← Back to Machine Learning (Statistics) stat.ML
When should a swarm of weak language models ask for help?
Prasanjit Dubey, Xiaoming Huo
May 27, 2026
Federated Conformal RAG combines weak language models in bandwidth-limited swarms while guaranteeing that retrieved answers stay accurate—but only worked at fixed stopping times. This work extends that guarantee to *any* moment: you can dynamically decide when to escalate retrieval bandwidth or refresh models without breaking coverage promises. The key insight is tracking a "betting budget" that stays positive even when models recalibrate adversarially. On MMLU, DBpedia, and AG News, the method flags true coverage failures exactly when they occur while cutting communication costs by 14–57%.
Read the original paper →