📚 Research Digest

Finding: Designers collaborate with AI in sketch-based ideation through four distinct interaction modes (Human-Only, Human-Lead, AI-Lead, and Co-Evolution), with frequent dynamic shifts between human-led and AI-led roles rather than relying on a single mode throughout the design process.

Why it matters: This provides crucial empirical insights for designing AI-powered creative tools and understanding how human-AI collaboration actually unfolds in visual design thinking and sketching workflows.

Method: User study with 12 participants analyzing automatically recorded interaction logs and post-task interviews to identify patterns in human-MLLM collaborative sketching.

Jargon: MLLM = Multimodal Large Language Model (AI that can process both text and visual inputs like sketches).

Finding: Large Language Models systematically erase cultural and linguistic markers from non-native English varieties when used to "professionalize" text, with an overall Identity Erasure Rate of 10.26% while maintaining semantic similarity, creating a "Semantic Preservation Paradox."

Why it matters: This directly impacts how AI tools are reshaping workplace communication and human expression, revealing critical biases in AI systems that affect diverse users' linguistic identity and cultural representation.

Method: Analyzed 22,350 LLM outputs from 1,490 culturally marked texts across Indian, Singaporean, and Nigerian English using novel metrics (Identity Erasure Rate and Semantic Preservation Score) with five different models under three prompt conditions.

Jargon: Cultural Ghosting = systematic erasure of linguistic markers unique to non-native English varieties; Identity Erasure Rate (IER) = percentage of cultural markers removed; Semantic Preservation Score (SPS) = measure of meaning retention despite cultural marker changes.

Finding: Participatory annotation workshops using feminist epistemology can transform data work from extractive labor into collaborative knowledge-building, showing how tactical consensus and bounded context enable meaningful participation while disrupting traditional AI development hierarchies.

Why it matters: This demonstrates concrete methods for redesigning human-AI collaboration processes to be more equitable and participatory, directly relevant to HCI design and team dynamics in AI systems.

Method: Multilingual workshops with journalists and activists using iterative, participatory annotation of gender-violence news narratives, analyzed through feminist lens.

Jargon: Tactical consensus - working agreement that acknowledges different perspectives rather than forcing uniformity; bounded context - deliberately limiting scope to make pluralistic participation manageable.

Finding: A 7-month codesign process with 13 teachers revealed that LLM-assisted assessment authoring tools can shift teacher practices from generation to curation, enabling creation of formative assessments they wouldn't otherwise make while improving their reflection on assessment quality.

Why it matters: This demonstrates how AI can augment rather than replace human expertise in creative work, showing concrete ways AI tools can enhance professional workflows through thoughtful HCI design.

Method: Extended codesign methodology with teachers to develop both a conceptual model of assessment workflows and a working system, followed by user study validation with 15 additional teachers.

Jargon: Codesign - collaborative design process involving end users as design partners; formative assessments - evaluations used during learning to provide feedback rather than final grades.

Finding: A framework enables LLMs to dynamically adapt their personality during conversations using state machines and continuous personality scoring, successfully demonstrated in medical de-escalation training where the system both responds to and influences user behavior.

Why it matters: This directly addresses adaptive AI systems for education and HCI, showing how AI can be designed to respond dynamically to human interaction patterns rather than maintaining static personas.

Method: Uses state machines with transition probabilities adapted to conversational context, plus a modular personality scoring pipeline that works across different personality models and LLMs.

Jargon: Interpersonal Circumplex (IPC) - a psychological model mapping interpersonal behaviors on two axes (dominance-submission, hostility-friendliness); State machines - computational models that change behavior based on current state and inputs.

Finding: This paper reframes speculative design through information theory, proposing "epiplexity" as a measure to distinguish between meaningful design insights and superficial aesthetic surprise, with a practical framework for evaluating speculative design quality.

Why it matters: This directly advances HCI and design pedagogy by providing rigorous criteria for assessing speculative design work, helping educators teach more effective design thinking and future-oriented problem-solving.

Method: Theoretical framework development with a practical self-assessment questionnaire for designers to audit their speculative work.

Jargon: Epiplexity = structured, learnable information that can be extracted by observers with limited cognitive resources; Provotypes = provocative prototypes used in speculative design to explore "what if" scenarios.

