Week 45 / November 2025

Constraint Architecture Solves the Flexibility-Reliability Tradeoff

From physical signal obfuscation to bounded UI generation, systems preserve adaptability by enforcing boundaries

Synthesized using AI

Analyzed 136 papers. AI models can occasionally hallucinate, please verify critical details.

The strongest work this week solves a problem most systems ignore: how to preserve flexibility while enforcing boundaries. PrivyWave demonstrates this at the physical layer—cryptographic obfuscation that generates decoy heartbeat signals only authorized sensors can filter. Unauthorized mmWave radar shows 21.3 BPM mean absolute error while authorized devices maintain 5.8 BPM. The same principle appears in UI generation: Portal UX Agent lets LLMs plan structure but constrains execution to schema-validated components, avoiding the security and design-system violations that kill free-form code generation. GUI-AIMA extends this by treating grounding as visual patch selection rather than coordinate generation, achieving state-of-the-art accuracy with 85k training examples instead of millions. The pattern holds in accessibility—FocusView reveals that ADHD viewers need customization infrastructure but too many options become the distraction themselves, suggesting curated presets over granular controls.

The misinformation work complicates this by showing where constraint fails. The Pakistan study with low-SES adults found that Western media literacy frameworks don't transfer because they miss local trust structures and communal decision-making patterns. The Scaffolded Support Model works by surfacing these structures through participatory design, not by imposing external frameworks. Separately, research on LLMs as information sources reveals they produce distinctive failure modes—misprioritization and exaggeration—that traditional fact-checking misses. You can't constrain what you can't detect, and detection strategies designed for human-generated content systematically fail on LLM outputs.

The implication: constraint architecture works when you can define the boundary in advance—component libraries for UI, cryptographic keys for sensing, curated profiles for accessibility. It fails when the problem space is culturally situated or when new failure modes emerge faster than you can encode rules. The design challenge is knowing which category you're in before you commit to the architecture.

Featured(1/6)

2511.02993

PrivyWave: Privacy-Aware Wireless Sensing of Heartbeat

Yixuan Gao, Tanvir Ahmed, Zekun Chang, Thijs Roumen, Rajalakshmi Nandakumar

Preprint·2025-11-04

Deploy this for health monitoring in shared spaces where you need selective access—authorized caregivers get accurate readings while blocking surveillance. Works across 30-150 cm distances and 120° field of view.

Wireless sensors can now detect heartbeats through walls using RF and acoustic signals. Existing privacy solutions either block all sensing (killing utility) or operate post-capture (too late for selective access control).

Method: A key-based physical obfuscation system that generates cryptographically-determined decoy heartbeat signals. Unauthorized mmWave radar sensors show 21.3 BPM mean absolute error while authorized sensors maintain 5.8 BPM. For acoustic sensing, unauthorized error jumps to 42.0 BPM while authorized sensors achieve 9.7 BPM. The system works across modalities without per-modality customization by mixing real and decoy signals that only authorized devices can filter.

Caveats: Tested with 13 participants in indoor environments. Real-world deployment at scale unverified.

Reflections: Can adversaries with multiple synchronized sensors triangulate to defeat the decoy system? · What's the energy cost of continuous decoy generation for battery-powered wearables? · How does the system perform with multiple people in the same space?

privacy-securitywearableshealthcare

2511.06147

Towards Misinformation Resilience in Pakistan: A Participatory Study with Low-Socioeconomic Status Adults

Muhammad Abdullah Sohail, Amna Hassan, Shaheer Hammad, Salaar Masood, Suleman Shahid

Preprint·2025-11-08

Stop exporting Western media literacy frameworks. Build misinformation tools using participatory methods that surface local trust structures and communal decision-making patterns. The Scaffolded Support Model works where lecture-based interventions fail.

Misinformation interventions designed for Western contexts fail with low-SES populations in the Global South. Existing tools report limited success, particularly among marginalized communities with different information ecologies.

Method: Three-phase participatory study with 41 low-SES Pakistani adults revealed a sophisticated moral economy of sharing and layered trust ecology prioritizing communal welfare. This informed the Scaffolded Support Model—combining on-demand assistance with gradual inoculation-based skill building. Prototype 'Pehchaan' tested with 15 users showed strong acceptance and cultural resonance, validating the culturally grounded approach.

Caveats: Tested only in Pakistan with low-SES adults. Transfer to other Global South contexts requires local participatory research.

