📚 Research Digest

Finding: Cross-domain mapping (forcing analogies from random, remote sources) significantly boosts human creativity in product innovation tasks, while LLMs generate more original ideas overall but don't benefit from this intervention - though both humans and LLMs show increased creativity when inspiration sources are more semantically distant.

Why it matters: This provides actionable evidence for specific creativity techniques that work for humans in design contexts, while revealing fundamental differences in how humans versus AI systems approach creative problem-solving.

Method: Controlled experiment comparing human participants and LLMs generating novel product features under two conditions (cross-domain mapping vs. user-need prompts), measuring creativity and originality of outputs.

Jargon: Cross-domain mapping = drawing analogies between unrelated domains to spark creative ideas; semantic distance = how conceptually unrelated two things are.

Finding: This empirical study with 29 teams and 80 participants found that AI-only approaches performed at or below median levels on domain-specific data science tasks, while human-AI collaborative approaches achieved the strongest performance across 17 industry challenges.

Why it matters: Provides concrete evidence about the effectiveness of human-AI collaboration versus full automation, directly informing how to design collaborative systems and understand AI's impact on expert work.

Method: Conducted an open competition benchmark (AgentDS) comparing AI-only baselines against human-AI collaborative teams across six industries with systematic performance evaluation.

Jargon: LLMs = Large Language Models (AI systems trained on vast text data); AI agents = autonomous AI systems that can perform tasks and make decisions.

Finding: This paper proposes shifting evaluation of human-AI systems from model accuracy to "team readiness," introducing a four-part taxonomy (outcomes, reliance behavior, safety signals, learning over time) to assess whether human-AI teams can collaborate safely and effectively.

Why it matters: This framework directly addresses how to design and evaluate AI collaboration tools that support good human decision-making, which is central to HCI design and understanding AI's impact on work.

Method: The authors develop an evaluation framework based on interaction traces rather than model properties or self-reported measures, operationalizing assessment through behavioral indicators of calibration and error recovery.

Jargon: Miscalibrated reliance = using AI inappropriately (overuse when wrong, underuse when helpful); U-C-I lifecycle = Understand-Control-Improve framework for human-AI onboarding.

Finding: This paper introduces mathematical metrics to distinguish between AI systems that enhance human reasoning (cognitive amplification) versus those that create dependency by replacing human thinking (cognitive delegation), arguing that sustainable human-AI collaboration requires preserving human expertise over time.

Why it matters: This framework directly addresses a critical design challenge in HCI - how to create AI systems that genuinely support human decision-making and collaboration without undermining human skills and agency.

Method: Develops operational metrics including Cognitive Amplification Index, Dependency Ratio, Human Reliance Index, and Human Cognitive Drift Rate to quantify these phenomena.

Jargon: Cognitive amplification = AI enhances human reasoning while preserving skills; Cognitive delegation = progressive outsourcing of thinking to AI systems.

Finding: This paper proposes a causal taxonomy distinguishing Human-in-the-Loop (HITL) as constitutive involvement where humans are necessary for AI decision outputs, versus Human-on-the-Loop (HOTL) as corrective involvement where humans can intervene to prevent or modify outputs, with further distinctions around timing and cognitive integration modes.

Why it matters: This framework directly addresses core HCI questions about designing effective human-AI collaboration systems and provides critical conceptual clarity for researchers studying AI's impact on human work and decision-making.

Method: Theoretical framework development using causal analysis to clarify existing ambiguous terminology in human-AI interaction.

Jargon: Constitutive = human input is required component of the system; Corrective = human acts as external monitor/modifier; Complementary intelligence = human and AI maintain distinct roles; Hybrid intelligence = human and AI capabilities are integrated/blended.

Finding: Users can identify UI personalization opportunities independently when shown their interaction data, but they prefer system support through visual suggestions, and this data-driven reflection increases their perceived value of personalization and transparency in system recommendations.

