Finding: An LLM-based educational agent (DBagent) significantly improved learning achievement over traditional instruction, produced a distinctive "Query-Evaluation-Query" cognitive verification loop in student interactions, and sustained engagement through a perception→satisfaction→engagement pathway confirmed via SEM.
Why it matters: Directly informs how AI agents function as learning tools versus passive chatbots, with implications for HCI design of AI systems, learning sciences, and behavior change (engagement/satisfaction loops).
Method: Quasi-experiment (n=313, 4 classes), lag sequential analysis (LSA) of interaction logs to reveal cognitive behavior sequences, plus structural equation modeling (SEM) of survey data — a notably rigorous multi-method design.
Jargon: *Lag sequential analysis (LSA)*: a method for identifying statistically significant sequential patterns in behavioral event streams; *autonomous LLM-based agent*: an LLM with tool use, memory, and goal-directed reasoning beyond simple chat; *SEM*: structural equation modeling, tests causal pathways among latent variables.
⚠️ Self-reported data only | No control group mentioned | Single study (not replicated) | Convenience sample (e.g., MTurk, students)
Xiaoru Teng, Ka Wai LAU, NG Tsz King et al. (7 authors) • PsyArXiv
Finding: Successfully replicates McKenzie et al. (2006), confirming that defaults function as implicit endorsements—people infer policymakers' preferences and recommendations from whatever option is set as default, even inferring attitudes when *no* default is set.
Why it matters: Directly relevant to teaching choice architecture and nudges—this replication strengthens the evidentiary basis that default effects work partly through perceived endorsement, a key mechanism students need to understand when designing behavioral interventions.
Method: Pre-registered replication (N=602, Prolific) of three original studies, using adjusted Cramér's V for effect size comparison across nine tests, plus an added no-default control condition as an extension.
Jargon: *Default effect*: tendency to stick with the pre-selected option in a choice set. *Choice architect*: person/institution that designs the choice environment. *Adjusted Cramér's V*: standardized effect size for categorical data comparisons.
⭐ Pre-registered | ⭐ Includes replication ⚠️ Convenience sample (e.g., MTurk, students) | Limited statistical detail in abstract
Xiaohe Bie, Bo Wang, Xinyu Long • Semantic Scholar
Finding: Proposes a four-layer integrative framework (micro-cognitive, meso-affective, macro-social, self-constructive) for understanding psychological theories in HCI, arguing psychology has shifted from post-hoc evaluation to proactive design.
Why it matters: Directly relevant to HCI teaching and synthesizes behavioral science, cognitive psychology, trust, and AI-human collaboration into a unified framework useful for structuring course content.
Jargon: *Mind symbiosis* = deep integration of human cognition with AI systems; *bidirectional theory of mind* = both humans and AI modeling each other's mental states; *agency erosion* = gradual loss of human autonomy through AI dependency.
Yi-Hsin Su, Xiamin Leng, Amitai Shenhav • PsyArXiv
Finding: Choice deferral (delaying decisions) is primarily predicted by the overall value of options rather than how difficult it is to identify the best option; people avoid choosing when options are incongruent with their choice goal (e.g., all options are bad when you want the best).
Why it matters: Directly relevant to behavioral science, decision-making, and choice architecture — understanding when and why people defer decisions informs nudge design, persuasive technology, and goal-congruent framing of choices.
Method: Six experiments across three domains (consumer goods, artwork, political candidates) with a dynamic decision model comparing option appraisal against a criterion vs. a time-limit model.
Jargon: *Choice deferral* = postponing a decision rather than making it now; *relative value* = difficulty of identifying the best option among alternatives; *choice goal congruence* = how well available options match what the decision-maker is trying to achieve (e.g., choose-best vs. choose-worst).
⚠️ No control group mentioned | Limited statistical detail in abstract
Harrison J. Goldwyn, Graham Johnson, Christopher Ibarra et al. (5 authors) • Semantic Scholar
Finding: The paper proposes using Active Inference (a probabilistic cognitive framework) to simulate how users read and misinterpret data visualizations, modeling Type 1 (fast/heuristic) and Type 2 (slow/analytic) agents to predict errors before empirical studies are run.