Reflections: How does the moral economy of sharing change as users gain digital literacy? · Can the Scaffolded Support Model scale beyond prototype deployment? · What happens when communal trust structures themselves become vectors for coordinated misinformation?

social-computingtrust-safetyeducation

2511.00810

GUI-AIMA: Aligning Intrinsic Multimodal Attention with a Context Anchor for GUI Grounding

Shijie Zhou, Viet Dac Lai, Hao Tan, Jihyung Kil, Wanrong Zhu, Changyou Chen, Ruiyi Zhang

Preprint·2025-11-02

Stop training massive models on millions of screenshots. Light fine-tuning on 85k examples triggers native grounding capability. Use attention alignment instead of coordinate generation for GUI agents—it's faster and more accurate.

Computer-use agents need to map natural language instructions to screen regions. Existing MLLMs generate coordinates directly from visual inputs—computationally expensive and imprecise.

Method: Attention-based, coordinate-free framework that aligns MLLMs' intrinsic multimodal attention with patch-wise grounding signals calculated adaptively via multi-head aggregation. GUI-AIMA-3B trained on only 85k screenshots achieves 59.6% accuracy on ScreenSpot-Pro, 63.8% on OSWorld-G, and 91.5% on ScreenSpot-v2—state-of-the-art for 3B models. The coordinate-free approach enables plug-and-play zoom-in stages.

Caveats: Tested only on standard GUI grounding benchmarks. Performance on dynamic or non-standard interfaces unknown.

Reflections: Does the attention alignment approach transfer to mobile or AR interfaces? · How does performance degrade with UI elements not seen during training? · Can the same attention mechanism ground multi-step workflows across screens?

ai-interactiondesign-toolsevaluation-methods

1 / 6

Featured

Findings(1/5)

UI generation shifts from code synthesis to bounded composition·Privacy protection moves upstream from data collection to physical signal layer·Accessibility design inverts from universal solutions to user-controlled customization·Misinformation interventions fail when designed without participatory grounding in target communities·AI-assisted learning surfaces metacognitive bypass as the critical risk, not correctness

LLM-driven interfaces are moving away from free-form code generation toward constrained assembly systems. Portal UX Agent demonstrates this shift: the LLM plans at a high level while a deterministic renderer assembles components, preserving reliability and design-system compliance without sacrificing adaptability. GUI-AIMA extends this by treating grounding as visual patch selection rather than coordinate generation, reducing computational cost. The implication: expressiveness and governance aren't opposed—they require architectural separation of planning from execution.

2511.00843

Portal UX Agent -- A Plug-and-Play Engine for Rendering UIs from Natural Language Specifications

2511.00810

GUI-AIMA: Aligning Intrinsic Multimodal Attention with a Context Anchor for GUI Grounding

Surprises(1/3)

LLMs produce distinctive low-quality information, not just more of the same·VLMs infer sensitive attributes from everyday personal videos at levels that create new threat models·Religious stakeholders identify secular embeddings as non-neutral design choices, not universal defaults

The assumption was that LLMs would replicate existing patterns of misinformation at scale. They don't. Research on LLM outputs as information sources reveals that the types of low-quality information produced by LLMs differ qualitatively from traditional sources—not just in volume but in kind. This matters because detection and intervention strategies designed for human-generated misinformation may miss LLM-specific failure modes entirely.

2511.03198

Large Language Models as Information Sources: Distinctive Characteristics and Types of Low-Quality Information

TOOLBOX(8)

GUI-AIMA

Model

Attention-based, coordinate-free GUI grounding framework for MLLMs. Aligns intrinsic multimodal attention with patch-wise grounding signals via multi-head aggregation. GUI-AIMA-3B trained on 85k screenshots achieves 59.6% accuracy on ScreenSpot-Pro, 63.8% on OSWorld-G, 91.5% on ScreenSpot-v2. Includes plug-and-play zoom-in stage for efficient visual grounding.

2511.00810

FocusView

Tool

Video customization interface enabling viewers with ADHD to customize informational videos across multiple aspects. Evaluated with 12 ADHD participants, significantly improved video viewability by reducing distractions. Addresses diverse perceptions of video elements (e.g., background music) and provides ADHD-specific customization options while minimizing distraction from customization itself.

