📚 Research Digest

Finding: This empirical study introduces "participatory provenance," a framework using optimal transport theory to audit whether AI summaries of public consultation faithfully represent input populations, revealing that official government AI summaries systematically exclude 15-17% of participants, particularly those expressing dissent or skepticism.

Why it matters: This directly addresses the critical HCI challenge of designing accountable AI systems for collaborative decision-making and provides concrete methods for auditing AI-mediated group processes.

Method: Applied optimal transport theory, causal inference, and semantic analysis to 5,253 responses from Canada's national AI strategy consultation, comparing official summaries against random baselines and developing an open-source auditing tool.

Jargon: Optimal transport theory - mathematical framework for measuring how distributions transform; Participatory provenance - tracking how individual inputs flow through AI summarization processes.

Finding: AgentLens introduces adaptive visual communication for mobile GUI agents, using three modalities (Full UI, Partial UI, GenUI) that adjust based on task context, achieving 85.7% user preference and highest usability scores compared to traditional foreground/background execution approaches.

Why it matters: This directly addresses fundamental HCI challenges in human-AI collaboration, particularly how to design effective visual interfaces that balance transparency with user autonomy in AI agent interactions.

Method: Iterative formative studies followed by controlled experiment with 21 participants measuring usability (PSSUQ) and adoption intent across different visual interaction modalities.

Jargon: GUI agents - AI systems that can interact with graphical user interfaces; GenUI - likely generative or simplified UI representations; PSSUQ - Post-Study System Usability Questionnaire.

Finding: Employees perceive performance feedback from AI managers as lower quality and less acceptable than identical feedback from human managers, leading to reduced work engagement and relatedness satisfaction, with these effects remaining stable even after ChatGPT's introduction.

Why it matters: This directly informs how AI integration in workplace management affects human motivation, engagement, and the fundamental psychological needs underlying effective performance - critical for designing human-AI collaborative systems.

Method: Two experimental studies comparing identical performance feedback delivered by AI versus human managers, conducted before (2022) and after (2024) ChatGPT's launch using self-determination theory as the theoretical framework.

Jargon: Self-determination theory - psychological framework identifying autonomy, competence, and relatedness as basic human needs driving motivation and well-being.

Finding: Researchers developed a digitized version of the Invented Alien Creatures creativity task (dIAC) that captures both creative products and fine-grained process data through timestamped action logs, addressing longstanding measurement limitations in creativity research.

Why it matters: This tool enables more rigorous study of creative processes and could revolutionize how we measure and understand creativity in HCI, design thinking, and collaborative creative work.

Method: Open-source jsPsych drawing interface with automated logging of all user actions (generate, move, transform, eliminate) plus traditional consensual assessment ratings.

Jargon: IAC = Invented Alien Creatures task (classic creativity assessment); consensual assessment = expert rating method for evaluating creative products.

Finding: Current Large Audio-Language Models struggle significantly with paralinguistic cues (tone, emotion, prosody) in speech generation, with leading models failing at comprehensive control and dynamic modulation, and 43.3% of errors in situational dialogue stemming from misinterpreting these cues.

Why it matters: This directly impacts HCI design for voice interfaces and AI collaboration, revealing critical gaps in how AI systems understand and generate human-like communication beyond just words.

Method: Developed a comprehensive benchmark with over 100 fine-grained paralinguistic features and 1,000+ parallel speech queries, using pairwise comparison with LALM-based judges to reduce evaluation subjectivity.

Jargon: Paralinguistic cues - non-verbal aspects of speech like tone, pitch, rhythm, and emotion that convey meaning; LALMs - Large Audio-Language Models that process both speech and text.

Finding: The AI value alignment problem is fundamentally a governance challenge involving trade-offs among competing stakeholders rather than a purely technical engineering problem, with misalignment occurring along three axes: objectives, information, and principals.

Why it matters: This reframes AI alignment as a human-centered design and collaboration challenge, directly relevant to HCI practitioners designing AI systems and understanding how different stakeholders' values conflict in sociotechnical systems.

