Researchers have published four papers demonstrating multi-agent AI frameworks that apply large language models to specialized technical domains.
According to arxiv.org, AgenticRecTune introduces a five-agent framework for optimizing recommendation system configurations. The system uses Gemini to manage “pre-ranking, ranking, and re-ranking phases” through specialized agents: Actor, Critic, Insight, Skill, and Online. The framework includes a “self-evolving Skillhub” that summarizes historical results and updates skills based on A/B test outcomes.
In a separate arxiv.org paper, researchers presented MetaSymbO, a multi-agent system for metamaterial discovery that interprets natural language design intents. The framework comprises three agents—Designer, Generator, and Supervisor—and introduces “symbolic-driven latent evolution” for structural refinement. According to the paper, MetaSymbO “improves structural validity by up to 34% in symmetry and nearly 98% in periodicity compared to state-of-the-art baselines” and achieves “about 6-7% higher language-guidance scores” than advanced reasoning LLMs.
A third arxiv.org paper describes PRISM, a three-stage pipeline for large multimodal models that addresses “distributional drift” between supervised fine-tuning and reinforcement learning. The system uses “on-policy distillation” with a Mixture-of-Experts discriminator featuring “dedicated perception and reasoning experts.”
Finally, according to arxiv.org, PBio-Agent tackles biological perturbation prediction through “difficulty-aware task sequencing with iterative knowledge refinement.” The framework introduces LINCSQA, a benchmark for predicting gene regulation under chemical perturbations in bulk-cell environments, and “outperforms existing baselines on both LINCSQA and PerturbQA.”