Signal
Loading the stream…
WDSF 2026 results are on record — 10 awards · 11 winnersSee the record →
Loading the stream…
What’s moving in agentic workstations and workflows — drawn from a reviewed source list, scored, and kept at a permanent address you can cite.
Curated and full layers · newest first · scored, sourced, citable
The ledger as a map. Dashed edges are machine-suggested (embedding similarity and duplicate clusters); solid edges are editorial — they appear only where a blog post cites an entry.
Raw data: signal-graph.json
Filter by category
275 entries in the full stream matching the current filters
FGLGuard is a federated graph learning framework for safety in LLM-based multi-agent systems. It trains a graph attention detector on episode graphs locally per organization, sharing only model updates, and outperforms centralized in-domain baselines on Agent-SafetyBench, R-Judge, and AgentDojo while cutting attack-success rate by 43% at near-unguarded utility.
Cross-organization MAS safety with measurable federated-vs-centralized tradeoffs; relevant for anyone deploying multi-agent systems across trust boundaries.
ArXiv paper experimentally localizing reliability properties in an agent system built around an append-only ledger adjudicator. Interventions on institutional epistemic mechanisms and on cognition (ablation, mid-task reset, frontier-LLM substitution, false testimony) left five core properties intact: singular accepted reality, typed refusals, durable duties, no double-acceptance, and no false completions across 2,581 substituted-panel claims.
Provides empirical evidence that reliability in agent systems can live in institutional machinery rather than in the model, a structural claim with direct design implications for agentic workflows.
Proposes the EluDA framework for exploratory analysis of unstructured image data, combining querying, visualization, and active knowledge construction. A formative user study found users build faceted classifications bottom-up. Evaluation showed CLIP is unreliable for user-defined concept assignment but supports semantic grouping, leading to four identified human-AI collaboration opportunities.
Empirical formative study of how people actually structure image exploration with CLIP, worth reading if designing or evaluating AI-assisted data analysis tools. Concrete findings rather than generic vision.
Introduces Agent Flight Recorder, an audit system for long-horizon tool-using AI agents that captures each action as a structured event with eight semantic fields. Uses hash chaining and Merkle batching for tamper evidence, with periodic on-chain anchoring of epoch roots. Evaluation across five ablations shows ~48μs per-event latency, 512 bytes per event, $2.30 per 100K events for L2 anchoring, and 100% tamper detection with zero false positives.
Concrete numbers on a tamper-evident agent audit stack with on-chain anchoring. Useful for anyone designing accountability layers for production multi-agent systems.
Researchers formalize the 'epistemic Sybil problem' in multi-agent AI systems, showing that multiple agents sharing a common evidence root do not produce independent observations. Controlled experiments with over 20,000 LLM-agent calls demonstrate that naive aggregation collapses posterior coverage from 0.94 to 0.26 as report count rises from 1 to 32, while correlated extraction errors further degrade calibration.
Challenges the assumption that more agents equals more evidence, with empirical calibration data showing how multi-agent aggregation can mislead without tracking evidential ancestry.
A research paper proposing five design principles for multi-agent LLM systems in 6G networks, framed via Theory of Mind and cellular sheaves. Treats inter-agent messages as subjective reasoning traces rather than objective facts, and validates the framework on telecom language models under signaling-storm conditions.
Useful for builders of multi-agent LLM systems where hallucination propagation matters. The messages-as-traces framing and the depth-two ToM bound generalize beyond telecom.
Field study with 640 customer-facing employees across two randomized experiments. Finds that information retrievability — not just reviewer capability or engagement — is a precondition for catching LLM errors. Self-generated explanations improve detection; retrieval cues sustain it under repeated use.
Reframes LLM oversight failure as an information-access problem rather than a training or motivation problem, with actionable onboarding and cueing interventions backed by sizable field experiments.
The paper proposes control-data flow separation for multi-agent LLM systems: execution-critical protocols become typed, validated program objects while task content remains optimizable natural language, preventing prompt edits from corrupting routing or formatting logic. Tested on reasoning, review, and insurance rating workflows with 100% protocol validity.