2511.01248

PrivyWave

Tool

Key-based physical obfuscation system for privacy-aware wireless heartbeat sensing. Generates controlled decoy heartbeat signals at cryptographically-determined frequencies. Evaluated with 13 participants: unauthorized mmWave radar sensors show 21.3 BPM mean absolute error vs 5.8 BPM for authorized; acoustic sensing shows 42.0 BPM vs 9.7 BPM. Operates across 30-150cm distances with 120° field of view.

2511.02993

SnappyMeal

Tool

AI-powered multimodal dietary tracking system leveraging multiple input methods (images, text, voice). Features goal-dependent follow-up questions and information retrieval from grocery receipts and nutritional databases. Evaluated through 3-week in-the-wild deployment with multiple users, capturing over 500 logged food instances with strong perceived accuracy and user satisfaction.

2511.03907

Vitessce Link

Framework

Web-based hybrid framework uniting 3D stereoscopic mixed reality view with synchronized 2D display for 3D tissue map analysis. Enables volumetric navigation via hand gestures while controlling channels, filters, and derived data through Vitessce platform. Built on open web standards, runs entirely in browser, uses lightweight WebSocket architecture for cross-device synchronization. Supports computational notebook integration.

2511.04262

psiUnity

Framework

Open-source C# integration bridging Platform for Situated Intelligence (psi) with Unity 2022.3 and MRTK3 for HoloLens 2. Enables bidirectional real-time streaming of head pose, hand tracking, gaze, IMU, audio, and depth sensor data (AHAT and long-throw) with microsecond-level temporal precision. Embeds psi's native serialization, logging, and temporal coordination within Unity architecture.

2511.05304

AIDEN

Tool

AI assistant for visually impaired individuals with hybrid architecture integrating YOLO for real-time object detection and LLaVA for scene description and OCR. Features continuous haptic guidance using Geiger-counter metaphor for object centering without occupying auditory channel. Evaluated with visually impaired participants using Technology Acceptance Model, showing high satisfaction for intuitiveness and perceived autonomy.

2511.06080

Pehchaan

Tool

Misinformation resilience prototype implementing the Scaffolded Support Model for low-SES Pakistani adults. Integrates on-demand assistance with gradual inoculation-based skill acquisition. Developed through three-phase participatory study with 41 low-SES adults, validated via usability testing (N=15) showing strong acceptance and cultural resonance. Addresses moral economy of sharing and layered ecology of trust in information practices.

2511.06147

Revealing AI Reasoning Increases Trust but Crowds Out Unique Human Knowledge

Shows that explaining AI logic makes humans *more* trusting but *less* likely to contribute their own expertise—transparency as a collaboration killer.

2511.04964

Scientific judgment drifts over time in AI ideation

Discovers that scientists' ratings of AI-generated research ideas aren't stable—they shift over days, undermining the assumption of a fixed evaluation gold standard.

2511.01205

When Machines Join the Moral Circle: The Persona Effect of Generative AI Agents in Collaborative Reasoning

Tests whether giving AI a persona changes how students deliberate trolley problems in groups. Spoiler: it does, and the implications are unsettling.

2511.01233

Towards Reliable Human Evaluations in Gesture Generation: Insights from a Community-Driven State-of-the-Art Benchmark

Audits gesture generation research and finds evaluation chaos—no one knows what's state-of-the-art because human studies are incomparable. Proposes a fix.

2511.02378

Revisiting put-that-there, context aware window interactions via LLMs

Resurrects Bolt's 1980 "Put-That-There" demo for AR headsets using LLMs to fuse pointing, gaze, speech, and live app metadata into spatial commands.

2511.00945

"Less is More": Reducing Cognitive Load and Task Drift in Real-Time Multimodal Assistive Agents for the Visually Impaired

Identifies why vision-language assistants overwhelm blind users—too much info, too much drift—then redesigns the interaction model around selective, task-focused responses.

2511.01663

The Ghost in the Keys: A Disklavier Demo for Human-AI Musical Co-Creativity

Builds a self-playing piano that listens to you perform and responds in real time, replacing text prompts with embodied, improvisational musical dialogue.

2511.03117

Tracing Generative AI in Digital Art: A Longitudinal Study of Chinese Painters' Attitudes, Practices, and Identity Negotiation

Tracks 17 digital painters over five years as generative AI arrives—from resistance to adoption to existential reckoning about what it means to be an artist.