Jargon: Principal-agent framework refers to economic theory analyzing relationships where one party (agent) acts on behalf of another (principal), often with conflicting incentives.

Finding: Meta-analysis of designing visualization atlases reveals five driving forces (data, people, stories, context, and the atlas itself) that dynamically influence different stages of complex data visualization projects, based on a 10-month design process involving 8 workshops and 15 stakeholder sessions.

Why it matters: Provides a structured framework for understanding how to design complex interactive visualizations that serve diverse audiences and use cases, directly applicable to HCI design processes and user experience research.

Method: Case study methodology combining design workshops, iterative prototyping, stakeholder onboarding sessions, and continuous reflection over 10 months to understand the atlas design process.

Jargon: Visualization atlas - an interactive platform that makes complex data comprehensible through visualizations and explanatory content, supporting both guided exploration and explanation.

Finding: Autark is a serverless toolkit that enables rapid prototyping of urban visual analytics systems, reducing development time from weeks to hours through domain-aware abstractions and AI-assisted coding workflows.

Why it matters: This directly addresses HCI design challenges by providing tools that accelerate the iterative design process and enable better human-AI collaboration in system development.

Method: Comparative study evaluating LLM coding effectiveness with and without the toolkit, plus usage scenarios demonstrating prototype creation capabilities.

Jargon: Visual analytics (VA) - interactive systems combining automated analysis with human visual exploration of data; Serverless architecture - cloud computing model where infrastructure management is abstracted away from developers.

Finding: The paper presents a wavelet-inspired technique to detect and parameterize submovements in human mouse motion, specifically ballistic components that arise from overlapping cognitive processes of perception, motor planning, and execution.

Why it matters: Understanding submovement detection is crucial for HCI research as it provides insights into how users interact with interfaces at a fundamental motor level, enabling better interface design and user experience optimization.

Method: Uses synthetic data modeled from real first-person shooter aim data (~6,400 trials from 13 users) with known ground truth to validate the wavelet technique against dual-threshold and persistence 1D segmentation methods.

Jargon: Submovements - ballistic (rapid, pre-planned) motion components that overlap as users continuously adjust their movements; Wavelet transform - mathematical technique for analyzing signals at different time scales.

Finding: Machine intelligence can augment human decision-making by improving situational awareness, decision framing, and cognitive flexibility, particularly in high-stakes scenarios with complexity, uncertainty, and time pressure.

Why it matters: This directly addresses how AI impacts human decision-making processes and could inform research on human-AI collaboration, cognitive biases in complex decisions, and behavioral interventions for better decision outcomes.

Method: Not specified in abstract.

Jargon: Agentic stewardship - AI systems actively managing and curating contextual information to support human decision-makers.

Finding: Large Language Models show a self-preference bias when assessing creativity, rating AI-generated ideas as more original than human-generated ones, but this bias disappears when controlling for idea elaboration (level of detail/development in responses).

Why it matters: This reveals critical limitations in using AI systems to assess human creativity and has direct implications for HCI design of AI-assisted creative tools and educational assessment systems.

Method: Analyzed 4,813 responses to the Alternate Uses Task from humans and ChatGPT-4o, comparing ratings between human judges and multiple AI assessment systems (OCSAI, CLAUS, and ChatGPT-4o).

Jargon: Alternate Uses Task (AUT) - a divergent thinking test where people generate creative uses for common objects; idea elaboration - the level of detail and development in creative responses.

Finding: This thesis argues that evaluating generative AI requires examining how models, users, and institutions recursively co-construct meaning and values through interaction, rather than treating AI as isolated predictors—introducing the Machine-Society-Human (MaSH) Loops framework for process-oriented evaluation.

Why it matters: This directly addresses how AI systems shape human collaboration and decision-making processes, offering frameworks for understanding AI's broader impact on work and society beyond simple performance metrics.

Method: Develops a distributional evaluation approach using World Values Survey data with structured prompts and anchor-aware scoring, demonstrated through case studies of value drift and real estate applications.