Identifies a concrete failure mode in multi-agent prompt optimization and offers a typed separation pattern with empirical validation across three domains.
A three-week diary study with 8 blind screen-reader users evaluated OLLA, a computer-use agent prototype, across 1,258 commands in 12 desktop applications. GPT-5 achieved the highest success rate at 52.5%, with trace analysis identifying grounding, planning, constraint-tracking, and termination failures.
First diary study with blind users on real desktop workflows; names specific failure modes and benchmarks five models, useful for agent builders and accessibility practitioners.
Research paper presenting RecalibrateGPT, an interface system that addresses AI fatigue in LLM conversations through five cross-turn operators (Anchor, Replay, Delta, Scope, Steer) accessed via a single AssistiveButton. A study with 12 users identified four fatigue types and showed halved perceived cognitive workload (NASA-TLX 2.7) at high usability (SUS 86.5).
Frames AI fatigue as an interaction-flow cost rather than a model-quality problem, and offers a transferable operator taxonomy worth examining before building any LLM interface.
ErgoAssist is a head-worn ergonomic system that combines IMU-based posture tracking with consumer-grade EEG to estimate cognitive load. By detecting the user's mental state, it issues posture alerts only when the user is unlikely to be in deep focus. Lab results show 81% posture classification accuracy and 81% alert reduction alongside 38% better posture correction.
Couples posture detection with cognitive load to fix the core failure of ergonomic wearables: interrupting during focus. The 81% alert reduction with better outcomes reframes alert design as a context problem.
Research prototype of a fabric water-bottle sleeve with sensors and a small display showing a virtual pet. Drinking, standing, and refilling act as pet-care actions. A 20-student two-week study reported higher water intake and more movement episodes, alongside noted design tensions around guilt and focused work.
Concrete first-deployment data on a novel pet-based desk-side wellness device, with explicit design tensions flagged. Useful for anyone prototyping habit-formation hardware for desk workers.
A gesture elicitation study co-designed with geography and migration researchers produced a taxonomy of user-defined gestures using active tangibles and styluses for 16 geospatial data visualization tasks in multi-surface environments, plus a thematic analysis of user experiences with the interaction techniques.
Offers a user-validated gesture vocabulary for tangible and stylus input on geospatial visualizations, relevant to designers of collaborative MSE tools and large-display data work.
Proposes a 'cognitive cell' framework: a fixed, frozen small language model as a constant unit for measuring how population size, communication topology, message bandwidth, and coordination protocols shape collective behavior. Experiments with 1.5B and 3B models show gains only when cell errors are uncorrelated, and simple voting outperforms debate, shared blackboard, and chain revision at matched cost.
Empirical pushback on multi-agent hype: freezing a 1.5–3B LM as a fixed unit lets you measure when adding cells actually helps, and shows debate-style protocols losing to plain voting at equal cost.
AREAs-Lab presents an interactive benchmark for studying AI-driven requirement elicitation. An assistant iteratively refines incomplete requirements by analyzing datasets and asking clarification questions. The benchmark spans 16 public datasets, user profiles, and an automated evaluation pipeline using AI-simulated users to assess elicitation quality.
Academic testbed for studying how AI assistants convert vague goals into actionable requirements. Relevant to researchers building elicitation tools; limited immediate utility for practitioners.
Provenance-aware pipeline merges system telemetry and human-authored reports into a shared typed task-state representation for human-AI task handover. Evaluation on 13 paired task states in a controlled spatial environment shows structured reconciliation preserves more task-state utility than either source alone, and incurs less misinformation than an end-to-end LLM given the same inputs.
Empirical comparison showing raw LLM handover introduces measurable misinformation; a design pattern worth borrowing when delegating state to agents.
Research paper comparing ML and DL models for classifying balanced versus imbalanced postural states in VR using kinematic, EMG, and EDA signals. A Mamba-inspired CNN reached 96.76% accuracy; SHAP analysis showed kinematic features dominated and that a 33% feature reduction preserved performance.
