The Bits-over-Random (BoR) metric, highlighted in a recent ICLR blogpost, fundamentally shifts how practitioners view retrieval in RAG and agent systems. Author Victor Lavrenko, an Edinburgh-trained PhD in Information Retrieval, argues traditional metrics like recall and Success@K are incomplete. BoR measures retrieval selectivity relative to random chance, revealing when retrieved bundles degrade into context pollution – flooding the model with noise instead of delivering a clean, selective set.

This insight is crucial for agentic systems where the model processes the full retrieved set as prompt context. BoR helps identify when retrieval stops being genuinely selective, forcing the model to process irrelevant material and degrade performance.