Multiple research papers published on arXiv in late March 2026 examine multi-agent frameworks for improving large language model reasoning across different domains.
According to one arxiv.org paper, researchers introduced MAD-ACC (Multi-Agent Debate for Argument Component Classification), which uses a Proponent-Opponent-Judge model for argument mining. The system achieved a Macro F1 score of 85.7% on the UKP Student Essays corpus, “significantly outperforming single-agent reasoning baselines, without requiring domain-specific training,” according to the paper. The framework was accepted for publication at ACIIDS 2026.
However, another arxiv.org study revealed a counterintuitive finding about verification systems. The research found that adding verification improves outcomes when upstream feedback has less than 70% accuracy, but “degrades performance by 4-6 percentage points through over-specification when feedback is already reliable (>85%),” according to the paper. The study evaluated step-level feedback for propositional logic proofs using a benchmark of 516 unique proof states.
Meanwhile, a third arxiv.org paper presented PROClaim, a courtroom-style multi-agent framework for claim verification that integrates “Progressive RAG (P-RAG) to dynamically expand and refine the evidence pool during the debate.” According to the researchers, PROClaim achieved 81.7% accuracy on the Check-COVID benchmark, outperforming standard multi-agent debate by 10.0 percentage points.
A fourth arxiv.org paper described the Rhizomatic Research Agent (V3), a 12-agent pipeline designed for non-linear literature analysis in social sciences research.