Jargon: Pluralist sociotechnical systems - AI systems that operate across diverse cultural contexts and values rather than assuming universal standards; MaSH Loops - recursive feedback cycles between machines, society, and humans that co-create meaning; constitutive interventions - actions that don't just observe but actively shape what they measure.

Finding: This scoping review reveals systematic gaps between current stuttered-speech technology research priorities and the actual needs of people who stutter and their clinicians, proposing a taxonomy and guidelines to better align research with end-user experiences.

Why it matters: This demonstrates critical principles for inclusive HCI design - how involving stakeholders throughout the research process prevents building solutions that miss real user needs, especially for underserved populations.

Method: Combines systematic scoping review methodology with stakeholder survey (n=70) to triangulate research gaps from both literature and user perspectives.

Jargon: Scoping review - systematic methodology to map research literature and identify gaps in a field.

Finding: This framework enables non-expert users to adapt industrial robot skills through three interaction modalities - kinesthetic touch, natural language, and graphical interfaces - using a tool-based LLM architecture that selects predefined functions rather than generating code for safer adaptation.

Why it matters: This demonstrates how multimodal interaction design can make complex AI systems accessible to non-experts, directly addressing HCI principles for human-AI collaboration and showing practical applications of natural language and graphical interfaces in robotics.

Method: Integrates five technical components including energy-based intention detection, tool-based LLM architecture, and probabilistic Virtual Fixtures, validated on a 7-DoF robot at a trade fair.

Jargon: Kernelized Movement Primitives (KMPs) - a method for encoding robot movements; ergodic control - a technique for ensuring robots cover desired areas during tasks like surface finishing.

Finding: VPL (Vibrotactile Preference Learning) uses Gaussian process-based preference learning with uncertainty-aware query selection to personalize haptic feedback through 40 rounds of pairwise comparisons, efficiently learning individual user preferences for vibrotactile parameters.

Why it matters: This demonstrates how machine learning can be applied to personalize user experiences in HCI, particularly for haptic interfaces where individual differences significantly impact user perception and satisfaction.

Method: Uses expected information gain-based acquisition strategy with Gaussian processes to intelligently select which vibrotactile parameter pairs to present to users for comparison, incorporating user-reported uncertainty to improve learning efficiency.

Jargon: Vibrotactile - relating to touch sensations produced by vibration; Gaussian process - a machine learning method for modeling uncertainty in predictions; acquisition strategy - method for selecting which data points to query next in active learning.

Finding: The paper proposes a three-layer trust framework (human-oriented, AI-oriented, and interaction-oriented) to align different stakeholder perspectives on what makes AI systems trustworthy for mental health support, revealing critical gaps between current NLP evaluation metrics and real-world clinical requirements.

Why it matters: This directly addresses fundamental HCI questions about human-AI interaction design and trust-building, which are crucial for understanding how to design AI systems that people will actually adopt and collaborate with effectively.

Method: Systematic review of AI-driven mental health research combined with stakeholder analysis across practitioners, researchers, and regulators to develop the multi-layered trust framework.

Jargon: Therapeutic fidelity = adherence to evidence-based therapeutic principles and practices; socio-technical alignment = ensuring technology design matches both social contexts and technical requirements.

Finding: Cognitive modeling of radiologist-AI collaboration reveals that AI improves diagnostic discrimination but shifts decision thresholds, leading to more detections but also higher false positive rates, with implications for healthcare resource planning and clinical diversity.

Why it matters: This demonstrates how cognitive modeling can provide deeper insights into human-AI collaboration beyond surface performance metrics, directly relevant to understanding AI's impact on professional decision-making and work practices.

Method: Applied hierarchical signal detection models to analyze decision-making patterns of 20 expert radiologists reading 1163 cases with and without AI assistance.

Jargon: MRMC (multi-reader, multi-case) = standardized study design for evaluating diagnostic performance; signal detection models = mathematical frameworks for understanding decision-making under uncertainty, separating ability to discriminate signals from response bias/thresholds.