Finding: Researchers developed AllyAAC, a wearable system that combines body-based gestures (unaided AAC) with speech-generating devices (aided AAC) to improve communication for people with motor impairments, using 18 months of participatory design and creating the first dataset of atypical gestures with over 600,000 data points.

Why it matters: This demonstrates how participatory design and adaptive AI can create more inclusive and personalized human-computer interfaces, particularly relevant for understanding accessibility, user-centered design, and AI's role in augmenting human capabilities.

Method: Used participatory design over 18 months with AAC users, followed by field study with 14 participants, and applied Transformer-based machine learning models to recognize personalized gestures.

Jargon: AAC (Augmentative and Alternative Communication) - technologies that help people with communication disabilities express themselves; IMU - Inertial Measurement Unit, a sensor that detects motion and orientation.

Finding: ViSTAR is an AR system that uses 3D motion reconstruction and LLM-generated coaching feedback to help basketball players practice skills independently, with users preferring AI feedback over human coach feedback and reporting improved awareness of posture and balance issues.

Why it matters: This demonstrates how AI can enhance human skill acquisition through personalized feedback and represents a novel approach to combining motion analysis with natural language coaching for learning complex physical skills.

Method: Mixed-methods evaluation with N=16 participants comparing AI-generated feedback to human coach feedback, using 3D motion capture data converted to natural language coaching cues via LLM.

Jargon: BST (Behavioral Skills Training) - a four-step framework: instruction, modeling, rehearsal, and feedback; spatio-temporal joint data - 3D position and movement data of body joints over time.

Finding: Large language models show inconsistent biases when choosing between human experts and algorithmic agents - they rate humans as more trustworthy when asked directly, but actually choose algorithms more often when making incentivized decisions, even when the algorithms perform worse.

Why it matters: This reveals critical inconsistencies in AI decision-making that affect human-AI collaboration and trust, especially relevant for understanding how AI systems will integrate into team dynamics and decision-making processes.

Method: Experimental study using behavioral economics paradigms with eight different LLMs, comparing stated preferences (direct trust ratings) versus revealed preferences (incentivized betting scenarios).

Jargon: Algorithm aversion - human tendency to avoid using algorithmic predictions even when they outperform human judgment.

Finding: This theoretical analysis identifies three dominant values in data visualization research—universality, objectivity, and efficiency—and argues for embracing more diverse, pluralistic values in visualization tools and practices.

Why it matters: This directly impacts HCI and UX design by questioning fundamental assumptions about how we design visual interfaces and data tools, while connecting to broader questions about cognitive biases in how we present and interpret information.

Jargon: Pluralistic values - incorporating multiple, potentially conflicting perspectives rather than adhering to single universal principles.

Finding: SPGen uses deep learning with unsupervised domain adaptation to predict human scanpaths (eye movement sequences) when viewing paintings, outperforming existing methods by transferring knowledge from natural scenes to artwork.

Why it matters: This directly advances HCI/UX design by providing better models of visual attention and gaze behavior, which are fundamental to interface design and user experience research.

Method: Uses a Fully Convolutional Neural Network with differentiable fixation selection, learnable Gaussian priors, and gradient reversal layers for domain adaptation from photographs to paintings.

Jargon: Scanpaths = sequences of eye movements and fixations; Domain adaptation = transferring learned knowledge from one type of data (photos) to another (paintings).

Finding: This study uses LLMs to automatically detect UX problems ("UX smells") in Visual Studio Code by analyzing user-reported issues on GitHub, finding that most problems cluster around informativeness, clarity, intuitiveness, and efficiency.

Why it matters: Demonstrates a scalable method for identifying systematic UX issues in complex software interfaces, directly applicable to HCI research and UX design practice for developer tools.

Method: Novel application of LLMs for automated UX issue classification combined with expert validation using an established UX taxonomy.

Jargon: UX smells - recurring patterns of poor user experience design, analogous to "code smells" in software engineering.

Finding: Romantic AI platforms systematically convert users' intimate emotional disclosures into reusable data assets through their privacy policies, using mechanisms like default training appropriation and ownership reconstruction that expand platform rights while shifting risks to users.

Why it matters: This reveals critical design and ethical considerations for AI systems that involve emotional disclosure, highlighting how interface design and policy frameworks can exploit human psychological vulnerabilities in intimate interactions.

Method: Policy analysis of Privacy Policies and Terms of Service across six Western and Chinese romantic AI platforms.