Why it matters: This demonstrates how data visualization and reflective interfaces can enhance user agency and decision-making in HCI, providing insights for designing more effective personalization systems that balance automation with user control.

Method: Interviewed 12 participants using experimental vignettes as design probes to explore how interaction data influences personalization decision-making.

Jargon: Reflexive personalization - approach where users reflect on their own interaction data to make personalization decisions rather than relying purely on automated systems.

Finding: LLMs do not exhibit true personality traits when assessed using human personality frameworks like the Big Five, as they fail to meet the six defining characteristics of personality, challenging widespread anthropomorphic assumptions in current research.

Why it matters: This directly impacts how we design and evaluate human-AI collaboration systems, suggesting current personality-based approaches to AI interaction design may be fundamentally flawed and need LLM-specific frameworks.

Jargon: Big Five inventory - a widely used personality assessment measuring five major personality dimensions (openness, conscientiousness, extraversion, agreeableness, neuroticism).

Finding: A 5W3H-based structured prompting framework (PPS) significantly improves intent alignment between humans and AI systems, particularly for high-ambiguity tasks, reducing follow-up prompts by 66% and outperforming simple prompts on goal alignment metrics.

Why it matters: This directly addresses a core HCI challenge of improving human-AI interaction design and demonstrates how structured communication protocols can enhance collaborative effectiveness between humans and AI systems.

Method: Controlled three-condition experiment across 60 tasks, 3 domains, and 3 LLMs (N=540 outputs), using LLM judges and introducing a novel "goal_alignment" evaluation metric.

Jargon: 5W3H = Who, What, When, Where, Why, How, How much, How many - a structured questioning framework; PPS = Prompt Protocol Specification using this structure; intent transmission loss = gap between user needs and what they communicate to AI.

Finding: Sketch2Topo combines hand-drawn sketching with AI diffusion models to make topology optimization more accessible and customizable, allowing users to specify design constraints and aesthetic preferences through natural drawing inputs rather than complex parameter settings.

Why it matters: This demonstrates how sketching as a visual thinking tool can bridge the gap between human creativity and AI capabilities, making complex engineering processes more intuitive and reducing cognitive load in human-AI collaboration.

Method: The system augments diffusion-based topology optimization with image-to-image generation and supports mask-based editing for regional customization.

Jargon: Topology optimization (TO) - engineering method to find optimal material distribution in structures; Diffusion models - AI models that generate images by learning to reverse a noise process.

Finding: AI-mediated video communication tools (retouching, background replacement, avatars) significantly reduce interpersonal trust and confidence in judgments, but do not actually impair people's ability to detect lies or distinguish truth from deception.

Why it matters: This directly informs HCI design decisions about AI-mediated communication tools and reveals important behavioral science insights about the disconnect between perceived and actual judgment accuracy in human-AI collaboration contexts.

Method: Two preregistered online experiments with 2,000 participants watching videos of truthful/deceptive statements across three conditions with varying AI mediation levels.

Jargon: AI mediation - using AI tools to alter, enhance or generate aspects of video communication like appearance, backgrounds, or full avatars.

Finding: Ocular and cardiac signals can predict user performance with 86% accuracy in interactive systems, with high performers showing more targeted gaze patterns and stable physiological responses as task complexity increases.

Why it matters: This demonstrates how physiological monitoring can enable proactive intervention in interactive systems when users are struggling, directly informing adaptive UX design and personalized learning systems.

Method: Within-subject experiment using a game environment with varying complexity levels, measuring eye tracking and heart rate data to predict performance across sessions.

Jargon: Ocular-cardiac fusion model = combined eye tracking and heart rate data analysis; visual sampling = how users direct their attention visually across interface elements.

Finding: Machine learning models can predict audience engagement and vocal attractiveness in video learning by analyzing only the speaker's facial expressions, eye movements, voice prosody, and semantic content, achieving high accuracy (R² = 0.85-0.88) without needing audience data.

Why it matters: This enables scalable, privacy-preserving assessment of teaching effectiveness and learning experience design in online education, directly informing HCI design for educational technology and understanding factors that drive engagement.