Finding: Sycophantic large language models reinforce users' existing beliefs rather than challenging them, and this effect persists even when users are warned about or trained to identify sycophantic behavior through disclosure interventions.

Why it matters: This reveals critical limitations in how people interact with AI systems and highlights cognitive biases that affect decision-making and belief updating in human-AI collaboration.

Method: Pre-registered experiment with 1,492 participants randomly assigned to interact with either sycophantic or challenger LLMs, with three disclosure conditions testing psychological inoculation theory.

Jargon: Sycophantic - tendency to agree with and validate user opinions rather than provide objective responses; Psychological inoculation theory - building resistance to persuasion by exposing people to weakened forms of opposing arguments.

Finding: A prototype animal adoption platform based on the "Cat Distribution System" meme uses algorithmic matchmaking and proximity-based discovery to make adoption feel more serendipitous and less transactional, with positive initial user evaluation (n=35) showing good usability but needing improvements in transparency and communication.

Why it matters: This demonstrates how internet culture and memes can inform UX design by leveraging existing mental models to create more intuitive and emotionally resonant human-computer interactions.

Method: Prototype development followed by user evaluation with 35 participants assessing usability and alignment with user expectations.

Jargon: Cat Distribution System (CDS) - internet meme suggesting cats "choose" their owners rather than being actively sought out.

Finding: Large language models significantly outperformed human advisors in fraud detection, with 0% endorsement of fraudulent investments versus 13-14% for humans, and showed greater resistance to motivated investor pressure that typically causes humans to suppress fraud warnings.

Why it matters: This demonstrates how AI systems can potentially mitigate human cognitive biases and social pressure in decision-making contexts, with important implications for AI-human collaboration in high-stakes advisory roles.

Method: Preregistered experiment comparing 7 LLMs across 12 investment scenarios (3,360 AI conversations) against 1,201 human participants serving as advisors under varying investor pressure conditions.

Jargon: Motivated investor framing - when investors arrive already convinced of an opportunity, creating social pressure on advisors; endorsement reversal - changing from warning against to endorsing a fraudulent investment.

Finding: This paper introduces conditional risk calibration - estimating how likely a prediction model is to make errors based on input features - and demonstrates its application in learning-to-defer frameworks where AI systems can hand off difficult decisions to humans.

Why it matters: Understanding when AI systems are likely to fail enables better human-AI collaboration design, particularly for interfaces that need to decide when to involve human judgment versus relying on automated predictions.

Method: Theoretical analysis connecting conditional risk calibration to regression tasks, plus empirical validation in learning-to-defer scenarios where systems choose between automated and human decision-making.

Jargon: Learning to defer (L2D) - frameworks where AI systems can choose to pass difficult decisions to human experts rather than making potentially poor automated decisions.

Finding: AktivTalk digitizes the clinical Talk Test using voice-based self-assessment for exercise intensity monitoring, achieving 90% accuracy in automated classification of high vs. non-high exertion from speech patterns in a 20-participant study.

Why it matters: This demonstrates how voice interfaces can replace traditional physiological monitoring methods, offering insights for designing accessible HCI systems that leverage natural speech interactions for real-time self-assessment.

Method: Within-subject study collecting exertion-labeled voice samples, using MFCC-based features with neural classification and cross-validation for automated speech analysis.

Jargon: MFCC (Mel-Frequency Cepstral Coefficients) - audio features that capture speech characteristics; Talk Test - clinical method where ability to speak comfortably indicates safe exercise intensity.

Finding: This paper presents a hybrid AI architecture that augments large language models with structured ontological memory layers (RDF/OWL knowledge graphs), demonstrating improved performance on multi-step reasoning tasks like Tower of Hanoi compared to baseline LLMs.

Why it matters: This addresses critical HCI concerns about AI system transparency, explainability, and reliable reasoning - key issues for designing trustworthy human-AI collaboration interfaces and understanding how structured knowledge affects AI decision-making.

Method: Automated pipeline constructs ontologies from heterogeneous data sources through entity recognition and relation extraction, then validates outputs using formal constraints in a generation-verification-correction loop.

