Researchers Advance Agentic AI While Introducing Safety Benchmarks

Researchers have made significant advancements in agentic AI, with various studies focusing on tool-augmented reasoning, reinforcement learning, and large language models. A key finding is that tool-augmented reasoning does not always outperform native CoT, and a critical tradeoff exists between the gains from tools and the 'tool-use tax'. Another study proposes a framework for assessing and optimizing LLM tool calling, highlighting the importance of necessity, utility, and affordability. Additionally, researchers have developed a continuous benchmark for measuring inference at endpoint granularity, and a methodology for tracing the functional role played by AI in natural language generation.

The use of large language models in military contexts has raised concerns about safety and alignment with military doctrines. A new benchmark, ARMOR 2025, has been introduced to evaluate LLM safety in military-aligned scenarios. The benchmark is grounded in three core military doctrines and features a structured taxonomy and rigorous evaluation procedures. Researchers have also made progress in understanding jailbreak success in LLMs, introducing a method called LOCA that provides local, causal explanations of jailbreak success.

Other studies have focused on improving the performance of agentic AI systems, including the development of a framework for instance-aware parameter configuration in combinatorial optimization and a method for learning where to click from self-supervision in GUI grounding. These advancements have the potential to improve the reliability and efficiency of agentic AI systems in various applications.