We investigate whether explicit belief graphs improve LLM performance in cooperative multi-agent reasoning. Through 3,000+ controlled trials across four LLM families in the cooperative card game Hanabi, we establish four findings. First, integration architecture determines whether belief graphs provide value: as prompt context, graphs are decorative for strong models and beneficial only for weak models on 2nd-order Theory of Mind (80
Community college adjunct faculty, who teach a majority of first-year courses, face varied challenges related to limited access to ongoing support. This qualitative study examined the effects of an instructional coaching program on community college adjunct faculty's selfefficacy and ability to enhance instructor presence in online courses. Using a multiple case study research design, we investigated the impact of an instructional coaching program on two earlycareer (<7 years) and two mid- to late-career (>12 years) online adjunct faculty. Self-Efficacy Theory (Bandura 1997, 2005), the Community of Inquiry model (Garrison 2016a), and Richardson et al.'s (2015) operationalization of instructor presence provided the interpretive frameworks for analyzing coaching session data and faculty-student interactions recorded in preand post-intervention archived courses. Findings revealed self-efficacy gains across career stages, with early-career faculty demonstrating a stronger willingness to adopt new technologies such as GenAI, resulting in significant improvements in their self-efficacy and application of newly acquired skills. Mid- to late-career faculty expressed higher perceived confidence from their extensive face-to-face teaching experience, with a more cautious and thoughtful approach to implementing new teaching strategies in their course. These results suggest that when ongoing professional development is designed to meet the unique needs of adjunct faculty across different career stages, it can improve their sense of institutional belonging and retention.
The primary objective of this ethnographic systematic review and meta-analysis of language access interventions and healthcare outcomes among patients with Limited English Proficiency (LEP) was to determine whether system-level language access interventions improved healthcare outcomes for LEP patients compared with English-proficient populations. We included comparative quantitative studies conducted in hospital or health-system settings that evaluated LEP populations and reported patient-related outcomes. Exclusion criteria were non-peer-reviewed reports, artificial-intelligence-focused interventions, case studies (<50 participants), animal studies, and community-only studies. PubMed, Scopus, and the Cochrane Library were last searched on 25 September 2025. Risk of bias was assessed independently by 2 reviewers using ROB-2 and ROBINS-I V2. Random-effects meta-analysis following the PRISMA 2020 guidelines, and narrative synthesis via meta-ethnography in compliance to the eMERGe guidelines, were used to summarize outcomes. 45 studies (n = 618,822 participants) were eligible and included for meta-analysis, while 32 were deferred to a qualitative meta-ethnography synthesis due to a lack of numerical outcomes; studies were often observational cohorts and randomized controlled trials, with high risk of bias. Results, sorted into 12 subgroup analyses by outcome measure and effect type from, demonstrated a consistent trend toward worsened health and access disparities and poorer clinical outcomes among limited English proficiency (LEP) populations compared to non-LEP groups with language barriers and poor intervention, although effects varied substantially by intervention type and outcome. Subgroups 1(a) (n = 2) and 2(a) (n = 3) showed small, non-significant increases in disparities (pooled effects: -0.769, 95% CI [-4.274, 2.736], p < 0.001; and -0.492, 95% CI [-2.150, 1.166], p < 0.001, respectively), favoring non-LEP populations. Clinical outcomes in subgroup 1(b) (n = 2) demonstrated a larger negative effect (-7.829, 95% CI [-24.561, 8.902], p < 0.001), also favouring non-LEP groups. Health system performance outcomes were mixed. In subgroup 3(a) (n = 2), wait times and delays increased (-1442.426 minutes, 95% CI [-4260.021, 1375.169], p < 0.001), indicating worse performance, while subgroup 3(b) (n = 3) showed increased length of stay and revisits (510.533 minutes, 95% CI [-2015.983, 3037.050], p < 0.001). In contrast, language-focused system redesigns in subgroup 4(a) (n = 5) were associated with significant improvements (0.541, 95% CI [0.143, 0.939], p < 0.001), favouring intervention groups, whereas subgroup 4(b) (n = 2) showed a non-significant improvement in disparities (0.323, 95% CI [-0.633, 1280000], p < 0.001). Disparities increased significantly in subgroup 5(a) (n = 10) (-1.149, 95% CI [-1.532, -0.766], p < 0.001), with smaller, non-significant negative effects observed in 5(b) (n = 7) (-0.362, 95% CI [-1.474, 0.750], p < 0.001) and 5(c) (n = 3) (-0.537, 95% CI [-2.075, 1.001], p < 0.001), all favouring non-LEP populations. Similarly, subgroup 6(a) (n = 2) showed a non-significant decrease in positive clinical outcomes (-0.333, 95% CI [−0.991, 0.325], p = 0.009), while subgroup 6(b) (n = 2) demonstrated a significant worsening of disparities (-0.206, 95% CI [-0.274, -0.138], p = 0.98), again favouring non-LEP groups. Heterogeneity was high across most subgroups (I² range: 85%-100%), except for subgroup 6(b), which showed no observed heterogeneity (I² = 0%), indicating consistent effects across included studies. In the meta-ethnography, language was found to be a structural determinant affecting care processes at multiple levels, and the treatment of interpreters as episodic support reduces consistency of positive intervention outcomes. Process-level reforms must be combined with patient-level strategies, such as ethnic matching and equitable perception of disparities. Heterogeneity in this literature is driven by contextual and implementation factors rather than study quality, and language-related access disparities carry consistent pooled signals when they do not measure clinical or time-based outcomes. Patient-reported outcomes, including perceived linguistic safety, communication quality, and care satisfaction, represent an underdeveloped dimension of the evidence base. Institutional de-centering led to poor implementation and informal interpretation contributed to underestimate of adverse events. Future research should prioritize several methodological advances to address the limitations identified in this review, including quasi-experimental trials and culturally and linguistically sensitive standardization of LEP measurement and documentation. Evidence was limited by predominantly observational designs, high heterogeneity and wide confidence intervals in some analyses, small subgroup sample size, and moderate-serious risk of bias. This work was funded by ThinkNeuro, LLC (EIN: 33-2258068) and this review was registered on OSF Registries (7dvkb).