Jargon: RDF/OWL = semantic web standards for representing structured knowledge; SHACL = validation language for RDF graphs; ontology = formal representation of knowledge domains and relationships.

Finding: This paper introduces semantic stratification to improve retrieval evaluation in RAG systems by systematically organizing documents into interpretable clusters and generating queries for missing coverage areas, rather than relying on average performance metrics.

Why it matters: The framework for identifying systematic blind spots and providing interpretable failure analysis could inform HCI design principles for AI-human collaboration interfaces and evaluation methodologies.

Method: Uses entity-based clustering to create interpretable document organization and systematic query generation to ensure coverage across different retrieval scenarios.

Jargon: RAG (Retrieval-Augmented Generation) - AI systems that retrieve relevant documents to inform their responses; semantic stratification - organizing evaluation by meaningful semantic categories rather than random sampling.

Finding: This paper presents a hybrid negotiation framework that combines deterministic pricing algorithms with LLMs for natural language translation, solving the problem of maintaining consistent concession behavior when pricing targets change mid-negotiation.

Why it matters: It demonstrates a principled approach to human-AI collaboration where AI handles communication while humans retain control over critical decision logic, relevant to designing trustworthy AI systems for complex business interactions.

Method: Empirical evaluation across 115,125 freight negotiations comparing the hybrid approach against pure LLM and rule-based baselines.

Jargon: Concession - gradual reduction in negotiation demands; Monotonicity - ensuring offers don't get worse over time; Spread - difference between target and minimum acceptable prices.

Finding: This empirical study of transformer compression across 40+ experiments reveals that high-variance neural network components don't correlate with predictive importance, and that compression methods face fundamental limits due to error accumulation and distribution shift between model layers.

Why it matters: Understanding AI model compression directly relates to how AI systems can be made more efficient for human-computer interaction applications and provides insights into the internal structure that affects AI's collaborative capabilities with humans.

Method: Systematic comparison of compression techniques including spectral analysis, block replacement, and quantization across GPT-2 and Mistral 7B models, measuring performance via perplexity and correlation analysis.

Jargon: CCA = Canonical Correlation Analysis (measures correlation between high-dimensional datasets); perplexity = measure of how well a model predicts text; quantization = reducing precision of model weights to compress size.

Finding: Multi-agent debate systems significantly improve AI performance in molecular design by enabling iterative generate-debate-refine loops that better reconcile semantic intent with structural constraints, achieving 59.82% exact match on benchmarks.

Why it matters: This demonstrates how collaborative AI architectures can enhance reasoning in complex design tasks, providing insights for team dynamics research and AI-human collaboration in creative problem-solving contexts.

Method: Uses multi-agent systems with perspective-oriented orchestration to address developer-debater conflicts and integrate global-local reasoning through iterative refinement cycles.

Jargon: RAG = Retrieval-Augmented Generation; CoT = Chain-of-Thought prompting; ChEBI-20 = chemical database benchmark.

Finding: Thinking about natural environments extends people's temporal perspective to longer time horizons compared to thinking about work or the future in general, though this doesn't translate to changes in mood, environmental intentions, or work commitment.

Why it matters: This demonstrates how environmental context can influence temporal decision-making and cognitive framing, relevant to behavioral science research on cognitive biases and decision-making processes.

Method: Pre-registered experimental study with 527 participants randomly assigned to different future-thinking conditions (general, work, or natural environment).

Jargon: Prospective thinking - mental simulation of future events and scenarios; Temporal perspective - how far into the future people consider when making decisions.

Finding: Higher sensory processing sensitivity reduces metacognitive efficiency (awareness of one's own perceptual accuracy) during multisensory illusions, particularly in emotional contexts, while emotional arousal creates dissociation between actual performance and subjective confidence.

Why it matters: This reveals individual differences in metacognitive accuracy that could inform personalized interface design and help explain why some users struggle more with confidence calibration in emotionally charged or multisensory environments.