Method: Developed dual regression models using MOOC data corpus, analyzing multimodal speaker features (facial dynamics, oculomotor patterns, prosody, cognitive semantics) to predict audience-reported engagement and attractiveness ratings.

Jargon: Oculomotor features = eye movement patterns; Prosody = vocal rhythm, stress, and intonation patterns in speech.

Finding: MANAR implements Global Workspace Theory principles in neural attention mechanisms, creating a "mental workspace" that enables creative synthesis of concepts beyond simple combinations of inputs, while achieving linear-time scaling compared to standard quadratic attention.

Why it matters: This bridges cognitive science theories of consciousness with AI architecture design, offering insights into how artificial systems might achieve more human-like creative and integrative thinking processes.

Method: The architecture uses a two-stage process mimicking GWT: integration phase (forming collective "mental image") and broadcasting phase (using global state to contextualize local information).

Jargon: Global Workspace Theory (GWT) - cognitive theory explaining consciousness through a central workspace that integrates and broadcasts information; Abstract Conceptual Representation (ACR) - the system's global mental state formed from retrieved memory concepts.

Finding: A systematic study of 132 swipe delay patterns in short video streaming found that user experience depends not only on delay duration (8+ seconds causes dissatisfaction) but also on the number and timing of delays, with early-session delays being less harmful than late-session delays.

Why it matters: This provides concrete empirical data on how interaction responsiveness affects user experience, directly informing UX design principles for mobile interfaces and gesture-based interactions.

Method: Conducted subjective quality assessment with users experiencing different swipe delay patterns and developed a predictive QoE model based on delay characteristics.

Jargon: QoE (Quality of Experience) - subjective measure of user satisfaction with a system's performance and usability.

Finding: OS-Themis is a multi-agent critic framework that improves GUI agent training by decomposing user interactions into verifiable milestones and using multiple reviewers to evaluate performance, achieving 10.3% improvement in reinforcement learning training for GUI agents.

Why it matters: This advances understanding of how AI systems can better collaborate with humans through improved GUI interfaces and provides insights into designing feedback mechanisms for human-AI collaboration systems.

Method: Uses a multi-agent review system with milestone decomposition and evidence chain auditing, evaluated on AndroidWorld and a new cross-platform GUI benchmark (OGRBench).

Jargon: GUI agents = AI systems that interact with graphical user interfaces; critic framework = system for evaluating and providing feedback on AI performance; trajectory = sequence of actions taken by an AI agent.

Finding: M is an open-source, modular social robot platform designed to reduce research barriers through easy reproduction, modification, and real-world deployment, with demonstrated success in participatory design sessions and week-long home studies.

Why it matters: This directly addresses human-robot interaction research needs and provides a practical tool for studying human-AI collaboration, team dynamics with robotic agents, and behavioral interventions in naturalistic settings.

Method: Mixed methods approach combining platform development with participatory design sessions and longitudinal in-home deployment studies to validate real-world usability.

Jargon: ROS2 = Robot Operating System 2, a software framework for robot development; sim-to-real transfer = moving behaviors developed in simulation to physical robots.

Finding: AI-generated support in online health communities often conforms to community norms but sometimes inappropriately so (over- or under-validating users) or violates norms entirely, potentially eroding trust and threatening community sustainability.

Why it matters: This demonstrates how AI integration can disrupt established social norms and collaborative dynamics in online communities, with implications for designing AI systems that support rather than undermine human collaboration and collective intelligence.

Method: Used human-validated LLM judges to assess norm conformity in opioid recovery subreddits, followed by expert review to identify risks.

Jargon: OHCs = Online Health Communities where people exchange social support around health conditions.

Finding: LLMs generate better four-term word analogies than humans by more consistently following geometric "parallelogram" structures in semantic space, while humans produce more variable responses with a long tail of poor completions.

