Researchers are exploring multi-agent debate systems to improve the efficiency of large language model safety evaluation, according to a new paper on arXiv (arXiv:2511.06396v3). The study investigates whether structured multi-agent debate can enhance judge reliability while reducing costs, as LLM-as-a-judge pipelines currently rely on expensive models for safety assessments.
In the domain of code generation, another arXiv paper (arXiv:2603.15707v1) introduces SEMAG, a self-evolutionary multi-agent system designed to address limitations in current LLM programming approaches. According to the abstract, existing methods depend on manual model selection and fixed workflows, which the researchers argue restricts adaptability to varying task complexities. The paper states that large language models have made significant progress in complex programming tasks, though these structural limitations remain.
Meanwhile, researchers are raising concerns about LLM benchmarking practices for low and medium-resource languages. A paper focusing on Icelandic language evaluation (arXiv:2603.16406v1) identifies problems with current benchmarking methods and calls for improved evaluation approaches. The study specifically highlights issues with benchmarks that include synthetic or machine-generated content, suggesting these may not accurately assess model performance for such languages.