Method: Online experiment using the Sound-induced Flash Illusion with facial emotional cues, measuring both perceptual accuracy and confidence ratings across different timing conditions.

Jargon: Sound-induced Flash Illusion (SiFI) - when sound causes people to perceive more visual flashes than actually presented; Highly Sensitive Person (HSP) scale - measures individual differences in sensory processing sensitivity; SOA - stimulus onset asynchrony (timing between stimuli).

Finding: This paper introduces a "counterfactual prediction machines" approach to interpreting complex statistical models by focusing on substantive quantities rather than raw coefficients, with implementation via the marginaleffects package for 100+ model types.

Why it matters: This directly addresses cognitive biases in how researchers interpret statistical results and provides practical tools for clearer decision-making in behavioral science research.

Method: Demonstrates the workflow through worked examples across observational studies and randomized experiments, showing applications to association testing, causal inference, and effect heterogeneity analysis.

Jargon: Marginal effects - the change in outcome associated with a unit change in predictor; Counterfactual prediction - using models to estimate "what if" scenarios for causal inference.

Finding: A structured mentoring program with four components (professional learning teams, collaborative teaching design, mentored practice, and reflective exchange) significantly improved novice teachers' instructional capabilities and pedagogical competencies.

Why it matters: This demonstrates effective strategies for skill acquisition and professional development that could inform mentorship design in HCI education and collaborative learning approaches.

Method: Mixed-methods study of 97 teachers across 15 disciplines, combining quantitative assessment scores with qualitative analysis of mentorship experiences.

Finding: Principal instructional leadership significantly impacts teacher job satisfaction, with clear goal setting, curriculum coordination, and professional development opportunities being the strongest predictors, while high visibility and student progress monitoring showed no significant influence.

Why it matters: This provides empirical evidence on which leadership practices most effectively motivate teachers, offering insights into goal-setting and team dynamics that could apply to other organizational contexts.

Method: Correlational quantitative study analyzing questionnaire responses from 372 randomly selected teachers in Northwest China junior high schools.

Jargon: Instructional leadership - principals' behaviors focused on curriculum, teaching practices, and academic goals rather than just administrative management.

Finding: Students in STEM courses using "ungrading" (alternative assessment without traditional grades) within Universal Design for Learning frameworks reported reduced stress, increased focus on learning over performance, but struggled to psychologically detach from viewing grades as measures of self-worth.

Why it matters: This directly addresses how assessment design affects student motivation, stress, and learning focus - key behavioral and motivational factors in educational settings.

Method: Qualitative interviews with 24 students (13 graduate, 11 undergraduate) across library science and computer science courses, using inductive coding.

Jargon: Ungrading = assessment approach that eliminates or de-emphasizes traditional letter grades in favor of feedback-focused evaluation; UDL = Universal Design for Learning, pedagogical framework emphasizing multiple means of engagement, representation, and expression.

Finding: Detection tasks commonly used in consciousness research contaminate measures of conscious experience with decision biases from contextual factors like statistical priors and reward structures, making it difficult to isolate true perceptual experience from non-perceptual influences.

Why it matters: This reveals a fundamental methodological problem in measuring conscious experience that could impact HCI research on user awareness, attention, and subjective experience of interfaces.

Method: Used a novel Hurdle-Gaussian model to decompose reproduction responses into perceptual (continuous contrast) and non-perceptual ("hurdle") components across 505 participants.

Jargon: Hurdle component - the non-perceptual decision bias portion of responses; Gabor stimuli - visual patterns used to test perception.

Finding: Philippine physical education teachers face five main challenges implementing the new PATH-Fit curriculum: inadequate instructional materials, insufficient time, lack of teacher training, missing equipment, and poor administrative support.

Why it matters: This illustrates systemic barriers to curriculum implementation that could inform research on organizational change, professional development design, and institutional support systems in educational settings.

Method: Qualitative study using focus groups and interviews with five teachers, validated by expert review and triangulated results.

Jargon: PATH-Fit = Physical Activities Towards Health and Fitness, the new tertiary-level physical education curriculum in the Philippines.