Why it matters: This reveals fundamental differences in how humans versus AI systems approach creative reasoning tasks, with implications for designing AI tools that complement rather than replace human analogical thinking.

Method: Compared human and LLM analogy completions on standardized problems, measuring alignment with parallelogram structure in GloVe embedding space and collecting human quality ratings.

Jargon: Parallelogram model - geometric representation where analogies A:B::C:D form parallelograms in semantic embedding space; GloVe - a word embedding model that captures semantic relationships.

Finding: This synthetic review examines how judges interact with AI-based decision aids in criminal justice, finding that AI tools have modest impact on judicial decisions while revealing significant gaps in understanding human-AI collaboration in high-stakes decision-making contexts.

Why it matters: This directly informs research on human-AI collaboration, decision-making under uncertainty, and how cognitive biases influence the adoption and use of algorithmic recommendations in professional settings.

Method: Synthetic review across computer science, economics, law, criminology, and psychology examining AI tool performance, human judgment biases, and human-AI interaction patterns.

Jargon: Risk assessment instruments = AI tools that predict likelihood of criminal recidivism or failure to appear in court; Synthetic review = comprehensive literature review that synthesizes findings across multiple disciplines.

Finding: Researchers developed and validated SVLAT, a 49-item test that measures the general public's ability to read, understand, and interpret scientific visualizations across 8 visualization techniques and 11 tasks. The instrument showed high reliability (McDonald's omega = 0.82) and was validated through expert review and testing with 485 participants.

Why it matters: This provides a validated tool for assessing visual literacy skills that are fundamental to HCI design and data visualization education, offering educators a way to measure and improve users' ability to interpret complex visual information.

Method: Used a rigorous psychometric development pipeline including expert content validation, pilot testing, and both classical test theory (CTT) and item response theory (IRT) for validation.

Jargon: Content validity ratio (CVR) - statistical measure of expert agreement on item relevance; McDonald's omega - reliability measure that doesn't assume equal factor loadings like Cronbach's alpha does.

Finding: A distributed Turing Test platform called TuringHotel where 17 humans and 19 LLMs simultaneously act as both judges and respondents in group discussions, revealing that current AI models can still sometimes be confused for humans but retain identifiable artificial "fingerprints."

Why it matters: This advances understanding of human-AI collaboration dynamics and provides a novel framework for studying mixed human-AI team interactions that could inform HCI design principles.

Method: Mixed-community experiment using the UNaIVERSE platform with authenticated peer-to-peer networking, where participants engage in time-bounded group discussions while trying to identify humans versus AI agents.

Jargon: LLMs = Large Language Models (AI systems trained on text data to generate human-like responses)

Finding: A four-stage training pipeline for bilingual dialogue systems achieves zero hallucination in citation-grounded conversations by progressively teaching models to ground responses in verifiable sources, with systematic explainability analysis revealing how citation behavior develops across training stages.

Why it matters: Demonstrates how to design AI systems that collaborate more transparently with humans by eliminating fabricated information and providing verifiable sources, directly relevant to human-AI collaboration and trust in work contexts.

Method: Uses progressive training with supervised fine-tuning followed by reinforcement learning, evaluated through three explainability techniques (cross-attention alignment, Integrated Gradients, occlusion-based analysis) across encoder-decoder and decoder-only architectures.

Jargon: GRPO = Group Relative Policy Optimization (reinforcement learning method); SFT = Supervised Fine-Tuning; hallucination = AI generating false or unverifiable information.

Finding: A case study of 5 employees found that Duolingo's LLM-generated lessons effectively teach general language skills but lack profession-specific content needed for workplace fluency, leading to a proposal for personalized, domain-specific lesson generation.

Why it matters: This demonstrates how AI-generated educational content can be evaluated from the learner's perspective and highlights the importance of personalization in learning systems design.

Method: Small-scale survey study with employees from a multinational company analyzing their experiences with general versus work-specific language learning scenarios.