Jargon: Intimate history assetization = converting personal emotional exchanges into platform-owned data assets; Default training appropriation = automatically using user data for AI model training unless explicitly opted out.

Finding: List experiments (a social science method for measuring sensitive beliefs) successfully reveal that major LLMs harbor hidden approval for controversial practices like mass surveillance, torture, and discrimination that they wouldn't express when asked directly.

Why it matters: This demonstrates critical methodological insights for understanding AI systems' true beliefs versus their stated responses, essential for researchers studying AI impact on decision-making and human-AI collaboration.

Method: Adapted list experiments from social science research - participants estimate how many items from a list apply to them, with treatment groups seeing an additional sensitive item, allowing researchers to infer hidden attitudes without direct questioning.

Jargon: List experiment - indirect survey method that measures sensitive attitudes by comparing aggregate responses between control and treatment groups; Alignment faking - when AI systems give responses they think humans want to hear rather than their true "beliefs."

Finding: This paper introduces a computational framework showing how AI decision support can paradoxically lead to worse outcomes when human decision-makers have misaligned prior beliefs, demonstrating that providing AI predictions to users can backfire without proper training and model documentation.

Why it matters: This directly addresses the critical challenge of designing effective human-AI collaboration systems and understanding cognitive biases in AI-assisted decision making, core issues in HCI and behavioral science.

Method: Uses Bayesian causal inference modeling to simulate the two-step process of how AI predictions change human beliefs and subsequent decision outcomes.

Jargon: Bayesian agent - a decision maker who updates beliefs using probability theory based on new evidence; misaligned prior beliefs - existing assumptions that don't match reality.

Finding: An AR interface using 3D Gaussian Splatting enables disaster responders to visualize detailed scene reconstructions through lightweight interactions (World-in-Miniature navigation and semantic Points of Interest) while maintaining low cognitive load and supporting real-time coordination.

Why it matters: This demonstrates how thoughtful UX design can reduce cognitive burden in high-stakes scenarios, providing insights for designing interfaces that support decision-making under pressure and team coordination.

Method: User-centered design approach with preliminary user evaluation measuring usability and acceptance ratios of the AR interface.

Jargon: 3D Gaussian Splatting - a technique for creating detailed 3D scene reconstructions; World-in-Miniature (WIM) - a navigation method using miniature representations of the environment.

Finding: Flipping a stylus provides an intuitive mode switch between drawing and teleportation in VR, with gaze handling orientation while the stylus controls positioning, offering seamless integration without disrupting drawing workflows.

Why it matters: This demonstrates how to design natural interaction techniques that support creative workflows in VR without mode-switching friction, directly relevant to VR sketching tools and 3D design interfaces.

Method: User study comparing six teleportation techniques across two mode-switching methods (button vs. flip) and three orientation approaches in teleport-and-orient tasks.

Jargon: Teleportation - VR navigation method where users instantly move to selected locations rather than walking.

Finding: The paper introduces fEDM+, an enhanced fuzzy logic framework for ethical AI decision-making that provides principle-level explainability and validates decisions against multiple stakeholder perspectives rather than a single ethical standard.

Why it matters: This directly addresses critical HCI concerns about making AI systems' ethical reasoning transparent and accountable to users, which is essential for designing trustworthy human-AI collaboration systems.

Jargon: Fuzzy logic uses degrees of truth rather than binary true/false; Fuzzy Petri Nets are mathematical models for system verification; pluralistic validation means testing against multiple ethical frameworks simultaneously.

Finding: Large Vision-Language Models use different attention heads for truthfulness versus visual perception, and dynamic steering vectors can reduce hallucinations by selectively activating the right attention mechanisms based on semantic context during inference.

Why it matters: This reveals how AI systems can be made more reliable partners in human-computer interaction by reducing false information generation, which is crucial for understanding AI's impact on collaborative work and decision-making.

Method: Training-free intervention method that builds semantic-based steering vector databases and dynamically selects relevant vectors during inference based on input similarity.

Jargon: LVLMs = Large Vision-Language Models (AI systems that process both images and text); hallucination = when AI generates false or nonsensical information; attention heads = components in neural networks that focus on different aspects of input; steering vectors = mathematical tools to guide model behavior.

Finding: Women use conversations with LLMs as a catalyst to share gendered experiences on social media, with AI interactions helping them overcome both external fears (judgment, backlash) and internal barriers (self-imposed standards about what's appropriate to share publicly).

