A research team has found a way to stop language models from overriding accurate retrieved information with their own baked-in assumptions.
Large language models paired with retrieval-augmented generation - a setup where the model pulls in external documents before responding - still frequently contradict those documents in their outputs. Researchers at OpenBMB traced the problem to a specific subset of mid-to-deep feed-forward networks inside the model that activate heavily during unfaithful generation. Their proposed fix, ParamMute, selectively suppresses those layers during inference, nudging the model to trust what it retrieved rather than what it thinks it already knows. They also released CoFaithfulQA, a benchmark built to test faithfulness specifically when internal model knowledge conflicts with accurate external evidence - a gap the existing ConFiQA benchmark did not fully cover.
This matters because most existing RAG improvements focus on the retrieval side: better chunking, better search, better reranking. ParamMute cuts the other way, editing the generation side at the architectural level. If the technique holds up at production scale, it could reduce hallucination without requiring retraining or external guardrails.
The approach is surgical, not lobotomous - only the offending FFN layers are suppressed, not the whole network. That said, surgical interventions in neural networks have a history of unexpected side effects, and the benchmarks here are controlled scenarios; real-world RAG pipelines are considerably messier.