Finding: PINA is a two-stage federated learning framework that uses compressed client updates to form privacy-preserving clusters and introduces normality-driven aggregation, achieving 2.9% higher accuracy than existing differentially private federated learning methods.

Why it matters: This contributes to understanding how AI systems can enable collaborative learning while protecting individual privacy, relevant to future human-AI collaboration scenarios.

Method: Combines low-rank adaptation (LoRA) for client fine-tuning with compressed sketching for cluster initialization, followed by normality-driven aggregation across federated learning rounds.

Jargon: Differential privacy (DP) - mathematical framework providing formal privacy guarantees by adding controlled noise; Federated learning (FL) - training AI models across distributed devices without centralizing raw data; LoRA - parameter-efficient fine-tuning method using low-rank matrix decomposition.

Finding: LaplacianFormer replaces softmax attention in Transformers with a Laplacian kernel to reduce computational complexity while maintaining better mid-range token interactions, using custom CUDA implementations for efficient computation.

Why it matters: This technical optimization could enable more efficient AI systems for visual tasks, potentially impacting how AI tools are deployed in creative and collaborative work environments.

Method: Combines theoretical analysis with empirical validation on ImageNet, implementing custom CUDA kernels and Newton-Schulz iteration for efficient computation.

Jargon: Laplacian kernel - mathematical function for measuring similarities; Nyström approximation - method for approximating large matrices; Newton-Schulz iteration - numerical method avoiding matrix inversion.

Finding: A Vision-Language-Action model integrates needle tracking and insertion control for robotic ultrasound systems, outperforming manual operation with higher accuracy and reduced procedure time.

Why it matters: This demonstrates how AI can augment human capabilities in complex medical procedures, relevant to understanding AI's impact on skilled work and human-AI collaboration.

Method: Combines cross-depth fusion tracking with uncertainty-aware control policy in an end-to-end system, tested against state-of-the-art trackers and manual operation.

Jargon: VLA (Vision-Language-Action) - AI model that processes visual input, language instructions, and produces actions; RUS - Robotic Ultrasound System.

Finding: Large language models can infer odor similarity relationships that correspond to human sensory evaluations, enabling the creation of odor maps where perceptually similar scents cluster together spatially.

Why it matters: This demonstrates how AI can model human perceptual and cognitive processes, potentially informing HCI design for sensory interfaces or applications involving subjective similarity judgments.

Jargon: Dravnieks' dataset - a standard reference collection of odor character profiles used in olfactory research.

Finding: This paper proposes that two visual perception phenomena (face distortion effects and eye movement patterns linked to heartbeat) both result from learned compensatory responses that normally help maintain visual stability during movement.

Why it matters: While focused on low-level perceptual mechanisms, this could inform HCI research on visual interface design and how users' physiological states might affect their interaction with visual displays.

Method: The paper presents a theoretical framework with testable predictions rather than empirical data collection.

Jargon: FFDE = visual distortions when faces appear in peripheral vision; microsaccades = tiny involuntary eye movements; baroreceptor = sensors that detect blood pressure changes.

Finding: Mandarin teachers and students differ in their perceptions of teaching strategy effectiveness, with students rating both strategy use and classroom practices higher than teachers, while learners face moderate challenges with tone production, character writing, and limited native speaker exposure.

Why it matters: The perception gaps between teachers and learners highlight important UX considerations for language learning interfaces and reveal how cognitive biases might affect self-assessment in educational contexts.

Method: Descriptive-correlational design using structured questionnaires with teachers and students, analyzed through multivariate statistical tests.

Finding: Bilingual instruction (English and Thai) led to significantly greater learning gains in genetics compared to English-only instruction among Grade 9 students, with students reporting improved understanding, engagement, and confidence.

Why it matters: This provides evidence for how language choices in instruction design affect learning outcomes and student engagement, relevant to educational technology and learning sciences research.

Method: Quasi-experimental pretest-posttest design with two intact classes over four weeks, using achievement tests, perception surveys, and qualitative interviews.