Why it matters: This reveals how AI tools can transform human self-expression and social sharing behaviors, demonstrating AI's impact on personal communication, social interaction design, and digital collaboration patterns.

Method: Semi-structured interviews with 20 women who engage in this practice on Xiaohongshu (Chinese social media platform).

Jargon: Xiaohongshu - popular Chinese social media platform; digital feminism - using digital platforms to advance feminist causes and women's voices.

Finding: LLMs are increasingly functioning as constant counselors for high-stakes personal decisions despite being technically stateless and lacking genuine subjectivity, creating "advisory intimacy without a subject" that feels personally supportive while having no accountable human agent behind it.

Why it matters: This directly impacts HCI design decisions around AI interfaces, reveals critical behavioral science insights about human-AI interaction patterns, and highlights how AI is fundamentally changing collaboration between humans and artificial agents.

Method: Theoretical analysis combining technical AI architecture review with social critique of contemporary loneliness and advisory relationships.

Jargon: Transformer architectures - the underlying technical structure of LLMs; stateless - models don't retain memory between interactions; advisory intimacy without a subject - feeling counseled by something that has no actual consciousness or responsibility.

Finding: WatchHand enables continuous 3D hand pose tracking on commercial smartwatches using only built-in speakers and microphones with inaudible acoustic waves, achieving 7.87mm accuracy across diverse real-world conditions without requiring additional hardware.

Why it matters: This breakthrough enables rich gestural interactions on ubiquitous devices, opening new possibilities for HCI research and UX design while demonstrating how existing hardware can be repurposed for novel interaction modalities.

Method: Uses frequency-modulated continuous acoustic waves reflected from hands, processed by deep learning models to estimate 20 finger joint positions, evaluated across multiple device models and usage conditions.

Jargon: Frequency-modulated continuous waves - acoustic signals that vary in frequency over time to capture detailed reflection patterns from moving objects.

Finding: The paper introduces "symmetrical reality" as a theoretical framework for symbiotic human-AI coexistence that spans both physical and virtual worlds, moving beyond traditional human-centric interaction paradigms.

Why it matters: This framework directly addresses how AI's evolution is fundamentally changing human-computer interaction design and collaborative work relationships, which is central to understanding AI's impact on human collaboration and UX design.

Method: Theoretical framework development rather than empirical study.

Jargon: Symmetrical reality - a proposed framework where humans and AI entities coexist as equal partners across physical and virtual environments, rather than in traditional hierarchical human-machine relationships.

Finding: This system uses smartwatch sensors to automatically detect social interactions in real-world settings, achieving 77% accuracy in a deployment with 38 participants over 900+ hours, distinguishing between in-person, virtual, and hybrid interactions.

Why it matters: This enables context-aware systems that could adapt interfaces, notifications, or collaborative tools based on users' real-time social engagement, directly supporting HCI research on adaptive user experiences and team dynamics sensing.

Method: Combined foreground speech detection with multimodal smartwatch sensing (likely audio, motion, physiological signals) trained on 100K+ labeled instances and validated through participant self-reports in naturalistic settings.

Jargon: Foreground speech detection = identifying when the wearer is actively speaking vs. hearing background conversation; balanced accuracy = average of sensitivity and specificity to handle uneven class distributions.

Finding: SemVideo reconstructs videos from fMRI brain activity by using hierarchical semantic guidance (static descriptions, motion narratives, and holistic summaries) to achieve better visual consistency and temporal coherence than previous approaches.

Why it matters: This advances understanding of how the brain processes visual information and demonstrates sophisticated human-AI collaboration in interpreting neural signals, with implications for brain-computer interfaces and cognitive load assessment in HCI applications.

Method: Novel framework combining fMRI signal processing with CLIP embeddings and a tripartite attention fusion architecture, tested on CC2017 and HCP datasets.

Jargon: fMRI = functional magnetic resonance imaging (brain activity measurement); CLIP = vision-language model that connects images and text; tripartite attention = three-way attention mechanism.

Finding: A haptic teleoperation system allows therapists to remotely guide patients through arm movements via a robotic exoskeleton, resulting in faster movement completion, smoother execution, and fewer verbal instructions compared to visual demonstration methods.

Why it matters: This demonstrates how intuitive physical interfaces can enhance human-robot collaboration and has implications for designing more natural interaction modalities in HCI and therapeutic applications.