Niche VR balance-detection study with code release; useful as a reference for multimodal posture sensing but only tangentially relevant to conventional desk ergonomics.
An interview study of 12 students examines how they delegate professional emails to generative AI, finding that AI involvement removes the writing skill learning loop, decouples messages from their specific recipient, and redirects confidence from the writer to the system. Two risks are identified: stunted individual capacities and added labor of managing authenticity.
Worth reading for its side-by-side model of AI-mediated versus unaided writing, and the specific framing of who captures the returns from cognitive delegation.
Academic paper proposing a preserved evidence-bundle model and policy-controlled offline verifier for cross-organization agent-to-agent workflows. Each bundle carries authentication, log commitment, witness-backed checkpoints, continuity proofs, delegation authorization, and signed receipts. Prototype tested on 300 workflows; corrupted bundles correctly rejected.
Addresses a concrete gap in multi-agent accountability: offline, policy-controlled verification without trusting live systems. Useful reference for anyone building auditable cross-organization agent pipelines.
FedEHR-Agents proposes a federated framework where autonomous clinical LLM agents at different hospitals collaborate by sharing modeling experience and meta-prompts rather than patient data, using TextGrad-based prompt refinement and evidence-guided aggregation to improve automated EHR modeling across multi-hospital benchmarks.
Frames experience and prompt sharing as a federation object for multi-agent systems under privacy constraints, a transferable pattern beyond the clinical setting.
A position paper arguing that human-centered explainable AI should incorporate psychology of information-seeking. It proposes users evaluate explanations via instrumental, hedonic, and cognitive expected utilities, shaped by cognitive biases, leading to either excessive or insufficient information-seeking with agentic AI systems.
Provides a framework for designers of agentic tools to move from making explanations available to making them sought, grounded in documented cognitive biases rather than assumptions about user behavior.
An arXiv paper examines AI-assisted sensemaking of qualitative UX data through a case study of 20 user responses on video conferencing platforms. It proposes a framework positioning AI as an amplifier of human interpretive judgment rather than a replacement, identifying when algorithmic efficiency helps versus hinders designer agency.
Provides a concrete framework and case study for integrating AI into qualitative research workflows without losing interpretive depth. Useful for UX practitioners and researchers designing AI-assisted analysis pipelines.
Proposes the SCAN framework (Substitute, Complement, Aid, Non-Negotiable) for allocating clinical reasoning tasks between medical learners and generative AI. Reframes learner difficulty as misclassification of interaction mode rather than misuse, and identifies passive engagement as a detection-resistant path to mis-skilling.
Names a concrete failure pattern (passive engagement in correctly scaffolded tasks) and supplies a testable decision vocabulary. Useful for educators and curriculum designers, though the framework is conceptual and unvalidated.
A mixed-methods user study examining how people choose between world-anchored and body-anchored mixed reality interface elements across stationary and mobile contexts, finding anchoring preferences shift with mobility and depend on personal factors like accessibility, stability, and visual clutter.
Empirical data on MR anchoring trade-offs that can guide adaptive interface design for spatial computing workflows.
The paper shows that Learning to Defer (LtD) strategies create class-dependent sampling bias, disproportionately routing minority-class items to human experts. A user study (N=226) finds that imbalanced deferred item sets trigger human cognitive bias, reducing accuracy in majority-class classification.
Concrete evidence that how an AI chooses to defer decisions can distort human judgment downstream, relevant for designing human-AI collaboration workflows.
A user study (N=204) introduces Graphionale, a system that converts linear LLM rationales into interactive multi-level argument graphs. Graphical rationales improved trust calibration for verbal reasoning but felt more demanding; for visual reasoning, they impaired calibration yet felt more engaging and helpful.
Empirical rule for matching rationale format to task modality, with a working testbed and a 204-participant study behind the claim.
Research paper drawing on two studies of experienced Wikipedia editors that proposes an AI-Assisted Sensemaking Model for collaborative discussions, capturing arguments, group norms, participants, and shared history, with a spectrum of interpretive support from low to high.
