Researchers have built a fix for one of retrieval-augmented generation's oldest problems: knowing when to trust the evidence it just fetched.
The method, called Intent-Guided Decoding (IGD), targets a flaw in most RAG systems. They apply the same trust policy to every retrieved document, whether it is accurate, irrelevant, or flat-out wrong. IGD instead reads user intent and arbitrates, token by token, between what was retrieved and what the model already knows from training. The researchers tested it across five large language models and six benchmarks - three built to reward strict context-following, three built to expose factual conflicts. On the conflict benchmarks, IGD beat standard RAG by as much as 65.4 percentage points, without sacrificing accuracy on the tasks where sticking to the provided context is the whole point.
This matters because RAG's reputation problem is not retrieval, it is judgment. Every chatbot that cites a source but still gets the fact wrong is a trust-policy failure, not a search failure. A model that can dynamically decide when to override bad context - instead of parroting it or ignoring it wholesale - closes a gap that has made enterprise RAG deployments unreliable for anything high-stakes, from legal research to customer support.
The 65-point gain is on adversarial benchmarks designed to surface exactly this failure mode, so treat it as a ceiling, not a typical result. Still, it is a reminder that the next wave of RAG improvements will come from decoding strategy, not bigger retrieval indexes.