Agent Rosetta Advances Science While FactorEngine Mines Investments

Researchers are developing advanced AI agents capable of complex reasoning, planning, and interaction across diverse domains. In scientific discovery, Agent Rosetta and FactorEngine are enabling autonomous protein design and quantitative investment factor mining, respectively, by integrating LLMs with specialized software and knowledge bases. For code generation, IQuest-Coder-V1 and petscagent-bench showcase progress in agentic software engineering and evaluating AI-generated HPC code, while SQL-ASTRA and TRUST-SQL focus on improving multi-turn Text-to-SQL capabilities. Safety and alignment remain critical, with MOSAIC and MAC introducing modular control tokens and multi-agent learning for compositional safety, and CritiSense offering a multilingual app for digital literacy against misinformation. The Vlasov-Maxwell-Landau equilibrium has been formally verified using an AI-assisted loop, demonstrating a complete AI-driven mathematical research process.

In embodied AI and robotics, OpenVLA is being enhanced for better linguistic generalization through synthetic instruction augmentation, and AsgardBench evaluates visually grounded interactive planning under minimal feedback. For smart homes and IoT, the DS-IA framework ensures safe and efficient AIoT interactions by separating intent understanding from physical execution, while VIGIL deploys edge-extended agents for enterprise IT support, reducing interaction rounds and speeding up diagnosis. For customer service, a framework is proposed to manage safety gaps arising from specialized AI agents composing capabilities dynamically. In healthcare, a dual-component framework optimizes hospital capacity during pandemics through patient relocation, combining prediction and simulation models.

Several papers address the challenges of memory and context management in AI agents. CraniMem offers a neurocognitively motivated memory design for long-running workflows, while NextMem and Compiled Memory focus on latent factual memory and compiling experience into agent instructions, respectively. POaaS provides minimal-edit prompt optimization for on-device sLLMs, improving accuracy and reducing hallucinations. The Context Alignment Pre-processor (C.A.P.) aims to enhance human-LLM dialogue coherence by pre-processing user input to align context. For multimodal agents, TraceR1 introduces anticipatory planning by forecasting trajectories, and SocialOmni benchmarks audio-visual social interactivity in omni-modal models. Research also explores the fundamental nature of attention in LLMs with the QV paradigm and investigates how AI agents acquire scientific taste from institutional traces, outperforming human experts.

Further advancements include the development of IRAM-Omega-Q for uncertainty regulation in artificial agents, and ARISE for agent reasoning with intrinsic skill evolution in hierarchical reinforcement learning. The challenge of data contamination in LLM benchmarks is addressed by rigorous audits, revealing significant performance gains due to training data leakage. Additionally, research explores persona-conditioned risk behavior in LLMs, demonstrating human-like cognitive patterns, and the development of CUBE, a standard for unifying agent benchmarks to reduce fragmentation. The need for robust AI governance is highlighted, with papers on runtime governance for AI agents, formal frameworks for capability-based AI systems, and the design of bounded autonomy for embodied AI in critical infrastructure.