Researchers claim to have found a way to stop a language model from making up facts it should already know.
The method, called Fact Grounded Attention (FGA), doesn't fact-check text after it's generated or stuff retrieved documents into the prompt, the way retrieval-augmented generation does. Instead it edits the attention mechanism itself, adjusting the pre-softmax attention scores so verified facts get baked directly into how the model weighs its own output. The team tested it on 1,107 technical questions about smartphones, laptops, and electric vehicles. A stock Llama 3.2 model answered accurately 6.3% of the time; the same model with FGA hit 99.7%, and the paper says updating its knowledge base takes under a second, versus hours for approaches that edit model parameters directly.
If those numbers hold up outside this paper, it is a bigger claim than another RAG variant: it targets hallucination at the mechanism level instead of patching it after the fact, and near-instant updates would matter for anything that changes quickly, like prices or specs. That is also exactly the kind of framing, claiming to "eliminate hallucination entirely," that should make you want independent verification before you believe it.
A jump from 6.3% accuracy to 99.7% is the kind of number that belongs in a follow-up study on harder, broader knowledge, not a victory lap over gadget trivia.