Method: Controlled experiment with 32 participants in trainer-trainee pairs, comparing haptic vs. visual demonstration conditions with quantitative movement analysis and automated speech analysis using large language models.

Jargon: Haptic teleoperation - remotely controlling a robot while receiving tactile/force feedback; Exoskeleton - wearable robotic device that assists or guides human movement.

Finding: Researchers developed a system where backcountry skiing guides can annotate 3D maps of terrain with safety information (hazards, speed zones) that are then displayed to skiers through augmented reality heads-up displays, with a VR prototype for safe testing of the interaction design.

Why it matters: This demonstrates spatial annotation interfaces, AR visualization design, and safety-critical communication systems that directly inform HCI design for high-stakes collaborative environments.

Method: Uses photogrammetry mapping, tablet-based spatial annotation tools, AR overlays, and VR prototyping for participatory design studies.

Jargon: Photogrammetry - creating 3D maps from photographs; in situ - in the actual location/context of use.

Finding: VCC-Net uses eye-tracking and mouse interactions to capture radiologists' visual attention patterns during chest X-ray diagnosis, integrating this human cognitive data with AI models to improve diagnostic accuracy and create more interpretable, collaborative human-AI decision-making systems.

Why it matters: This demonstrates a sophisticated approach to human-AI collaboration that leverages natural human cognitive processes (visual attention) as input to improve AI systems, directly relevant to HCI design and understanding how humans and AI can work together effectively.

Method: The system captures radiologists' visual search traces through eye-tracking or mouse interfaces during real diagnostic tasks, then uses this data to guide AI attention mechanisms and build disease-aware graphs that align model representations with human cognitive patterns.

Jargon: Visual cognition (VC) refers to the cognitive processes involved in visual perception and attention; cognition-graph co-editing integrates human attention patterns with AI inference to create structured representations of diagnostic reasoning.

Finding: An AI-powered conversational system for finding food pantry information reveals fundamental challenges in information retrieval for fragmented, inconsistent public service data, including poor handling of vague queries and unreliable knowledge bases.

Why it matters: This demonstrates critical HCI design challenges when building AI systems for real-world social problems, showing how technical limitations create barriers to accessing essential services.

Method: Used Retrieval-Augmented Generation (RAG) with web scraping to build a conversational interface, then conducted pilot evaluation with community-sourced queries to test real-world performance.

Jargon: RAG (Retrieval-Augmented Generation) - AI approach that combines information retrieval with text generation to answer questions using relevant documents.

Finding: Persona4Rec uses LLMs to create offline "persona" representations of items based on user motivations inferred from reviews, enabling fast real-time recommendations while maintaining interpretability and performance comparable to expensive online LLM reasoning.

Why it matters: Demonstrates how AI can enhance human-computer interaction through interpretable recommendations while addressing the critical UX challenge of system latency in real-world deployment.

Method: Combines offline LLM reasoning over item reviews to generate persona representations with online lightweight encoding that matches user profiles to item personas.

Jargon: Rerankers - systems that reorder candidate items by relevance; Persona representations - interpretable profiles capturing different user motivations for engaging with items.

Finding: The paper develops a logic-based explainable AI system that makes the Framingham Cardiovascular Risk Score transparent by identifying minimal sets of patient attributes that explain risk classifications and suggesting actionable interventions to reduce risk.

Why it matters: This demonstrates practical approaches to making AI decision-making transparent and actionable for users, which is relevant to HCI design principles for AI systems and understanding how to present complex algorithmic outputs to support human decision-making.

Method: Uses first-order logic and XAI principles to analyze all possible input combinations of the Framingham Risk Score (22,000+ samples) to generate explanations and actionable recommendations.

Jargon: XAI = Explainable Artificial Intelligence; Framingham Risk Score = widely-used cardiovascular disease risk prediction tool.

Finding: SPG-LLM uses large language models to pre-filter irrelevant elements in automated planning tasks before the computationally expensive grounding step, achieving orders of magnitude speedup in planning benchmarks.

Why it matters: This demonstrates how AI can augment human-like reasoning processes (filtering irrelevant information before detailed analysis) and shows potential for AI systems that better collaborate with human planning and decision-making workflows.

Jargon: Grounding = converting abstract planning operators into specific actions; PDDL = Planning Domain Definition Language used in automated planning systems.