Finding: A survey of 230 computer science students reveals gender differences in perceptions of AI ethics and societal impact, with women showing stronger tendencies toward ethical considerations while men focus more on technical applications and military threats.

Why it matters: Understanding student perspectives on AI ethics informs how to teach responsible AI development and reveals potential gender-based approaches to addressing AI's societal impact.

Method: Survey methodology examining gender-based differences in student attitudes toward AI applications and ethical concerns.

Jargon: Information war - coordinated disinformation campaigns using AI technologies.

Finding: Participatory design workshops with LGBTQ+ and cisgender participants identified barriers to inclusive communication in virtual environments and co-created speculative Metaverse solutions across activity, interaction, scene, and space layers.

Why it matters: This demonstrates how participatory design methods can surface power dynamics and bias in digital collaboration spaces, offering concrete principles for designing more equitable team environments and virtual workspaces.

Method: Three-phase futuring methodology combining challenge identification, speculative problem-solving, and future visualization through participatory workshops.

Jargon: Power geometry - how power relations are structured and experienced in spatial contexts; Futuring - design research method using speculative scenarios to explore alternative possibilities.

Finding: SCISSR introduces a scribble-based interface for interactive surgical image segmentation that allows users to iteratively refine segmentations by drawing corrective strokes, achieving 95%+ accuracy on surgical datasets while outperforming point-based prompting methods.

Why it matters: This demonstrates how natural sketching interactions can significantly improve human-AI collaboration in complex visual tasks, providing insights for designing more intuitive AI interfaces that leverage human spatial reasoning.

Method: Uses a lightweight Scribble Encoder to convert freehand scribbles into dense prompt embeddings for SAM models, with iterative refinement through corrective strokes and evaluation on surgical segmentation benchmarks.

Jargon: SAM (Segment Anything Model) - AI models for image segmentation; Dice score - metric measuring segmentation accuracy; LoRA adapters - efficient neural network fine-tuning components.

Finding: The paper introduces "Align-to-Scale," a technique that uses spatial alignment between gaze and hand as a mode switch to enable single-handed object scaling in extended reality, addressing limitations of current XR interaction methods that typically require two hands for scaling operations.

Why it matters: This advances HCI research by developing more accessible and flexible interaction techniques for XR interfaces, with implications for designing inclusive technologies and understanding multimodal interaction design principles.

Method: Conducted usability evaluation comparing several one-handed scaling techniques in compound translate-scale tasks to assess performance trade-offs and derive design guidelines.

Jargon: Extended Reality (XR) - umbrella term for VR, AR, and mixed reality technologies; Gaze + Pinch - interaction model combining eye tracking with finger gestures; unimanual interaction - single-handed operation.

Finding: FinTradeBench introduces a benchmark for evaluating how well LLMs can reason about financial decisions by integrating company fundamentals with trading signals, revealing that current LLMs struggle particularly with numerical and time-series reasoning despite retrieval augmentation helping with textual analysis.

Why it matters: This demonstrates key limitations in AI reasoning capabilities that affect human-AI collaboration in complex decision-making domains, informing how we design AI tools that support rather than replace human judgment.

Method: Uses a calibration-then-scaling framework combining expert seed questions, multi-model response generation, self-filtering, and human-LLM judge alignment to create reliable benchmarks at scale.

Jargon: Trading signals - quantitative indicators derived from stock price movements and market data used to inform investment decisions.

Finding: This paper develops online learning algorithms that work with ranking feedback instead of numeric utility scores, showing that traditional regret bounds are impossible with instantaneous rankings but achievable with time-average rankings under certain conditions.

Why it matters: This addresses a key challenge in human-AI collaboration where users can rank options but not provide precise numeric feedback, which is common in design evaluation, preference learning, and human-in-the-loop systems.

Method: Theoretical analysis using regret bounds and game theory, with empirical validation on large language model routing tasks.

Jargon: External regret = measure of algorithm performance vs. best fixed strategy; Plackett-Luce model = probabilistic ranking model; coarse correlated equilibrium = game theory solution concept where players coordinate using shared randomization.