Identifies a real gap: most AI discussion tools ignore the social context that gives group conversations meaning. The model and its tradeoff analysis are directly useful for anyone designing AI-mediated collaboration tools.
MemToC is a controlled benchmark for evaluating how LLM agents arbitrate between parametric memory and tool returns when the two conflict. Across five 7-9B open-weight models, instruction-tuned variants retain a verified-correct answer against an incorrect tool in only 6.5-17.1% of cases. SFT and DPO improve arbitration on some backbones but reduce abstention after tool errors.
Quantifies a specific failure mode in tool-using agents and offers a benchmark plus fine-tuning baseline for anyone designing agent arbitration policies.
Paper tests a zero-shot manager-worker multi-agent scaffold across nine LLMs on 100 hard LiveCodeBench problems without training. Gains are model-dependent: up to +42 points for some models, null or negative for others. The scaffold roughly triples token cost but is often cheaper per unit of accuracy than scaling to a larger model.
A controlled, per-model breakdown of when multi-agent scaffolding actually helps coding — directly useful for deciding whether to add orchestration to a coding workflow.
A research paper introducing Dynamic Tree Colors, a dynamic hierarchical color map for visualizing large hierarchical data. It balances discriminability and color stability via configurable tradeoffs, with defined quality metrics and an 18-participant user study, though it underperforms the Cuttlefish algorithm in that algorithm's specific scenario.
Offers a configurable color-mapping algorithm and explicit quality metrics for hierarchical visualization, useful for analysts working with large tree-structured datasets.
RegulAR is an AR task assistant that models procedural tasks as hierarchical dependency graphs and combines this with a multimodal LLM to track progress, classify errors by type, estimate their downstream impact, and deliver recovery guidance through a head-up display. A 12-person within-subject study reported better task-structure understanding and recovery support versus an MLLM-only baseline.
Concrete pattern for pairing structured task graphs with an MLLM so agents can reason about procedural deviations, not just the next step. Relevant to anyone designing workflow agents that must detect and recover from user errors.
Research paper designing user-centric Chain-of-Thought reasoning that uses XML-like tags to structure reasoning traces into self-contained, verifiable steps. Evaluation on mathematical reasoning tasks shows equivalent performance to standard CoT while improving perceived usefulness and ease of use in user studies.
Tests whether tagged, self-contained reasoning steps help users inspect and correct LLM outputs without losing accuracy on math tasks.
A peer-reviewed study examining how prolonged standing in occupational footwear affects perceived discomfort, standing balance, and gait biomechanics in young adults.
Quantifies footwear effects on balance and discomfort during extended standing, relevant for sit-stand desk users and workplace ergonomics guidelines.
Academic paper applying the REBA postural assessment tool alongside organisational analysis to evaluate biomechanical overload risk among physiotherapists, published in the journal Ergonomics.
Niche peer-reviewed ergonomics study; methodology may interest readers designing workplace risk assessments, but population (physiotherapists) and empty available content limit direct desk-work applicability.
Peer-reviewed study in the journal Ergonomics examining an ergonomic intervention aimed at reducing physical and physiological stress associated with office standing workstations. Full content not available.
Adds a primary-source data point to the standing-desk ergonomics record, though the intervention's specific contributions cannot be assessed from metadata alone.
Peer-reviewed study comparing complexity-based eye movement metrics against conventional features for recognizing mental workload, evaluating their relative performance in classification tasks.
Empirical head-to-head of feature families for workload detection, useful for researchers selecting eye-tracking inputs for HCI or adaptive systems.
Peer-reviewed study examining how strategic music listening affects subjective attention and workload across tasks of varying complexity, contributing empirical data to the music-in-workplace literature.
Useful empirical input for readers deciding when music helps or hurts focused desk work; topic is familiar but the 'strategic' framing across task complexity adds nuance.
A scoping review published in the journal Ergonomics surveying emerging technologies and automation methods for musculoskeletal ergonomic assessments, synthesizing recent developments in the field.
Maps the current landscape of automated ergonomic assessment tools; useful for practitioners selecting methods and for researchers identifying gaps.
