News and views on the sources and systems behind what we know and believe.

Six Misconceptions About Large Language Models: A Minimal Model and Diagnostic Taxonomy

2026-08-24 · arXiv (cs.CY)

This paper argues that public and expert debates about large language models are shaped by persistent 'folk theories,' including both deflationary slogans such as 'stochastic parrots' and anthropomorphic framings. The authors propose a minimal model and diagnostic taxonomy intended to identify and correct these misconceptions.

From Urban Mobility to Epidemic Dynamics: A Mixture-of-Experts Framework with Preference Alignment for Policy Scenario Simulation

2026-08-24 · arXiv (cs.CY)

The paper introduces UrbanShare-MoE-PA, a data-driven agent-level framework designed to model how non-pharmaceutical interventions alter epidemic risk through behavioral changes rather than simple aggregate mobility reductions. The framework translates alternative policy calendars into plausible activity and mobility trajectories for use in downstream epidemic outcome simulations.

ExploraTwin: A Non-Profit Research Platform for Digital Twin Simulations

2026-08-24 · arXiv (cs.CY)

This commentary introduces ExploraTwin, an open-access, non-profit research platform intended to lower barriers for researchers and practitioners testing and deploying digital twin survey simulations. The platform supports multiple simulation modes and is available at exploratwin.org.

Interaction Effects Between Learner Characteristics and Dialogue Format in TTS Dialogue-Based Lessons

2026-08-24 · arXiv (cs.CY)

This study investigated how learner characteristics interact with three LLM- and text-to-speech-generated dialogue formats—teacher–student, student–student, and teacher–teacher—affecting motivation, learning outcomes, and overall evaluation. The researchers focused specifically on the role of experiential learning style, particularly the Concrete Experience factor, as a moderating variable.

Fine-Tuning LLMs for Tourist Trajectory Prediction Using Field Experiment Data

2026-08-24 · arXiv (cs.CY)

The paper proposes fine-tuning large language models on field experiment data to predict tourist movement under varying contextual conditions such as weather and fatigue, addressing limitations of traditional mobility models. The approach leverages commonsense knowledge encoded during LLM pretraining to enable reasoning about context-dependent visitor decisions.

Atom Learning Model (ALM): How a Real Classroom Got Tokenised

2026-08-24 · arXiv (cs.CY)

The Atom Learning Model tokenizes a school curriculum by parsing two secondary mathematics textbooks into 1,934 discrete learning units called atoms, linked by 4,616 machine-generated prerequisite relationships. Both instructional questions and individual learner ability are represented within this unified graph structure, with learner proficiency scored between 0 and 1 for each atom.

Invisible Agents, Uninformed Patients: Towards Responsible Deployment of Autonomous AI Diagnostic Agents in Sub-Saharan Africa

2026-08-24 · arXiv (cs.CY)

The paper examines the deployment of autonomous AI diagnostic agents on eHealth platforms in sub-Saharan Africa, arguing that rollout has outpaced the governance infrastructure needed to oversee such systems. The authors contend that existing frameworks addressing AI accountability, transparency, and explainability in healthcare do not adequately address the specific conditions of this context.

Self-Reported AI Usage for Learning in Computer Science Education: Relationships with Goal Orientation and Academic Help-Seeking

2026-08-25 · arXiv (cs.CY)

This study examines associations between university students' goal orientation, academic help-seeking behavior, and self-reported AI use for learning in a computer science setting, drawing on data from 236 students. The research also accounts for individual, behavioral, and contextual characteristics as part of its analysis.

PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks

2026-08-25 · arXiv (cs.CY)

PersonaMem-v3 is presented as a benchmark and framework addressing the challenge of building AI agents that understand users across multiple digital contexts, including preferences, habits, social relationships, and evolving needs over time. The work targets gaps in measuring cross-context personal intelligence, which current single-app or single-task personalization systems do not address.

Determinants of Starting Salaries for Filipino Graduates: An Explainable Machine Learning Approach

2026-08-25 · arXiv (cs.CY)

The study applies explainable machine learning to a crowd-sourced, self-reported survey dataset of Filipino graduates to identify the determinants of starting salaries, addressing a gap left by descriptive tracer studies that document employment rates without explaining pay variation. The authors treat the noisy nature of the self-reported data as a methodological challenge to be addressed rather than avoided.

Interrupting the Chain: Human Perception of AI-Generated Disinformation Through a Kill Chain Lens

2026-08-25 · arXiv (cs.CY)

This paper reports empirical findings from a human-subject study in which 504 participants made 2,438 judgments classifying news fragments by origin (human vs. machine) and veracity (real vs. fake). Results are organized using an adapted cybersecurity kill chain taxonomy to map human perception data onto stages of a cognitive attack lifecycle and identify potential intervention points.

A Survey Instrument to Assess Students' AI and Generative AI Knowledge

2026-08-25 · arXiv (cs.CY)

This research-to-practice paper presents a survey instrument designed to assess students' knowledge of artificial intelligence and generative AI, responding to the proliferation of AI literacy frameworks and the need for corresponding assessment tools. The instrument is intended for use alongside existing AI literacy frameworks that outline essential knowledge for students.

Generative Gap Filling

2026-08-25 · arXiv (cs.CY)

The paper examines how courts fill gaps in contracts when textual interpretation is insufficient to resolve disputes, challenging the scholarly assumption that remaining contract text provides only thin evidence of the parties' actual agreement on disputed points. The authors argue that the existing contract text may offer more guidance for gap-filling than conventional accounts suggest.

Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?

2026-08-25 · arXiv (cs.CY)

The paper argues that the analogy between LLM performance on medical licensing exams and legal competence is misleading, because legal truth is contingent on jurisdiction, temporal validity, and source hierarchy rather than grounded in stable empirical reality. The authors investigate whether LLMs exhibit sycophantic or fragile behavior when confronted with juridical authority and evolving legal standards.

View full edition and past editions