Researchers recover suppressed answers by altering an AI model's internal signals

Researchers testing Minerva-7B found that its answers often concealed distinctions represented inside the model. In a preprint, they report that the model responded identically to 63.7% of 124 prompt pairs designed to test professional risk, despite showing internal differences between the paired requests. The team also tested 25 facts presented under five linguistic framings. When false assumptions suppressed a correct answer, removing the internal signal associated with the falsehood restored that answer in 11 of 25 cases. The results show a possible way to test whether an AI system represents a fact it fails to state, but the intervention was demonstrated on one model.