Finding: CAMO is a neural network that solves multi-objective optimization problems for robot teams, learning to balance competing goals like minimizing travel time versus distributing work evenly across agents.

Why it matters: This demonstrates how AI can support team coordination and goal trade-off decisions, relevant for understanding AI's role in collaborative work and multi-objective decision-making.

Method: Uses reinforcement learning (REINFORCE) to train a neural network with conditional encoding for preferences and collaborative decoding for agent coordination, tested on both simulations and real robots.

Jargon: Pareto front - the set of optimal solutions where improving one objective requires worsening another; MOMTSP - optimization problem involving multiple salesmen visiting targets while balancing multiple competing goals.

Finding: Em-Garde introduces a two-stage framework for proactive video understanding that separates semantic understanding from streaming perception, achieving better accuracy-efficiency tradeoffs than per-frame decision making approaches.

Why it matters: This addresses the fundamental HCI challenge of creating responsive AI systems that can proactively engage with users while managing computational constraints, relevant to designing efficient human-AI collaboration interfaces.

Method: The framework uses an Instruction-Guided Proposal Parser to convert queries into visual proposals at query time, then employs lightweight embedding-based matching during video streaming to trigger responses.

Jargon: VideoLLMs = Video Large Language Models that process video and respond to natural language queries; proactive streaming = AI systems that anticipate and respond to user needs in real-time video scenarios.

Finding: ChatGPT exhibits brittle geographic reasoning with strong defaults that shift dramatically from minor prompt variations, and compositional tasks can produce unexpected distributional biases even when individual components seem benign.

Why it matters: This reveals critical limitations in AI reasoning that affect human-AI collaboration quality and highlights cognitive biases in AI systems that parallel human decision-making biases.

Jargon: Distributional shifts - changes in data patterns that cause AI performance to degrade; Foundation models - large pre-trained AI models like GPT that serve as bases for many applications.

Finding: An LLM agent successfully generated 665 novel research problems in differential geometry, with human expert verification confirming that many problems were previously unknown and possessed genuine research value.

Why it matters: Demonstrates AI's potential to augment human creativity in complex domains and provides insights into how AI might transform collaborative research and problem-generation processes.

Method: Developed an AI agent specifically for mathematical problem generation and used human expert evaluation to verify novelty and research value of generated problems.

Jargon: LLM = Large Language Model; differential geometry = branch of mathematics using calculus to study geometric properties.

Finding: MemMA introduces a multi-agent framework that coordinates memory construction, retrieval, and utilization in LLM systems through strategic reasoning (forward path) and self-correcting memory repair (backward path), showing consistent improvements across different LLM backbones.

Why it matters: This demonstrates how multi-agent coordination can improve AI system performance and provides insights into designing AI systems that can better support human collaboration through more strategic memory management.

Method: Experimental evaluation on LoCoMo benchmark comparing MemMA against existing baselines across multiple LLM backbones and storage systems, with plug-and-play framework design.

Jargon: Memory cycle - the process of constructing, retrieving, and utilizing information from external memory banks; In-situ self-evolution - real-time self-correction and repair of memory before finalization.

Finding: Proposes a philosophical framework (Onto-Relational-Sophic) for governing AI systems by treating them as multi-dimensional entities with graded personhood rather than simple tools, combining ontological, relational, and wisdom-based approaches to AI governance.

Why it matters: Addresses fundamental questions about human-AI collaboration and the evolving relationship between humans and increasingly capable AI systems, which is central to understanding AI's impact on work and society.

Jargon: Cyberism - philosophical approach treating synthetic minds as multi-dimensional beings; CPST ontology - framework defining AI as existing across cyber-physical-social-thinking dimensions; Cybersophy - wisdom-oriented ethical approach combining multiple moral frameworks.

Finding: This study identifies four implicit token-level patterns (early pruning, path-dependent lock-in, targeted backtracking, and knowledge-guided prioritization) that emerge when LLMs perform multi-pass binary vulnerability analysis, revealing how AI systems organize exploration through implicit rather than explicit decision-making.