Why it matters: Directly relevant to HCI and visualization design, and the dual-process (Type 1/Type 2) cognitive modeling connects to behavioral science and cognitive biases in decision-making.
Method: Implements executable agent simulations using Active Inference to model visual search, belief updating, and cognitive effort trade-offs in bar-chart reading tasks.
Jargon: *Active Inference* — a Bayesian brain framework where agents minimize "free energy" (surprise/uncertainty) by updating beliefs and selecting actions; *tick-salience bias* — tendency to anchor on visually prominent axis markers rather than computing accurate averages; *in silico evaluation* — testing via computational simulation rather than human participants.
Jordan Gutterman, Gillian Gold, Peter Brusilovsky et al. (4 authors) • PsyArXiv
Finding: Students in online CS courses tend to follow the default sequential order of programming activities ("straight-line behavior"), and those who do complete more of the curriculum — suggesting interface design shapes learning navigation choices.
Why it matters: Directly relevant to choice architecture and nudges (default effects in learning interfaces) and learning sciences — the implied sequence acts as a structural nudge influencing student behavior.
Method: Interaction log data from 247 students across three undergraduate CS courses analyzed for navigation patterns.
Jargon: "Straight-line behavior" = tendency to select the next activity in the displayed sequence rather than jumping around; essentially a default-following effect in educational UX.
V. Pimenova, Chris Lee, Baramee Bhakdibhumi et al. (5 authors) • Semantic Scholar
Finding: Temporal video segmentation (breaking instructional videos into single-instruction chunks with pauses) reduces cognitive load for ADHD learners in computing education, equalizing performance between ADHD and non-ADHD participants.
Why it matters: Relevant to learning sciences and UX/interface design for educational technology, particularly around cognitive load reduction and inclusive design principles.
Method: Within-participants controlled study (n=27, 17 ADHD, 10 without) comparing standard vs. segmented instructional videos.
Jargon: *Extraneous cognitive load* = mental effort imposed by poor instructional design rather than the material itself; *UDL (Universal Design for Learning)* = framework for designing instruction accessible to diverse learners; *post-hoc* = intervention applied after content creation, not during.
⚠️ No control group mentioned | Single study (not replicated)
Finding: Motivations to watch gambling streams align with motivations to gamble (enhancement, coping, social), and viewing frequency of slots/esports/loot box streams predicts problem gambling severity via actual gambling behavior.
Why it matters: Tangentially relevant to behavior change design and addictive design patterns (loot boxes, engagement mechanics), but the focus is on gambling pathology rather than HCI, persuasive technology, or learning sciences.
Method: Structural equation modeling across pooled data from three online surveys (n=544).
⚠️ Self-reported data only | No control group mentioned | Single study (not replicated) | Limited statistical detail in abstract
Finding: Introduces "Narrative Response Models" as a validity tool for psychological measurement instruments, using self-efficacy as a case study to clarify how survey responses map to construct definitions.
Why it matters: Self-efficacy is a core construct in learning sciences, motivation, and behavior change research, but this paper is primarily a psychometrics/measurement methodology contribution rather than an applied behavioral science paper.
Ariel Zylberberg, Ian Krajbich, M. Shadlen • Semantic Scholar
Finding: Attention during decision-making may reflect post-decisional processes rather than causally driving choices, challenging the dominant attentional drift-diffusion model in preferential choice research.
Why it matters: Tangentially relevant to behavioral science and decision-making, but the focus is narrow (eye-tracking/gaze behavior in food choice tasks) and not connected to HCI, behavior change design, or learning contexts.
Method: Reanalysis of existing food-choice datasets testing predictions of attentional drift-diffusion models (aDDM).
Jargon: *Gaze-cascade effect* — tendency for gaze to shift toward the chosen option immediately before a decision; *aDDM* — attentional drift-diffusion model positing that attention multiplicatively amplifies subjective value during choice formation.