Curated from knowledge graph (201 nodes, 244 edges) · All articles published within the last 7 days

• AI readiness on campus: How to strive for durability — University Business explores how higher ed institutions can build durable AI readiness programs, noting that employer-desired AI skills evolve rapidly and institutions must prepare students for continuous learning rather than static competencies. This aligns with the TritonAI program's focus on scaffolding sustainable AI adoption across campus. 🔗 Graph: higher-ed-ai, tritonai, brett-pollak 📅 Published: 2026-05-05 📰 https://universitybusiness.com/ai-readiness-on-campus-how-to-strive-for-durability/

• Jailbroken Frontier Models Retain Their Capabilities — New arXiv research demonstrates that advanced jailbreak methods on frontier LLMs impose effectively no performance degradation, challenging assumptions about the "jailbreak tax." This has direct implications for AI governance and security frameworks in enterprise deployments like TritonGPT. 🔗 Graph: ai-governance, ai-security, litellm-enterprise 📅 Published: 2026-05-05 📰 https://arxiv.org/abs/2605.00267

• When Do Diffusion Models learn to Generate Multiple Objects? — DeepMind researchers on arXiv investigate why text-to-image diffusion models struggle with multi-object generation, examining whether limitations stem from training data or architectural constraints. Relevant for understanding generative AI capabilities as institutions evaluate creative AI tools. 🔗 Graph: google, ai-adoption 📅 Published: 2026-05-05 📰 https://arxiv.org/abs/2605.00273

• Skills as Verifiable Artifacts: A Trust Schema for Human-in-the-Loop Agent Runtimes — arXiv paper proposing a trust framework for agent skills—structured packages that augment LLMs without model modification. This maps directly to the TritonAI App Builder's modular approach and the governance challenges of deploying agentic AI in higher ed. 🔗 Graph: agentic-ai, tritonai-app-builder, ai-governance 📅 Published: 2026-05-05 📰 https://arxiv.org/abs/2605.00424

• AI Adoption Among Teachers: Insights on Concerns, Support, Confidence, and Attitudes — arXiv study of 260 Philippines teachers examines factors driving AI tool adoption in education, finding institutional support and teacher confidence are key mediators. Parallels TritonGPT's campus-wide adoption strategy and the Service Desk integration work underway. 🔗 Graph: ai-adoption, higher-ed-ai, tritongpt 📅 Published: 2026-05-05 📰 https://arxiv.org/abs/2605.00343

• Rethinking Network Topologies for Cost-Effective Mixture-of-Experts LLM Serving — arXiv research questions whether expensive high-bandwidth scale-up networks are strictly necessary for MoE LLM serving, presenting cross-layer analysis of network cost-effectiveness. Relevant to on-prem infrastructure decisions and SDSC hosting strategy. 🔗 Graph: san-diego-supercomputer-center, enterprise-monitoring, infrastructure-topic 📅 Published: 2026-05-05 📰 https://arxiv.org/abs/2605.00254

• LLM-Oriented Information Retrieval: A Denoising-First Perspective — arXiv perspective paper argues that modern IR must be redesigned for LLM consumption rather than human users, as noise directly causes hallucinations and reasoning failures in RAG systems. Core concept for enterprise data agents and document AI layers. 🔗 Graph: onyx, enterprise-data-agent, model-context-protocol 📅 Published: 2026-05-05 📰 https://arxiv.org/abs/2605.00505

💡 Signal: This week's feed shows continued convergence of agentic AI research and higher ed implementation concerns—especially around governance, verifiable skills, and RAG reliability. The Berkeley/Stanford arXiv pipeline continues to dominate foundational research, while EDUCAUSE and University Business provide the institutional framing. Worth noting: multiple papers addressing the "how do we trust agents?" question, which will be central to the April 6 cabinet presentation and the upcoming Developer API Program launch.