HyMEM Advances GUI Agents While DxEvolve Improves Clinical Diagnosis

Advancements in AI are enabling more sophisticated human-computer interaction and complex decision-making across various domains. For GUI agents, Hybrid Self-evolving Structured Memory (HyMEM) enhances performance by coupling symbolic nodes with trajectory embeddings, boosting models like Qwen2.5-VL-7B by +22.5% and outperforming Gemini2.5-Pro-Vision and GPT-4o. In resource-constrained game AI, a hybrid framework integrating Graph Attention Autoencoders with LLMs like GPT-4o-mini achieved a 45.0%-66.5% win rate in Amazons, demonstrating weak-to-strong generalization. For LLM safety, IH-Challenge fine-tuning improved instruction hierarchy robustness by +10.0% and reduced unsafe behavior from 6.6% to 0.7%.

Continuous control tasks in AI-native networks are being addressed by self-finetuning agents that internalize experience into parameters, bypassing explicit rewards. This approach, evaluated on a dynamic RAN slicing task, outperforms standard RL and LLM-based agents in sample efficiency and multi-metric optimization. Similarly, agent execution trajectories are being leveraged for self-improvement; a framework extracting learnings from these trajectories achieved up to 14.3 pp gains in scenario goal completion on the AppWorld benchmark. For clinical diagnosis, DxEvolve, a self-evolving diagnostic agent, improved accuracy by 11.2% over backbone models on MIMIC-CDM, reaching 90.4% and outperforming competitive methods on external cohorts.

Evaluating AI reasoning is moving beyond scalar probabilities. TRACED assesses reasoning quality through geometric kinematics, distinguishing correct reasoning (high-progress, stable trajectories) from hallucinations (low-progress, unstable patterns). Imprecise probabilities are being used to verbalize higher-order uncertainty in LLMs, improving credibility in ambiguous settings. For distilling reasoning from large models, HEAL uses hindsight entropy-assisted learning to repair broken trajectories and overcome the 'Teacher Ceiling', significantly outperforming traditional SFT distillation. Furthermore, FAME offers formal abstract minimal explanations for neural networks, scaling to large models while reducing explanation size and runtime.

AI agents are being developed with a focus on encoding domain expertise. Nurture-First Development (NFD) uses conversational knowledge crystallization to progressively grow agents through interaction with practitioners, illustrated by a financial research agent. In prescription verification, PharmGraph-Auditor uses a hybrid knowledge base and a Chain of Verification (CoV) reasoning paradigm to transform LLMs into transparent engines for evidence-grounded auditing, enhancing safety and traceability. Finally, automated data product improvement is achieved through specialized AI agents in an optimization loop, balancing automation with human oversight.