AI Advances Decision Making While Agents Improve Efficiency

Recent advancements in AI are enhancing complex decision-making and operational efficiency across diverse fields. In public health, an AI agent system designed for HPV vaccination in Japan provides verified information and analyzes social media discourse, achieving high scores for relevance and correctness in both single and multi-turn evaluations. For epidemic forecasting, the Mobility-Informed Causal Adapter (MiCA) integrates human mobility data to improve lightweight temporal models, reducing relative error by 7.5% across various disease datasets. In sports analytics, BoxMind, an AI expert system validated at the 2024 Olympics, analyzes boxing match footage to predict outcomes with 69.8% accuracy and provides strategic recommendations, contributing to significant medal wins. Furthermore, LLMs are being leveraged for complex planning and optimization; the LEG framework integrates expert knowledge with optimization for health facility location in Ethiopia, balancing coverage guarantees with qualitative guidance. For job seekers, AI agents enhance referral requests, with Retrieval-Augmented Generation (RAG) increasing predicted success rates by 14% for weaker requests without degrading stronger ones.

Autonomous agents are being benchmarked and refined for real-world applications. AgencyBench evaluates 6 core agentic capabilities across 32 scenarios requiring extensive tool calls and tokens, revealing that closed-source models outperform open-source ones (48.4% vs 32.1%). To enable automated agent creation, ReCreate uses interaction histories to systematically improve agents, outperforming human-designed agents and existing methods. For multi-agent systems, SCALE offers a low-cost framework for task-level workflow generation, reducing token usage by up to 83% while maintaining competitive performance. Constrained Temporal Hierarchical Architecture (CTHA) enhances coordination stability in multi-time-scale agent systems, reducing failure cascades by 47%. AdaMARP, an adaptive multi-agent interaction framework, improves role-playing immersion and adaptability in interactive narratives, with an 8B actor outperforming commercial LLMs. For reliable agentic search, Boundary-Aware Policy Optimization (BAPO) cultivates boundary awareness, significantly enhancing reliability without compromising accuracy.

AI is also advancing reasoning, efficiency, and interpretability. For long-context LLM serving, ORBITFLOW dynamically reconfigures KV caches to meet latency Service Level Objectives (SLOs), improving SLO attainment by up to 66% and reducing 95th percentile latency by 38%. In multimodal hate speech detection, TANDEM uses a tandem reinforcement learning strategy for interpretable, temporally grounded analysis, achieving a 30% improvement in target identification. For embodied exploration, MemoryExplorer fine-tunes multimodal LLMs via reinforcement learning to encourage active memory querying and proactive exploration, showing significant advantages in long-horizon tasks. Efficient deep reasoning is addressed by Think-with-Me, a test-time interactive paradigm that uses external feedback to guide reasoning, achieving superior accuracy with reduced reasoning length. Hyperparameter optimization for constraint programming solvers is automated by a probe-and-solve algorithm, yielding robust improvements over default configurations. Policy-based deep reinforcement learning hyperheuristics are applied to job-shop scheduling problems, outperforming traditional and recent neural network methods. Finally, LLMs implicitly encode psychologically grounded trust signals, offering a foundation for credible AI systems, and demonstrate higher-order reasoning in predictive process monitoring for small-scale event logs, surpassing benchmarks in data-scarce settings.