REFLECTION(4)

Transparency breeds confidence, not judgment

Across 22 papers on AI interaction, education, and healthcare, showing users how systems reason doesn't calibrate their trust—it inflates it. The paradox is sharp: explainability, designed to empower, instead creates a credibility illusion that users mistake for understanding.

Transparent AI reasoning increases user confidence without improving decision accuracy. If explainability is backfiring as a trust mechanism, are we solving the wrong problem—should we be designing for *appropriate skepticism* rather than informed consent?

1 / 4

Week 44October 2025

Week 01January 2026

ABOUT THIS ISSUE

How was this newsletter synthesized?

Methodology

This newsletter is generated by an AI pipeline (leveraging Anthropic Sonnet 4.5 & Haiku 4.5) that processes the metadata and abstracts of every new arXiv HCI paper from the past week—136 this issue. Each paper is scored on three dimensions: Practice (applicability for practitioners), Research (scientific contribution), and Strategy (industry implications), with scores from 1-5. Papers passing threshold are grouped into topic clusters, and each cluster is summarized to capture what that body of research is exploring.

Selection Criteria

The pipeline builds a curated selection that balances high scores with topic diversity—and deliberately includes at least one 'contrarian' paper that challenges prevailing assumptions. This selection is then analyzed to identify key findings (patterns across multiple papers) and surprises (results that contradict conventional wisdom). A narrative synthesis ties the week's research together under a unifying frame.

Key Themes Discovered

Field Report: ai-interaction

Trust, Calibration, and Collaboration

This cluster examines how humans evaluate, trust, and collaborate with AI systems across high-stakes domains. Core questions: How do users calibrate trust when AI reasoning is revealed? Do AI credibility signals override institutional judgment? How do expert standards drift when evaluating AI outputs? Research spans trust dynamics in human-AI teams, information quality assessment, and the gap between perceived and desired AI behavior. Methodologically diverse—combining experiments, longitudinal studies, and qualitative analysis—the work reveals systematic biases in human judgment and the paradox that transparency often induces over-trust rather than calibration.

1/12

SnappyMeal: Design and Longitudinal Evaluation of a Multimodal AI Food Logging Application

2511.04478

Generate, Evaluate, Iterate: Synthetic Data for Human-in-the-Loop Refinement of LLM Judges

2511.01205

Synthesized using AI

PrivyWave: Privacy-Aware Wireless Sensing of Heartbeat

Towards Misinformation Resilience in Pakistan: A Participatory Study with Low-Socioeconomic Status Adults

GUI-AIMA: Aligning Intrinsic Multimodal Attention with a Context Anchor for GUI Grounding

Portal UX Agent -- A Plug-and-Play Engine for Rendering UIs from Natural Language Specifications

GUI-AIMA: Aligning Intrinsic Multimodal Attention with a Context Anchor for GUI Grounding

Large Language Models as Information Sources: Distinctive Characteristics and Types of Low-Quality Information

GUI-AIMA

FocusView

PrivyWave

SnappyMeal

Vitessce Link

psiUnity

AIDEN

Pehchaan

Revealing AI Reasoning Increases Trust but Crowds Out Unique Human Knowledge

Scientific judgment drifts over time in AI ideation

When Machines Join the Moral Circle: The Persona Effect of Generative AI Agents in Collaborative Reasoning

Towards Reliable Human Evaluations in Gesture Generation: Insights from a Community-Driven State-of-the-Art Benchmark

Revisiting put-that-there, context aware window interactions via LLMs

"Less is More": Reducing Cognitive Load and Task Drift in Real-Time Multimodal Assistive Agents for the Visually Impaired

The Ghost in the Keys: A Disklavier Demo for Human-AI Musical Co-Creativity

Tracing Generative AI in Digital Art: A Longitudinal Study of Chinese Painters' Attitudes, Practices, and Identity Negotiation

Transparency breeds confidence, not judgment

How was this newsletter synthesized?

Methodology

Selection Criteria

Key Themes Discovered

Field Report: ai-interaction

Trust, Calibration, and Collaboration

Top Papers in this Theme

Scientific judgment drifts over time in AI ideation

Revealing AI Reasoning Increases Trust but Crowds Out Unique Human Knowledge

SnappyMeal: Design and Longitudinal Evaluation of a Multimodal AI Food Logging Application

Generate, Evaluate, Iterate: Synthetic Data for Human-in-the-Loop Refinement of LLM Judges

When Machines Join the Moral Circle: The Persona Effect of Generative AI Agents in Collaborative Reasoning