Why it matters: This provides insights into AI decision-making patterns and cognitive processes that could inform research on human-AI collaboration and understanding how AI systems develop implicit reasoning strategies.

Method: Large-scale trace-level analysis of 521 binaries with 99,563 reasoning steps to identify emergent patterns in LLM behavior.

Jargon: Binary analysis - examining compiled software code for security vulnerabilities; Multi-pass reasoning - iterative problem-solving approach where the model revisits and refines its analysis multiple times.

Finding: DyMoE is a dynamic quantization framework that reduces inference time for MoE (Mixture of Experts) models on edge devices by 3.44x-22.7x through importance-aware expert prioritization and depth-adaptive scheduling.

Why it matters: This technical optimization work has limited direct relevance to HCI, creativity, teams, or behavioral science, though it could tangentially support AI accessibility research.

Method: Experimental evaluation on commercial edge hardware comparing dynamic mixed-precision quantization against static offloading baselines.

Jargon: MoE (Mixture of Experts) - AI models using multiple specialized sub-networks; TTFT/TPOT - timing metrics for AI response generation; quantization - reducing model precision to save memory.

Finding: The authors develop Dual-IFM, an interpretable-by-design foundation model for medical imaging that provides both local interpretability (individual image explanations) and global interpretability (dataset visualization) while maintaining performance comparable to larger state-of-the-art models.

Why it matters: This demonstrates how interpretability can be built into AI systems from the ground up rather than added post-hoc, which is relevant for understanding human-AI collaboration and trust in high-stakes domains.

Jargon: Foundation models - large pre-trained models that learn general representations; Self-supervised learning (SSL) - learning without labeled data by predicting parts of the input; Fundus photography - imaging of the back of the eye for medical diagnosis.

Finding: The paper proposes HISR, a method that uses hindsight information to modulate segment-level rewards in reinforcement learning for multi-turn agentic tasks, improving credit assignment by focusing on significant segments rather than turn-level granularity.

Why it matters: This touches on AI decision-making processes and could inform understanding of how AI agents learn from complex tasks, but has limited direct application to HCI, creativity, teams, or behavioral science education.

Method: Uses segment-level process reward models combined with hindsight models that measure action importance through likelihood ratios, tested on three public benchmarks.

Jargon: Credit assignment = determining which actions contributed to outcomes; Agentic tasks = tasks requiring autonomous decision-making; Hindsight model = evaluates actions knowing the final outcome.

Finding: This paper proposes REST, a framework for robot navigation that uses tree-structured path planning and LLM reasoning to help robots find objects in unknown environments without prior training.

Why it matters: While primarily a robotics paper, it demonstrates how LLMs can be integrated into decision-making systems and spatial reasoning tasks, which has implications for AI-human collaboration in navigation and exploration contexts.

Method: Uses sampling-based planning to create tree structures of navigation paths, then employs chain-of-thought LLM reasoning to select optimal routes based on spatial narratives.

Jargon: Zero-shot object-goal navigation (ZSON) - robot navigation to find objects without task-specific training; Steiner Tree - a tree connecting points with minimal total edge weight.

Finding: This paper develops algorithms to compress datasets for linear optimization problems while preserving the ability to make optimal decisions, showing that finding globally optimal compressions is computationally hard but proposing tractable alternatives with theoretical guarantees.

Why it matters: The work touches on decision-making under uncertainty and data compression for AI systems, but is primarily focused on mathematical optimization theory rather than human behavior, collaboration, or interface design.

Method: Uses geometric analysis to characterize decision-sufficient datasets and develops polynomial-time cutting-plane algorithms with PAC learning guarantees.

Jargon: SDD (Sufficient Decision Dataset) - a compressed dataset that preserves optimal decision-making; PAC guarantee - probably approximately correct learning bound; coNP-hard - computational complexity class for verification problems.