Finding: In timing games where players compete to act fastest after a randomly appearing opportunity (observed with delay), the unique symmetric equilibrium leads to costly "spam" actions as players attempt to time their responses optimally.

Why it matters: This provides insights into competitive dynamics and inefficient behaviors that emerge in time-sensitive decision-making contexts, relevant to understanding team coordination challenges and behavioral economics.

Jargon: Probabilistic backrunning - strategy of placing orders immediately after price-impacting trades; symmetric equilibrium - solution where all players use identical strategies.

Finding: This study compares machine learning models for predicting kidney discard rates, finding that consistent data preprocessing and evaluation methods matter more than algorithm choice, with ensemble methods achieving the best performance.

Why it matters: While primarily a medical application, it demonstrates important principles about ML model comparison, feature engineering, and evaluation strategies that are relevant when teaching about AI decision-making systems and their impact on critical human decisions.

Method: Systematic comparison of five ML models using unified benchmarking, standardized preprocessing, Bayesian hyperparameter optimization, and SHAP explainability analysis on 4,080 donor cases.

Jargon: SHAP = method for explaining individual predictions by computing feature importance; MCC = Matthews Correlation Coefficient, a balanced accuracy metric.

Finding: A neural network architecture called StrassenNet successfully rediscovered the optimal Strassen algorithm for 2×2 matrix multiplication and provided evidence that rank-23 may be optimal for 3×3 matrix multiplication through consistent convergence patterns across training runs.

Why it matters: This demonstrates how AI systems can autonomously discover mathematical algorithms, relevant to understanding AI's potential for creative problem-solving and algorithm discovery in computational tasks.

Method: Neural architecture designed to learn tensor decompositions through training on matrix multiplication tasks with different rank constraints.

Jargon: Tensor decomposition - breaking down multi-dimensional arrays into simpler components; rank - the minimum number of simple tensors needed to express a given tensor; border rank - a generalization allowing approximate decompositions.

Finding: This paper develops sum-of-squares optimization methods for computing optimal strategies in imperfect-recall extensive-form games, where players can forget past information, proving the methods converge and can solve certain game classes with a single semidefinite program.

Why it matters: While primarily theoretical computer science, it relates to decision-making under uncertainty and strategic interaction, which connects to behavioral science research on how people make sequential decisions with incomplete information.

Jargon: Extensive-form games = mathematical models of sequential decision-making with multiple players; imperfect recall = players forget previous information; sum-of-squares = polynomial optimization technique; semidefinite program = convex optimization method.

Finding: C²TC introduces a training-free framework for condensing tabular datasets that improves computational efficiency by 2+ orders of magnitude while maintaining performance, using class-adaptive clustering instead of gradient-based optimization.

Why it matters: This addresses a key challenge in data-driven decision-making systems by enabling more efficient processing of large-scale tabular data, which is fundamental to many HCI applications and behavioral analytics.

Method: Reformulates dataset condensation as a class-adaptive cluster allocation problem (CCAP) solved through heuristic local search, with hybrid categorical feature encoding for heterogeneous data.

Jargon: Dataset condensation (DC) = creating smaller synthetic datasets that preserve the utility of larger original datasets for training machine learning models.

Finding: The paper presents adaptations of active automata learning algorithms that leverage error-awareness and domain knowledge to accelerate learning of behavioral models by 1-2 orders of magnitude.

Why it matters: This relates to AI-assisted learning systems and could inform how educational technologies might adapt based on error patterns, though the connection to behavioral science applications is indirect.

Method: Theoretical extensions to the L# algorithm with empirical validation on challenging problems from literature.

Jargon: Active automata learning (AAL) - algorithms that learn behavioral models of systems through interaction; L# - a state-of-the-art AAL algorithm.

Finding: Researchers developed a machine learning method to create fast surrogate models that can replace slow agent-based epidemic models, reducing computation time by 10,000x while maintaining accuracy for hospital planning decisions.

Why it matters: This demonstrates how AI can augment human decision-making in high-stakes scenarios by making complex simulations accessible for real-time "what-if" analysis and threshold-aware planning.

Method: Uses Universal Differential Equations with neural networks to learn from large-scale agent-based model outputs, employing multiple shooting and prediction-error methods for stable training.

Jargon: ABM = agent-based model (individual-level simulation), UDE = Universal Differential Equations (physics-informed neural networks), SEIR = epidemiological model framework (Susceptible-Exposed-Infected-Recovered).