NightFeats @ MMU-RAGent NeurIPS 2025: A Context-Optimized Multi-Agent RAG System for the Text-to-Text Track

NightFeats, a multi-agent retrieval-augmented generation system, won Best Dynamic Evaluation at NeurIPS 2025's MMU-RAGent competition by prioritizing architectural transparency and evidence grounding over benchmark optimization. The system outperformed proprietary models like Claude-SonnetV2 and Nova-Pro through a three-phase pipeline combining retrieval, curation, and composition with explicit intermediate representations.

NightFeats represents a meaningful shift in how AI systems approach knowledge synthesis and evaluation. Rather than chasing higher scores on automatic similarity metrics, the system demonstrates that human judges and real-world performance reward verifiable, transparent reasoning processes. This distinction matters because it reveals a growing divergence between what benchmark metrics measure and what users actually value in AI outputs.

The competition victory reflects broader industry recognition that opaque, high-performing black boxes carry inherent risks. The three-phase decomposition—retrieval, curation, and composition—creates clear handoff points where errors can be identified and corrected. Temporal-semantic reranking and bounded contradiction reconciliation ensure that retrieved information is not merely relevant but contextually appropriate and internally consistent. Citation-preserving composition means users can trace claims back to sources, building trust through verifiability.

For AI practitioners and organizations deploying RAG systems, this work suggests that architectural choices favoring interpretability can yield competitive advantages. The human evaluation preferences demonstrated here indicate that end-users prioritize trustworthiness and explainability alongside quality. This has implications for enterprise adoption, where regulatory compliance and auditability increasingly matter. The result also challenges the notion that proprietary models automatically outperform structured, modular approaches—technical design can offset scale advantages. As RAG systems become more prevalent in production environments, the principles embedded in NightFeats—explicit contracts between components, bounded error reconciliation, and evidence grounding—may become standard best practices rather than novel differentiators.

• →Architectural transparency and evidence grounding outperformed proprietary baselines on human evaluation metrics at NeurIPS 2025
• →Three-phase decomposition with explicit intermediate representations enables error identification and system auditability
• →Structured multi-agent design can achieve competitive results without relying solely on model scale or benchmark chasing
• →Temporal-semantic reranking and contradiction reconciliation improve contextual relevance beyond raw retrieval quality
• →Citation-preserving composition builds user trust through verifiable source attribution and claim traceability