Peer-reviewed article in the journal Ergonomics examining how cognitive and perceptual science can inform typeface selection for reading tasks.
Cited academic findings on typeface legibility offer evidence-based guidance for selecting fonts in documents, code, and presentations.
A peer-reviewed article in the journal Ergonomics proposing a 4S framework for investigating Human Factors and Ergonomics research. Only the volume, issue, and page range are provided, with no abstract or findings available.
A meta-framework for mapping the HFE research field, relevant to researchers but the bare citation gives no actionable detail without the full text.
Proposes a conceptual architecture using Model Context Protocol (MCP) as a unified transport and schema layer between platform-specific accessibility APIs (Windows UI Automation, macOS Accessibility, Android, ARIA) and LLM-based screen reader agents, with analysis of protocol extensibility, latency-versus-semantic trade-offs, and persistent user accessibility profiles.
Conceptual paper, not an implementation. Offers a structured comparison of accessibility APIs and a concrete MCP-based design reference for cross-platform assistive agents.
Paper introduces Agentic World Analysis (AWA), a framework combining agentic AI, simulation modeling, and expert elicitation to study complex socio-environmental systems. Authors implement WEGA and test it on the Netherlands Nitrogen Crisis, producing two 2026–2041 scenario pathways evaluated for coherence and quality.
Concrete pattern for structuring agentic AI as an auditable expert panel on complex policy problems. The iterative scenario-tree loop and self-built context are transferable to other domains.
Peer-reviewed paper in Applied Ergonomics proposing an interrogative approach that uses physiological measures to assess and diagnose team performance, authored by researchers including Eduardo Salas.
Contributes empirical grounding for evaluating team states through physiological signals rather than self-report alone. Useful for researchers and designers of collaborative work systems.
Aura is a framework that estimates user emotional state from facial expressions, uses a probabilistic belief model to select interventions, and applies LoRA adapters to modulate LLM output mid-turn. A within-subjects study (N=20) on information-seeking tasks reported higher normalized perceived learning gains than a Llama-3 baseline and 21% lower interaction time versus GPT-4o and Llama-3 baselines.
Concrete architecture and measured interaction-time gains for real-time emotion-aware LLM adaptation, relevant to designing responsive agent workflows that adjust mid-turn rather than between turns.
Peer-reviewed paper in the journal Ergonomics examining how dynamic lighting affects neurobehavioral performance across different mental states in a space station working area.
Rare peer-reviewed look at lighting and cognition in a confined, extreme environment; a reference point for alertness-related design claims, though not directly transferable to office desks.
A peer-reviewed study examining the validity of EEG and fNIRS as physiological measures of situation awareness during performance and automation monitoring in a complex task, published in Ergonomics.
Useful methodological reference for researchers validating neurophysiological measures of situation awareness; limited direct application to desk work practices.
An editorial or review article from the journal Ergonomics surveying the current state of research on inclusive design and digital health interventions, published by Taylor & Francis. No detailed content or findings were provided in the source.
Broad state-of-science piece on inclusive design and digital health; relevance to desk-work ergonomics is indirect and the source carried no extractable detail.
A peer-reviewed study presenting a probabilistic regression framework to automatically estimate Hand Activity Level (HAL) from upper-limb motion trajectories, replacing manual ergonomic rating of repetitive work.
Automates a standard ergonomic exposure metric from motion data, cutting the manual effort behind HAL ratings in workplace assessments.
A peer-reviewed review paper examining single-human multiple-robot (SHMR) systems, analyzing their effects on occupational safety, worker health, and overall system performance. Published in Ergonomics, Volume 69, Issue 8, 2026.
Synthesizes existing research on one-operator-many-robot configurations, useful for designers and ergonomists planning human-robot collaborative workstations and safety protocols.
Peer-reviewed validation study examining whether force-estimating insoles can accurately calculate center of pressure and vertical ground reaction forces during occupational tasks, published in Ergonomics (Vol 69, Issue 8, 2026).
Methodological reference for researchers selecting wearable foot-pressure instruments in occupational biomechanics; narrow utility for general workstation readers.
