A new research paper diagnoses a basic flaw in retrieval-augmented AI: the model reads the question and skims the sources.
Researchers describe a failure mode they call query dominance, where a RAG system's encoder favors the user's question so heavily that retrieved documents barely shape the answer. Their fix, GRIP (Grounded Reasoning via Information-Restricted Premises), gives the decoder full access to the query but forces retrieved evidence through a severe stochastic bottleneck, so retrieved text can only pass along information the query didn't already supply. Tested across five reasoning benchmarks, GRIP beat existing iterative RAG baselines, cut a diagnostic measuring how much the model's internal state depends on the query alone by roughly 30 times, from 14.8 bits to 0.47, and reduced hallucination by 73 percent. A follow-up analysis found the bottleneck's output lands in representational space further from the query than in baseline models, evidence the fix is doing what it claims.
RAG is the standard answer to chatbots making things up: fetch real documents, ground the response in them. But if the query still dominates the model's internal state, that grounding is theater, the system looks like it consulted its sources while mostly pattern-matching the question. GRIP's bottleneck is a structural constraint rather than a prompting trick, which matters for any product claiming its answers are sourced rather than just decorated with citations.
Still, this is a benchmark result on five academic test sets, not a shipped feature, and a 73 percent drop in hallucinations on paper rarely survives contact with messy production data and real user queries.