Separating Semantic Competition from Context Length in RAG: Why Readers Fail Isn't What You Think

arXiv·medium signal

Repantis et al. introduce a matched-control protocol that disentangles two failure modes in RAG readers: longer context vs. semantically competing passages. By keeping passage count and length constant while varying semantic competition, they show that reader accuracy drops from genuine competition — not just context length. This finding reframes the RAG reader problem and suggests optimization should target passage selection quality, not just context window size.