrastructure) and revealing gaps in interoperability, real-world implementation, and governance.

Why it matters: The human-AI collaboration aspects and decision-making frameworks in healthcare contexts provide some insights for HCI and behavioral science research, though the focus is primarily on technical infrastructure rather than human factors.

Jargon: ScoRBA = Scoping Review with Bibliometric Analysis; ESD = Enhanced Strategic Diagram; PAGER = framework linking Patterns, Advances, Gaps, Research directions, and Evidence-based policy implications.

Finding: ScoringBench introduces a benchmark for evaluating tabular foundation models using proper scoring rules (like CRPS, Energy Score) rather than just point estimates, showing that model rankings vary significantly depending on the evaluation metric chosen.

Why it matters: This highlights how evaluation choices can bias our understanding of AI system performance, which is relevant for behavioral science research on decision-making biases and AI impact assessment.

Method: Compares multiple tabular foundation models (TabPFN, TabICL) across regression benchmarks using comprehensive scoring rules that better capture probabilistic forecast quality than traditional RMSE/R2 metrics.

Jargon: CRPS = Continuous Ranked Probability Score; TabPFN/TabICL = tabular foundation models; proper scoring rules = metrics that evaluate full probability distributions rather than just point predictions.

Finding: This study develops an optimization framework for blood donation centers to strategically invite donors to collection sessions, balancing supply-demand matching with donor convenience and safety constraints.

Why it matters: While primarily focused on healthcare logistics, it demonstrates decision support system design and addresses behavioral considerations like donor fatigue from over-solicitation.

Method: Compares binary integer linear programming against a greedy heuristic using real data from Portuguese blood services across 4-month periods.

Jargon: BILP = Binary Integer Linear Programming (optimization method using 0/1 variables)

Finding: This paper proposes a hierarchical multi-agent reinforcement learning architecture (SDA-MARL) for coordinating multiple autonomous underwater vehicles in target tracking, addressing challenges of non-stationarity, sparse rewards, and environmental robustness through supervised diffusion models and structured experience replay.

Why it matters: While focused on underwater robotics, it contributes to understanding multi-agent coordination and AI collaboration challenges that apply broadly to human-AI teamwork and distributed decision-making systems.

Method: Combines hierarchical MARL architecture with supervised diffusion models and behavioral cloning to improve multi-agent coordination and learning efficiency.

Jargon: MARL = Multi-Agent Reinforcement Learning; AUV = Autonomous Underwater Vehicle; non-stationarity = when the learning environment changes as other agents update their policies.

Finding: The paper proposes a "Monodense" deep neural network to estimate price elasticity from retail transaction data, showing superior performance compared to other machine learning methods in predicting consumer demand responsiveness to price changes.

Why it matters: Understanding consumer response to pricing could inform behavioral science research on decision-making and economic behavior, though the connection to core HCI/design topics is limited.

Method: Tests three ML approaches (including the novel Monodense neural network, double machine learning, and gradient boosting) on millions of retail transactions using backtesting validation.

Jargon: Price elasticity - measure of how much consumer demand changes when prices change; Monodense - the authors' hybrid neural architecture combining different layer types.

Finding: This paper develops algorithms for the ski rental problem (deciding when to buy vs. continue renting) that use distributional predictions rather than point estimates, with both robust performance guarantees and consistency when predictions are accurate.

Why it matters: The decision-making framework could inform research on behavioral biases in sequential choices and algorithmic advice systems, though the connection to HCI/behavioral science applications is indirect.

Method: Combines theoretical algorithm development with experimental validation across different probability distributions (Gaussian, geometric, bi-modal).

Jargon: Ski rental problem = classic online decision model about rent-vs-buy timing; competitive ratio = algorithm performance measure comparing to optimal offline solution; learning-augmented = algorithms that use predictive advice.